Merge 4ffae801be into b94a7167a8

Bump libpqxx to v7.7.5

README.md

@@ -40,17 +40,8 @@ Every month we get together with the community (users, contributors, customers,
 
 Keep an eye out for upcoming meetups and events around the world. Somewhere else you want us to be? Please feel free to reach out to tyler `<at>` clickhouse `<dot>` com. You can also peruse [ClickHouse Events](https://clickhouse.com/company/news-events) for a list of all upcoming trainings, meetups, speaking engagements, etc.
 
-The following upcoming meetups are featuring creator of ClickHouse & CTO, Alexey Milovidov:
+Upcoming meetups
 
-* [Raleigh Meetup (Deutsche Bank)](https://www.meetup.com/triangletechtalks/events/302723486/) - September 9
-* [New York Meetup (Rokt)](https://www.meetup.com/clickhouse-new-york-user-group/events/302575342) - September 10
-* [Chicago Meetup (Jump Capital)](https://lu.ma/43tvmrfw) - September 12
-
-Other upcoming meetups
-
-* [Toronto Meetup (Shopify)](https://www.meetup.com/clickhouse-toronto-user-group/events/301490855/) - September 10
-* [Austin Meetup](https://www.meetup.com/clickhouse-austin-user-group/events/302558689/) - September 17
-* [London Meetup](https://www.meetup.com/clickhouse-london-user-group/events/302977267) - September 17
 * [Bangalore Meetup](https://www.meetup.com/clickhouse-bangalore-user-group/events/303208274/) - September 18
 * [Tel Aviv Meetup](https://www.meetup.com/clickhouse-meetup-israel/events/303095121) - September 22
 * [Jakarta Meetup](https://www.meetup.com/clickhouse-indonesia-user-group/events/303191359/) - October 1
@@ -62,13 +53,20 @@ Other upcoming meetups
 * [Dubai Meetup](https://www.meetup.com/clickhouse-dubai-meetup-group/events/303096989/) - November 21
 * [Paris Meetup](https://www.meetup.com/clickhouse-france-user-group/events/303096434) - November 26
 
-Recently completed events
+Recently completed meetups
+
 * [ClickHouse Guangzhou User Group Meetup](https://mp.weixin.qq.com/s/GSvo-7xUoVzCsuUvlLTpCw) - August 25
 * [Seattle Meetup (Statsig)](https://www.meetup.com/clickhouse-seattle-user-group/events/302518075/) - August 27
 * [Melbourne Meetup](https://www.meetup.com/clickhouse-australia-user-group/events/302732666/) - August 27
 * [Sydney Meetup](https://www.meetup.com/clickhouse-australia-user-group/events/302862966/) - September 5
 * [Zurich Meetup](https://www.meetup.com/clickhouse-switzerland-meetup-group/events/302267429/) - September 5
 * [San Francisco Meetup (Cloudflare)](https://www.meetup.com/clickhouse-silicon-valley-meetup-group/events/302540575) - September 5
+* [Raleigh Meetup (Deutsche Bank)](https://www.meetup.com/triangletechtalks/events/302723486/) - September 9
+* [New York Meetup (Rokt)](https://www.meetup.com/clickhouse-new-york-user-group/events/302575342) - September 10
+* [Toronto Meetup (Shopify)](https://www.meetup.com/clickhouse-toronto-user-group/events/301490855/) - September 10
+* [Chicago Meetup (Jump Capital)](https://lu.ma/43tvmrfw) - September 12
+* [London Meetup](https://www.meetup.com/clickhouse-london-user-group/events/302977267) - September 17
+* [Austin Meetup](https://www.meetup.com/clickhouse-austin-user-group/events/302558689/) - September 17
 
 ## Recent Recordings
 * **Recent Meetup Videos**: [Meetup Playlist](https://www.youtube.com/playlist?list=PL0Z2YDlm0b3iNDUzpY1S3L_iV4nARda_U) Whenever possible recordings of the ClickHouse Community Meetups are edited and presented as individual talks. Current featuring "Modern SQL in 2023", "Fast, Concurrent, and Consistent Asynchronous INSERTS in ClickHouse", and "Full-Text Indices: Design and Experiments"

base/poco/Crypto/include/Poco/Crypto/EVPPKey.h

@@ -188,8 +188,9 @@ namespace Crypto
                     pFile = fopen(keyFile.c_str(), "r");
                     if (pFile)
                     {
-                        pem_password_cb * pCB = pass.empty() ? (pem_password_cb *)0 : &passCB;
+                        pem_password_cb * pCB = &passCB;
-                        void * pPassword = pass.empty() ? (void *)0 : (void *)pass.c_str();
+                        static constexpr char * no_password = "";
+                        void * pPassword = pass.empty() ? (void *)no_password : (void *)pass.c_str();
                         if (readFunc(pFile, &pKey, pCB, pPassword))
                         {
                             fclose(pFile);
@@ -225,6 +226,13 @@ namespace Crypto
         error:
             if (pFile)
                 fclose(pFile);
+            if (*ppKey)
+            {
+                if constexpr (std::is_same_v<K, EVP_PKEY>)
+                    EVP_PKEY_free(*ppKey);
+                else
+                    EC_KEY_free(*ppKey);
+            }
             throw OpenSSLException("EVPKey::loadKey(string)");
         }
 
@@ -286,6 +294,13 @@ namespace Crypto
         error:
             if (pBIO)
                 BIO_free(pBIO);
+            if (*ppKey)
+            {
+                if constexpr (std::is_same_v<K, EVP_PKEY>)
+                    EVP_PKEY_free(*ppKey);
+                else
+                    EC_KEY_free(*ppKey);
+            }
             throw OpenSSLException("EVPKey::loadKey(stream)");
         }
 

base/poco/NetSSL_OpenSSL/include/Poco/Net/Context.h

@@ -248,6 +248,9 @@ namespace Net
         SSL_CTX * sslContext() const;
         /// Returns the underlying OpenSSL SSL Context object.
 
+        SSL_CTX * takeSslContext();
+        /// Takes ownership of the underlying OpenSSL SSL Context object.
+
         Usage usage() const;
         /// Returns whether the context is for use by a client or by a server
         /// and whether TLSv1 is required.
@@ -401,6 +404,13 @@ namespace Net
         return _pSSLContext;
     }
 
+    inline SSL_CTX * Context::takeSslContext()
+    {
+        auto * result = _pSSLContext;
+        _pSSLContext = nullptr;
+        return result;
+    }
+
 
     inline bool Context::extendedCertificateVerificationEnabled() const
     {

base/poco/NetSSL_OpenSSL/src/Context.cpp

@@ -106,6 +106,11 @@ Context::Context(
 
 Context::~Context()
 {
+    if (_pSSLContext == nullptr)
+    {
+        return;
+    }
+
 	try
 	{
 		SSL_CTX_free(_pSSLContext);

contrib/libpqxx

@@ -1 +1 @@
-Subproject commit c995193a3a14d71f4711f1f421f65a1a1db64640
+Subproject commit 41e4c331564167cca97ad6eccbd5b8879c2ca044

contrib/libpqxx-cmake/CMakeLists.txt

@@ -1,9 +1,9 @@
 set (LIBRARY_DIR "${ClickHouse_SOURCE_DIR}/contrib/libpqxx")
 
 set (SRCS
-    "${LIBRARY_DIR}/src/strconv.cxx"
     "${LIBRARY_DIR}/src/array.cxx"
     "${LIBRARY_DIR}/src/binarystring.cxx"
+    "${LIBRARY_DIR}/src/blob.cxx"
     "${LIBRARY_DIR}/src/connection.cxx"
     "${LIBRARY_DIR}/src/cursor.cxx"
     "${LIBRARY_DIR}/src/encodings.cxx"
@@ -12,59 +12,25 @@ set (SRCS
     "${LIBRARY_DIR}/src/field.cxx"
     "${LIBRARY_DIR}/src/largeobject.cxx"
     "${LIBRARY_DIR}/src/notification.cxx"
+    "${LIBRARY_DIR}/src/params.cxx"
     "${LIBRARY_DIR}/src/pipeline.cxx"
     "${LIBRARY_DIR}/src/result.cxx"
     "${LIBRARY_DIR}/src/robusttransaction.cxx"
+    "${LIBRARY_DIR}/src/row.cxx"
     "${LIBRARY_DIR}/src/sql_cursor.cxx"
+    "${LIBRARY_DIR}/src/strconv.cxx"
     "${LIBRARY_DIR}/src/stream_from.cxx"
     "${LIBRARY_DIR}/src/stream_to.cxx"
     "${LIBRARY_DIR}/src/subtransaction.cxx"
+    "${LIBRARY_DIR}/src/time.cxx"
     "${LIBRARY_DIR}/src/transaction.cxx"
     "${LIBRARY_DIR}/src/transaction_base.cxx"
-    "${LIBRARY_DIR}/src/row.cxx"
-    "${LIBRARY_DIR}/src/params.cxx"
     "${LIBRARY_DIR}/src/util.cxx"
     "${LIBRARY_DIR}/src/version.cxx"
+    "${LIBRARY_DIR}/src/wait.cxx"
 )
 
-# Need to explicitly include each header file, because in the directory include/pqxx there are also files
+add_library(_libpqxx ${SRCS})
-# like just 'array'. So if including the whole directory with `target_include_directories`, it will make
-# conflicts with all includes of <array>.
-set (HDRS
-    "${LIBRARY_DIR}/include/pqxx/array.hxx"
-    "${LIBRARY_DIR}/include/pqxx/params.hxx"
-    "${LIBRARY_DIR}/include/pqxx/binarystring.hxx"
-    "${LIBRARY_DIR}/include/pqxx/composite.hxx"
-    "${LIBRARY_DIR}/include/pqxx/connection.hxx"
-    "${LIBRARY_DIR}/include/pqxx/cursor.hxx"
-    "${LIBRARY_DIR}/include/pqxx/dbtransaction.hxx"
-    "${LIBRARY_DIR}/include/pqxx/errorhandler.hxx"
-    "${LIBRARY_DIR}/include/pqxx/except.hxx"
-    "${LIBRARY_DIR}/include/pqxx/field.hxx"
-    "${LIBRARY_DIR}/include/pqxx/isolation.hxx"
-    "${LIBRARY_DIR}/include/pqxx/largeobject.hxx"
-    "${LIBRARY_DIR}/include/pqxx/nontransaction.hxx"
-    "${LIBRARY_DIR}/include/pqxx/notification.hxx"
-    "${LIBRARY_DIR}/include/pqxx/pipeline.hxx"
-    "${LIBRARY_DIR}/include/pqxx/prepared_statement.hxx"
-    "${LIBRARY_DIR}/include/pqxx/result.hxx"
-    "${LIBRARY_DIR}/include/pqxx/robusttransaction.hxx"
-    "${LIBRARY_DIR}/include/pqxx/row.hxx"
-    "${LIBRARY_DIR}/include/pqxx/separated_list.hxx"
-    "${LIBRARY_DIR}/include/pqxx/strconv.hxx"
-    "${LIBRARY_DIR}/include/pqxx/stream_from.hxx"
-    "${LIBRARY_DIR}/include/pqxx/stream_to.hxx"
-    "${LIBRARY_DIR}/include/pqxx/subtransaction.hxx"
-    "${LIBRARY_DIR}/include/pqxx/transaction.hxx"
-    "${LIBRARY_DIR}/include/pqxx/transaction_base.hxx"
-    "${LIBRARY_DIR}/include/pqxx/types.hxx"
-    "${LIBRARY_DIR}/include/pqxx/util.hxx"
-    "${LIBRARY_DIR}/include/pqxx/version.hxx"
-    "${LIBRARY_DIR}/include/pqxx/zview.hxx"
-)
-
-add_library(_libpqxx ${SRCS} ${HDRS})
-
 target_link_libraries(_libpqxx PUBLIC ch_contrib::libpq)
 target_include_directories (_libpqxx SYSTEM BEFORE PUBLIC "${LIBRARY_DIR}/include")
 

contrib/postgres

@@ -1 +1 @@
-Subproject commit 665ff8c164d56d012e359735efe4d400c0564b44
+Subproject commit cfd77000af28469fcb650485bad65a35e7649e41

contrib/postgres-cmake/pg_config.h

@@ -1,18 +1,6 @@
 /* src/include/pg_config.h.  Generated from pg_config.h.in by configure.  */
 /* src/include/pg_config.h.in.  Generated from configure.in by autoheader.  */
 
-/* Define to the type of arg 1 of 'accept' */
-#define ACCEPT_TYPE_ARG1 int
-
-/* Define to the type of arg 2 of 'accept' */
-#define ACCEPT_TYPE_ARG2 struct sockaddr *
-
-/* Define to the type of arg 3 of 'accept' */
-#define ACCEPT_TYPE_ARG3 size_t
-
-/* Define to the return type of 'accept' */
-#define ACCEPT_TYPE_RETURN int
-
 /* Define if building universal (internal helper macro) */
 /* #undef AC_APPLE_UNIVERSAL_BUILD */
 
@@ -49,6 +37,9 @@
 /* Define to the default TCP port number as a string constant. */
 #define DEF_PGPORT_STR "5432"
 
+/* Define to the file name extension of dynamically-loadable modules. */
+#define DLSUFFIX ".so"
+
 /* Define to build with GSSAPI support. (--with-gssapi) */
 //#define ENABLE_GSS 0
 
@@ -122,6 +113,9 @@
    don't. */
 #define HAVE_DECL_SNPRINTF 1
 
+/* Define to 1 if you have the declaration of `sigwait', and to 0 if you don't. */
+#define HAVE_DECL_SIGWAIT 1
+
 /* Define to 1 if you have the declaration of `strlcat', and to 0 if you
    don't. */
 #if OS_DARWIN
@@ -257,6 +251,9 @@
 /* Define to 1 if you have the `inet_aton' function. */
 #define HAVE_INET_ATON 1
 
+/* Define to 1 if you have the `inet_pton' function. */
+#define HAVE_INET_PTON 1
+
 /* Define to 1 if the system has the type `int64'. */
 /* #undef HAVE_INT64 */
 
@@ -323,6 +320,9 @@
 /* Define to 1 if you have the `z' library (-lz). */
 #define HAVE_LIBZ 1
 
+/* Define to 1 if you have the `zstd' library (-lzstd). */
+/* #undef HAVE_LIBZSTD */
+
 /* Define to 1 if constants of type 'long long int' should have the suffix LL.
    */
 #define HAVE_LL_CONSTANTS 1
@@ -378,6 +378,9 @@
 /* Define to 1 if you have the <poll.h> header file. */
 #define HAVE_POLL_H 1
 
+/* Define to 1 if you have a POSIX-conforming sigwait declaration. */
+/* #undef HAVE_POSIX_DECL_SIGWAIT */
+
 /* Define to 1 if you have the `posix_fadvise' function. */
 #define HAVE_POSIX_FADVISE 1
 
@@ -408,9 +411,6 @@
 /* Define to 1 if you have the <pwd.h> header file. */
 #define HAVE_PWD_H 1
 
-/* Define to 1 if you have the `random' function. */
-#define HAVE_RANDOM 1
-
 /* Define to 1 if you have the <readline.h> header file. */
 /* #undef HAVE_READLINE_H */
 
@@ -426,10 +426,6 @@
 /* Define to 1 if you have the `rint' function. */
 #define HAVE_RINT 1
 
-/* Define to 1 if you have the global variable
-   'rl_completion_append_character'. */
-/* #undef HAVE_RL_COMPLETION_APPEND_CHARACTER */
-
 /* Define to 1 if you have the `rl_completion_matches' function. */
 #define HAVE_RL_COMPLETION_MATCHES 1
 
@@ -439,6 +435,9 @@
 /* Define to 1 if you have the `rl_reset_screen_size' function. */
 /* #undef HAVE_RL_RESET_SCREEN_SIZE */
 
+/* Define to 1 if you have the `rl_variable_bind' function. */
+#define HAVE_RL_VARIABLE_BIND 1
+
 /* Define to 1 if you have the <security/pam_appl.h> header file. */
 #define HAVE_SECURITY_PAM_APPL_H 1
 
@@ -451,6 +450,9 @@
 /* Define to 1 if you have the `shm_open' function. */
 #define HAVE_SHM_OPEN 1
 
+/* Define to 1 if the system has the type `socklen_t'. */
+#define HAVE_SOCKLEN_T 1
+
 /* Define to 1 if you have the `sigprocmask' function. */
 #define HAVE_SIGPROCMASK 1
 
@@ -466,9 +468,6 @@
 /* Define to 1 if you have spinlocks. */
 #define HAVE_SPINLOCKS 1
 
-/* Define to 1 if you have the `srandom' function. */
-#define HAVE_SRANDOM 1
-
 /* Define to 1 if you have the `SSL_CTX_set_num_tickets' function. */
 /* #define HAVE_SSL_CTX_SET_NUM_TICKETS */
 
@@ -885,6 +884,9 @@
 /* Define to select Win32-style shared memory. */
 /* #undef USE_WIN32_SHARED_MEMORY */
 
+/* Define to 1 to build with ZSTD support. (--with-zstd) */
+/* #undef USE_ZSTD */
+
 /* Define to 1 if `wcstombs_l' requires <xlocale.h>. */
 /* #undef WCSTOMBS_L_IN_XLOCALE */
 

docs/en/sql-reference/aggregate-functions/reference/quantileddsketch.md

@@ -9,7 +9,7 @@ Computes an approximate [quantile](https://en.wikipedia.org/wiki/Quantile) of a
 **Syntax**
 
 ``` sql
-quantileDDsketch[relative_accuracy, (level)](expr)
+quantileDD(relative_accuracy, [level])(expr)
 ```
 
 **Arguments**

src/Access/AccessBackup.cpp

@@ -29,6 +29,7 @@ namespace DB
 namespace ErrorCodes
 {
     extern const int CANNOT_RESTORE_TABLE;
+    extern const int ACCESS_ENTITY_ALREADY_EXISTS;
     extern const int LOGICAL_ERROR;
 }
 
@@ -175,9 +176,46 @@ namespace
         return res;
     }
 
-    std::unordered_map<UUID, UUID> resolveDependencies(const std::unordered_map<UUID, std::pair<String, AccessEntityType>> & dependencies, const AccessControl & access_control, bool allow_unresolved_dependencies)
+    /// Checks if new entities (which we're going to restore) already exist,
+    /// and either skips them or throws an exception depending on the restore settings.
+    void checkExistingEntities(std::vector<std::pair<UUID, AccessEntityPtr>> & entities,
+                               std::unordered_map<UUID, UUID> & old_to_new_id,
+                               const AccessControl & access_control,
+                               RestoreAccessCreationMode creation_mode)
+    {
+        if (creation_mode == RestoreAccessCreationMode::kReplace)
+            return;
+
+        auto should_skip = [&](const std::pair<UUID, AccessEntityPtr> & id_and_entity)
+        {
+            const auto & id = id_and_entity.first;
+            const auto & entity = *id_and_entity.second;
+            auto existing_id = access_control.find(entity.getType(), entity.getName());
+            if (!existing_id)
+            {
+                return false;
+            }
+            else if (creation_mode == RestoreAccessCreationMode::kCreateIfNotExists)
+            {
+                old_to_new_id[id] = *existing_id;
+                return true;
+            }
+            else
+            {
+                throw Exception(ErrorCodes::ACCESS_ENTITY_ALREADY_EXISTS, "Cannot restore {} because it already exists", entity.formatTypeWithName());
+            }
+        };
+
+        std::erase_if(entities, should_skip);
+    }
+
+    /// If new entities (which we're going to restore) depend on other entities which are not going to be restored or not present in the backup
+    /// then we should try to replace those dependencies with already existing entities.
+    void resolveDependencies(const std::unordered_map<UUID, std::pair<String, AccessEntityType>> & dependencies,
+                             std::unordered_map<UUID, UUID> & old_to_new_ids,
+                             const AccessControl & access_control,
+                             bool allow_unresolved_dependencies)
     {
-        std::unordered_map<UUID, UUID> old_to_new_ids;
         for (const auto & [id, name_and_type] : dependencies)
         {
             std::optional<UUID> new_id;
@@ -188,9 +226,9 @@ namespace
             if (new_id)
                 old_to_new_ids.emplace(id, *new_id);
         }
-        return old_to_new_ids;
     }
 
+    /// Generates random IDs for the new entities.
     void generateRandomIDs(std::vector<std::pair<UUID, AccessEntityPtr>> & entities, std::unordered_map<UUID, UUID> & old_to_new_ids)
     {
         Poco::UUIDGenerator generator;
@@ -203,27 +241,12 @@ namespace
         }
     }
 
-    void replaceDependencies(std::vector<std::pair<UUID, AccessEntityPtr>> & entities, const std::unordered_map<UUID, UUID> & old_to_new_ids)
+    /// Updates dependencies of the new entities using a specified map.
+    void replaceDependencies(std::vector<std::pair<UUID, AccessEntityPtr>> & entities,
+                             const std::unordered_map<UUID, UUID> & old_to_new_ids)
     {
         for (auto & entity : entities | boost::adaptors::map_values)
-        {
+            IAccessEntity::replaceDependencies(entity, old_to_new_ids);
-            bool need_replace = false;
-            for (const auto & dependency : entity->findDependencies())
-            {
-                if (old_to_new_ids.contains(dependency))
-                {
-                    need_replace = true;
-                    break;
-                }
-            }
-
-            if (!need_replace)
-                continue;
-
-            auto new_entity = entity->clone();
-            new_entity->replaceDependencies(old_to_new_ids);
-            entity = new_entity;
-        }
     }
 
     AccessRightsElements getRequiredAccessToRestore(const std::vector<std::pair<UUID, AccessEntityPtr>> & entities)
@@ -314,7 +337,9 @@ std::pair<String, BackupEntryPtr> makeBackupEntryForAccess(
 
 AccessRestorerFromBackup::AccessRestorerFromBackup(
     const BackupPtr & backup_, const RestoreSettings & restore_settings_)
-    : backup(backup_), allow_unresolved_access_dependencies(restore_settings_.allow_unresolved_access_dependencies)
+    : backup(backup_)
+    , creation_mode(restore_settings_.create_access)
+    , allow_unresolved_dependencies(restore_settings_.allow_unresolved_access_dependencies)
 {
 }
 
@@ -362,7 +387,9 @@ std::vector<std::pair<UUID, AccessEntityPtr>> AccessRestorerFromBackup::getAcces
 {
     auto new_entities = entities;
 
-    auto old_to_new_ids = resolveDependencies(dependencies, access_control, allow_unresolved_access_dependencies);
+    std::unordered_map<UUID, UUID> old_to_new_ids;
+    checkExistingEntities(new_entities, old_to_new_ids, access_control, creation_mode);
+    resolveDependencies(dependencies, old_to_new_ids, access_control, allow_unresolved_dependencies);
     generateRandomIDs(new_entities, old_to_new_ids);
     replaceDependencies(new_entities, old_to_new_ids);
 

src/Access/AccessBackup.h

@@ -17,6 +17,7 @@ using BackupPtr = std::shared_ptr<const IBackup>;
 class IBackupEntry;
 using BackupEntryPtr = std::shared_ptr<const IBackupEntry>;
 struct RestoreSettings;
+enum class RestoreAccessCreationMode : uint8_t;
 
 
 /// Makes a backup of access entities of a specified type.
@@ -45,7 +46,8 @@ public:
 
 private:
     BackupPtr backup;
-    bool allow_unresolved_access_dependencies = false;
+    RestoreAccessCreationMode creation_mode;
+    bool allow_unresolved_dependencies = false;
     std::vector<std::pair<UUID, AccessEntityPtr>> entities;
     std::unordered_map<UUID, std::pair<String, AccessEntityType>> dependencies;
     std::unordered_set<String> data_paths;

src/Access/AccessControl.cpp

@@ -544,9 +544,9 @@ scope_guard AccessControl::subscribeForChanges(const std::vector<UUID> & ids, co
     return changes_notifier->subscribeForChanges(ids, handler);
 }
 
-bool AccessControl::insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists)
+bool AccessControl::insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
-    if (MultipleAccessStorage::insertImpl(id, entity, replace_if_exists, throw_if_exists))
+    if (MultipleAccessStorage::insertImpl(id, entity, replace_if_exists, throw_if_exists, conflicting_id))
     {
         changes_notifier->sendNotifications();
         return true;

src/Access/AccessControl.h

@@ -243,7 +243,7 @@ private:
     class CustomSettingsPrefixes;
     class PasswordComplexityRules;
 
-    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists) override;
+    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id) override;
     bool removeImpl(const UUID & id, bool throw_if_not_exists) override;
     bool updateImpl(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists) override;
 

src/Access/DiskAccessStorage.cpp

@@ -1,8 +1,6 @@
 #include <Access/DiskAccessStorage.h>
 #include <Access/AccessEntityIO.h>
 #include <Access/AccessChangesNotifier.h>
-#include <Backups/RestorerFromBackup.h>
-#include <Backups/RestoreSettings.h>
 #include <IO/WriteHelpers.h>
 #include <IO/ReadHelpers.h>
 #include <IO/ReadBufferFromFile.h>
@@ -418,7 +416,7 @@ void DiskAccessStorage::setAllInMemory(const std::vector<std::pair<UUID, AccessE
 
     /// Insert or update entities.
     for (const auto & [id, entity] : entities_without_conflicts)
-        insertNoLock(id, entity, /* replace_if_exists = */ true, /* throw_if_exists = */ false, /* write_on_disk= */ false);
+        insertNoLock(id, entity, /* replace_if_exists = */ true, /* throw_if_exists = */ false, /* conflicting_id = */ nullptr, /* write_on_disk= */ false);
 }
 
 void DiskAccessStorage::removeAllExceptInMemory(const boost::container::flat_set<UUID> & ids_to_keep)
@@ -507,14 +505,14 @@ std::optional<std::pair<String, AccessEntityType>> DiskAccessStorage::readNameWi
 }
 
 
-bool DiskAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists)
+bool DiskAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
     std::lock_guard lock{mutex};
-    return insertNoLock(id, new_entity, replace_if_exists, throw_if_exists, /* write_on_disk = */ true);
+    return insertNoLock(id, new_entity, replace_if_exists, throw_if_exists, conflicting_id, /* write_on_disk = */ true);
 }
 
 
-bool DiskAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, bool write_on_disk)
+bool DiskAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id, bool write_on_disk)
 {
     const String & name = new_entity->getName();
     AccessEntityType type = new_entity->getType();
@@ -533,9 +531,15 @@ bool DiskAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr & ne
     if (name_collision && !replace_if_exists)
     {
         if (throw_if_exists)
+        {
             throwNameCollisionCannotInsert(type, name);
+        }
         else
+        {
+            if (conflicting_id)
+                *conflicting_id = id_by_name;
             return false;
+        }
     }
 
     auto it_by_id = entries_by_id.find(id);
@@ -548,7 +552,11 @@ bool DiskAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr & ne
             throwIDCollisionCannotInsert(id, type, name, existing_entry.type, existing_entry.name);
         }
         else
+        {
+            if (conflicting_id)
+                *conflicting_id = id;
             return false;
+        }
     }
 
     if (write_on_disk)
@@ -727,25 +735,4 @@ void DiskAccessStorage::deleteAccessEntityOnDisk(const UUID & id) const
         throw Exception(ErrorCodes::FILE_DOESNT_EXIST, "Couldn't delete {}", file_path);
 }
 
-
-void DiskAccessStorage::restoreFromBackup(RestorerFromBackup & restorer)
-{
-    if (!isRestoreAllowed())
-        throwRestoreNotAllowed();
-
-    auto entities = restorer.getAccessEntitiesToRestore();
-    if (entities.empty())
-        return;
-
-    auto create_access = restorer.getRestoreSettings().create_access;
-    bool replace_if_exists = (create_access == RestoreAccessCreationMode::kReplace);
-    bool throw_if_exists = (create_access == RestoreAccessCreationMode::kCreate);
-
-    restorer.addDataRestoreTask([this, my_entities = std::move(entities), replace_if_exists, throw_if_exists]
-    {
-        for (const auto & [id, entity] : my_entities)
-            insert(id, entity, replace_if_exists, throw_if_exists);
-    });
-}
-
 }

src/Access/DiskAccessStorage.h

@@ -34,14 +34,13 @@ public:
     bool exists(const UUID & id) const override;
 
     bool isBackupAllowed() const override { return backup_allowed; }
-    void restoreFromBackup(RestorerFromBackup & restorer) override;
 
 private:
     std::optional<UUID> findImpl(AccessEntityType type, const String & name) const override;
     std::vector<UUID> findAllImpl(AccessEntityType type) const override;
     AccessEntityPtr readImpl(const UUID & id, bool throw_if_not_exists) const override;
     std::optional<std::pair<String, AccessEntityType>> readNameWithTypeImpl(const UUID & id, bool throw_if_not_exists) const override;
-    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists) override;
+    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id) override;
     bool removeImpl(const UUID & id, bool throw_if_not_exists) override;
     bool updateImpl(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists) override;
 
@@ -55,7 +54,7 @@ private:
     void listsWritingThreadFunc() TSA_NO_THREAD_SAFETY_ANALYSIS;
     void stopListsWritingThread();
 
-    bool insertNoLock(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, bool write_on_disk) TSA_REQUIRES(mutex);
+    bool insertNoLock(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id, bool write_on_disk) TSA_REQUIRES(mutex);
     bool updateNoLock(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists, bool write_on_disk) TSA_REQUIRES(mutex);
     bool removeNoLock(const UUID & id, bool throw_if_not_exists, bool write_on_disk) TSA_REQUIRES(mutex);
 

src/Access/IAccessEntity.cpp

@@ -9,4 +9,28 @@ bool IAccessEntity::equal(const IAccessEntity & other) const
     return (name == other.name) && (getType() == other.getType());
 }
 
+void IAccessEntity::replaceDependencies(std::shared_ptr<const IAccessEntity> & entity, const std::unordered_map<UUID, UUID> & old_to_new_ids)
+{
+    if (old_to_new_ids.empty())
+        return;
+
+    bool need_replace_dependencies = false;
+    auto dependencies = entity->findDependencies();
+    for (const auto & dependency : dependencies)
+    {
+        if (old_to_new_ids.contains(dependency))
+        {
+            need_replace_dependencies = true;
+            break;
+        }
+    }
+
+    if (!need_replace_dependencies)
+        return;
+
+    auto new_entity = entity->clone();
+    new_entity->replaceDependencies(old_to_new_ids);
+    entity = new_entity;
+}
+
 }

src/Access/IAccessEntity.h

@@ -50,7 +50,8 @@ struct IAccessEntity
     virtual std::vector<UUID> findDependencies() const { return {}; }
 
     /// Replaces dependencies according to a specified map.
-    virtual void replaceDependencies(const std::unordered_map<UUID, UUID> & /* old_to_new_ids */) {}
+    void replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) { doReplaceDependencies(old_to_new_ids); }
+    static void replaceDependencies(std::shared_ptr<const IAccessEntity> & entity, const std::unordered_map<UUID, UUID> & old_to_new_ids);
 
     /// Whether this access entity should be written to a backup.
     virtual bool isBackupAllowed() const { return false; }
@@ -66,6 +67,8 @@ protected:
     {
         return std::make_shared<EntityClassT>(typeid_cast<const EntityClassT &>(*this));
     }
+
+    virtual void doReplaceDependencies(const std::unordered_map<UUID, UUID> & /* old_to_new_ids */) {}
 };
 
 using AccessEntityPtr = std::shared_ptr<const IAccessEntity>;

src/Access/IAccessStorage.cpp

@@ -4,6 +4,8 @@
 #include <Access/User.h>
 #include <Access/AccessBackup.h>
 #include <Backups/BackupEntriesCollector.h>
+#include <Backups/RestorerFromBackup.h>
+#include <Backups/RestoreSettings.h>
 #include <Common/Exception.h>
 #include <Common/quoteString.h>
 #include <Common/callOnce.h>
@@ -14,10 +16,11 @@
 #include <base/FnTraits.h>
 #include <boost/algorithm/string/join.hpp>
 #include <boost/algorithm/string/replace.hpp>
+#include <boost/range/adaptor/map.hpp>
 #include <boost/range/adaptor/reversed.hpp>
+#include <boost/range/algorithm/copy.hpp>
 #include <boost/range/algorithm_ext/erase.hpp>
 
-
 namespace DB
 {
 namespace ErrorCodes
@@ -178,20 +181,20 @@ UUID IAccessStorage::insert(const AccessEntityPtr & entity)
     return *insert(entity, /* replace_if_exists = */ false, /* throw_if_exists = */ true);
 }
 
-std::optional<UUID> IAccessStorage::insert(const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists)
+std::optional<UUID> IAccessStorage::insert(const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
     auto id = generateRandomID();
 
-    if (insert(id, entity, replace_if_exists, throw_if_exists))
+    if (insert(id, entity, replace_if_exists, throw_if_exists, conflicting_id))
         return id;
 
     return std::nullopt;
 }
 
 
-bool IAccessStorage::insert(const DB::UUID & id, const DB::AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists)
+bool IAccessStorage::insert(const DB::UUID & id, const DB::AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
-    return insertImpl(id, entity, replace_if_exists, throw_if_exists);
+    return insertImpl(id, entity, replace_if_exists, throw_if_exists, conflicting_id);
 }
 
 
@@ -285,7 +288,7 @@ std::vector<UUID> IAccessStorage::insertOrReplace(const std::vector<AccessEntity
 }
 
 
-bool IAccessStorage::insertImpl(const UUID &, const AccessEntityPtr & entity, bool, bool)
+bool IAccessStorage::insertImpl(const UUID &, const AccessEntityPtr & entity, bool, bool, UUID *)
 {
     if (isReadOnly())
         throwReadonlyCannotInsert(entity->getType(), entity->getName());
@@ -611,12 +614,51 @@ void IAccessStorage::backup(BackupEntriesCollector & backup_entries_collector, c
 }
 
 
-void IAccessStorage::restoreFromBackup(RestorerFromBackup &)
+void IAccessStorage::restoreFromBackup(RestorerFromBackup & restorer)
 {
     if (!isRestoreAllowed())
         throwRestoreNotAllowed();
 
-    throw Exception(ErrorCodes::NOT_IMPLEMENTED, "restoreFromBackup() is not implemented in {}", getStorageType());
+    if (isReplicated() && !acquireReplicatedRestore(restorer))
+        return;
+
+    auto entities = restorer.getAccessEntitiesToRestore();
+    if (entities.empty())
+        return;
+
+    auto create_access = restorer.getRestoreSettings().create_access;
+    bool replace_if_exists = (create_access == RestoreAccessCreationMode::kReplace);
+    bool throw_if_exists = (create_access == RestoreAccessCreationMode::kCreate);
+
+    restorer.addDataRestoreTask([this, entities_to_restore = std::move(entities), replace_if_exists, throw_if_exists] mutable
+    {
+        std::unordered_map<UUID, UUID> new_to_existing_ids;
+        for (auto & [id, entity] : entities_to_restore)
+        {
+            UUID existing_entity_id;
+            if (!insert(id, entity, replace_if_exists, throw_if_exists, &existing_entity_id))
+            {
+                /// Couldn't insert `entity` because there is an existing entity with the same name.
+                new_to_existing_ids[id] = existing_entity_id;
+            }
+        }
+
+        if (!new_to_existing_ids.empty())
+        {
+            /// If new entities restored from backup have dependencies on other entities from backup which were not restored because they existed,
+            /// then we should correct those dependencies.
+            auto update_func = [&](const AccessEntityPtr & entity) -> AccessEntityPtr
+            {
+                auto res = entity;
+                IAccessEntity::replaceDependencies(res, new_to_existing_ids);
+                return res;
+            };
+            std::vector<UUID> ids;
+            ids.reserve(entities_to_restore.size());
+            boost::copy(entities_to_restore | boost::adaptors::map_keys, std::back_inserter(ids));
+            tryUpdate(ids, update_func);
+        }
+    });
 }
 
 

src/Access/IAccessStorage.h

@@ -64,6 +64,9 @@ public:
     /// Returns true if this entity is readonly.
     virtual bool isReadOnly(const UUID &) const { return isReadOnly(); }
 
+    /// Returns true if this storage is replicated.
+    virtual bool isReplicated() const { return false; }
+
     /// Starts periodic reloading and updating of entities in this storage.
     virtual void startPeriodicReloading() {}
 
@@ -153,8 +156,8 @@ public:
     /// Inserts an entity to the storage. Returns ID of a new entry in the storage.
     /// Throws an exception if the specified name already exists.
     UUID insert(const AccessEntityPtr & entity);
-    std::optional<UUID> insert(const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists);
+    std::optional<UUID> insert(const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id = nullptr);
-    bool insert(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists);
+    bool insert(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id = nullptr);
     std::vector<UUID> insert(const std::vector<AccessEntityPtr> & multiple_entities, bool replace_if_exists = false, bool throw_if_exists = true);
     std::vector<UUID> insert(const std::vector<AccessEntityPtr> & multiple_entities, const std::vector<UUID> & ids, bool replace_if_exists = false, bool throw_if_exists = true);
 
@@ -218,7 +221,7 @@ protected:
     virtual std::vector<UUID> findAllImpl(AccessEntityType type) const = 0;
     virtual AccessEntityPtr readImpl(const UUID & id, bool throw_if_not_exists) const = 0;
     virtual std::optional<std::pair<String, AccessEntityType>> readNameWithTypeImpl(const UUID & id, bool throw_if_not_exists) const;
-    virtual bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists);
+    virtual bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id);
     virtual bool removeImpl(const UUID & id, bool throw_if_not_exists);
     virtual bool updateImpl(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists);
     virtual std::optional<AuthResult> authenticateImpl(
@@ -240,6 +243,7 @@ protected:
     LoggerPtr getLogger() const;
     static String formatEntityTypeWithName(AccessEntityType type, const String & name) { return AccessEntityTypeInfo::get(type).formatEntityNameWithType(name); }
     static void clearConflictsInEntitiesList(std::vector<std::pair<UUID, AccessEntityPtr>> & entities, LoggerPtr log_);
+    virtual bool acquireReplicatedRestore(RestorerFromBackup &) const { return false; }
     [[noreturn]] void throwNotFound(const UUID & id) const;
     [[noreturn]] void throwNotFound(AccessEntityType type, const String & name) const;
     [[noreturn]] static void throwBadCast(const UUID & id, AccessEntityType type, const String & name, AccessEntityType required_type);

src/Access/MemoryAccessStorage.cpp

@@ -1,7 +1,5 @@
 #include <Access/MemoryAccessStorage.h>
 #include <Access/AccessChangesNotifier.h>
-#include <Backups/RestorerFromBackup.h>
-#include <Backups/RestoreSettings.h>
 #include <base/scope_guard.h>
 #include <boost/container/flat_set.hpp>
 #include <boost/range/adaptor/map.hpp>
@@ -63,14 +61,14 @@ AccessEntityPtr MemoryAccessStorage::readImpl(const UUID & id, bool throw_if_not
 }
 
 
-bool MemoryAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists)
+bool MemoryAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
     std::lock_guard lock{mutex};
-    return insertNoLock(id, new_entity, replace_if_exists, throw_if_exists);
+    return insertNoLock(id, new_entity, replace_if_exists, throw_if_exists, conflicting_id);
 }
 
 
-bool MemoryAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists)
+bool MemoryAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
     const String & name = new_entity->getName();
     AccessEntityType type = new_entity->getType();
@@ -86,9 +84,15 @@ bool MemoryAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr &
     if (name_collision && !replace_if_exists)
     {
         if (throw_if_exists)
+        {
             throwNameCollisionCannotInsert(type, name);
+        }
         else
+        {
+            if (conflicting_id)
+                *conflicting_id = id_by_name;
             return false;
+        }
     }
 
     auto it_by_id = entries_by_id.find(id);
@@ -97,9 +101,15 @@ bool MemoryAccessStorage::insertNoLock(const UUID & id, const AccessEntityPtr &
     {
         const auto & existing_entry = it_by_id->second;
         if (throw_if_exists)
+        {
             throwIDCollisionCannotInsert(id, type, name, existing_entry.entity->getType(), existing_entry.entity->getName());
+        }
         else
+        {
+            if (conflicting_id)
+                *conflicting_id = id;
             return false;
+        }
     }
 
     /// Remove collisions if necessary.
@@ -270,28 +280,7 @@ void MemoryAccessStorage::setAll(const std::vector<std::pair<UUID, AccessEntityP
 
     /// Insert or update entities.
     for (const auto & [id, entity] : entities_without_conflicts)
-        insertNoLock(id, entity, /* replace_if_exists = */ true, /* throw_if_exists = */ false);
+        insertNoLock(id, entity, /* replace_if_exists = */ true, /* throw_if_exists = */ false, /* conflicting_id = */ nullptr);
-}
-
-
-void MemoryAccessStorage::restoreFromBackup(RestorerFromBackup & restorer)
-{
-    if (!isRestoreAllowed())
-        throwRestoreNotAllowed();
-
-    auto entities = restorer.getAccessEntitiesToRestore();
-    if (entities.empty())
-        return;
-
-    auto create_access = restorer.getRestoreSettings().create_access;
-    bool replace_if_exists = (create_access == RestoreAccessCreationMode::kReplace);
-    bool throw_if_exists = (create_access == RestoreAccessCreationMode::kCreate);
-
-    restorer.addDataRestoreTask([this, my_entities = std::move(entities), replace_if_exists, throw_if_exists]
-    {
-        for (const auto & [id, entity] : my_entities)
-            insert(id, entity, replace_if_exists, throw_if_exists);
-    });
 }
 
 }

src/Access/MemoryAccessStorage.h

@@ -34,17 +34,16 @@ public:
     bool exists(const UUID & id) const override;
 
     bool isBackupAllowed() const override { return backup_allowed; }
-    void restoreFromBackup(RestorerFromBackup & restorer) override;
 
 private:
     std::optional<UUID> findImpl(AccessEntityType type, const String & name) const override;
     std::vector<UUID> findAllImpl(AccessEntityType type) const override;
     AccessEntityPtr readImpl(const UUID & id, bool throw_if_not_exists) const override;
-    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists) override;
+    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id) override;
     bool removeImpl(const UUID & id, bool throw_if_not_exists) override;
     bool updateImpl(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists) override;
 
-    bool insertNoLock(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists);
+    bool insertNoLock(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id);
     bool removeNoLock(const UUID & id, bool throw_if_not_exists);
     bool updateNoLock(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists);
 

src/Access/MultipleAccessStorage.cpp

@@ -353,7 +353,7 @@ void MultipleAccessStorage::reload(ReloadMode reload_mode)
 }
 
 
-bool MultipleAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists)
+bool MultipleAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
     std::shared_ptr<IAccessStorage> storage_for_insertion;
 
@@ -376,7 +376,7 @@ bool MultipleAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr &
             getStorageName());
     }
 
-    if (storage_for_insertion->insert(id, entity, replace_if_exists, throw_if_exists))
+    if (storage_for_insertion->insert(id, entity, replace_if_exists, throw_if_exists, conflicting_id))
     {
         std::lock_guard lock{mutex};
         ids_cache.set(id, storage_for_insertion);

src/Access/MultipleAccessStorage.h

@@ -67,7 +67,7 @@ protected:
     std::vector<UUID> findAllImpl(AccessEntityType type) const override;
     AccessEntityPtr readImpl(const UUID & id, bool throw_if_not_exists) const override;
     std::optional<std::pair<String, AccessEntityType>> readNameWithTypeImpl(const UUID & id, bool throw_if_not_exists) const override;
-    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists) override;
+    bool insertImpl(const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id) override;
     bool removeImpl(const UUID & id, bool throw_if_not_exists) override;
     bool updateImpl(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists) override;
     std::optional<AuthResult> authenticateImpl(const Credentials & credentials, const Poco::Net::IPAddress & address, const ExternalAuthenticators & external_authenticators, bool throw_if_user_not_exists, bool allow_no_password, bool allow_plaintext_password) const override;

src/Access/Quota.cpp

@@ -24,7 +24,7 @@ std::vector<UUID> Quota::findDependencies() const
     return to_roles.findDependencies();
 }
 
-void Quota::replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
+void Quota::doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
 {
     to_roles.replaceDependencies(old_to_new_ids);
 }

src/Access/Quota.h

@@ -47,7 +47,7 @@ struct Quota : public IAccessEntity
     AccessEntityType getType() const override { return TYPE; }
 
     std::vector<UUID> findDependencies() const override;
-    void replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
+    void doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
     bool isBackupAllowed() const override { return true; }
 };
 

src/Access/ReplicatedAccessStorage.cpp

@@ -5,10 +5,9 @@
 #include <Access/AccessChangesNotifier.h>
 #include <Access/AccessBackup.h>
 #include <Backups/BackupEntriesCollector.h>
-#include <Backups/RestorerFromBackup.h>
-#include <Backups/RestoreSettings.h>
 #include <Backups/IBackupCoordination.h>
 #include <Backups/IRestoreCoordination.h>
+#include <Backups/RestorerFromBackup.h>
 #include <IO/ReadHelpers.h>
 #include <Interpreters/Context.h>
 #include <Common/ZooKeeper/KeeperException.h>
@@ -120,7 +119,7 @@ static void retryOnZooKeeperUserError(size_t attempts, Func && function)
     }
 }
 
-bool ReplicatedAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists)
+bool ReplicatedAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id)
 {
     const AccessEntityTypeInfo type_info = AccessEntityTypeInfo::get(new_entity->getType());
     const String & name = new_entity->getName();
@@ -128,7 +127,7 @@ bool ReplicatedAccessStorage::insertImpl(const UUID & id, const AccessEntityPtr
 
     auto zookeeper = getZooKeeper();
     bool ok = false;
-    retryOnZooKeeperUserError(10, [&]{ ok = insertZooKeeper(zookeeper, id, new_entity, replace_if_exists, throw_if_exists); });
+    retryOnZooKeeperUserError(10, [&]{ ok = insertZooKeeper(zookeeper, id, new_entity, replace_if_exists, throw_if_exists, conflicting_id); });
 
     if (!ok)
         return false;
@@ -143,7 +142,8 @@ bool ReplicatedAccessStorage::insertZooKeeper(
     const UUID & id,
     const AccessEntityPtr & new_entity,
     bool replace_if_exists,
-    bool throw_if_exists)
+    bool throw_if_exists,
+    UUID * conflicting_id)
 {
     const String & name = new_entity->getName();
     const AccessEntityType type = new_entity->getType();
@@ -167,27 +167,52 @@ bool ReplicatedAccessStorage::insertZooKeeper(
 
     if (res == Coordination::Error::ZNODEEXISTS)
     {
-        if (!throw_if_exists && !replace_if_exists)
+        if (!replace_if_exists)
-            return false; /// Couldn't insert a new entity.
-
-        if (throw_if_exists)
         {
             if (responses[0]->error == Coordination::Error::ZNODEEXISTS)
             {
-                /// To fail with a nice error message, we need info about what already exists.
+                /// Couldn't insert the new entity because there is an existing entity with such UUID.
-                /// This itself could fail if the conflicting uuid disappears in the meantime.
+                if (throw_if_exists)
-                /// If that happens, then we'll just retry from the start.
+                {
-                String existing_entity_definition = zookeeper->get(entity_path);
+                    /// To fail with a nice error message, we need info about what already exists.
+                    /// This itself can fail if the conflicting uuid disappears in the meantime.
+                    /// If that happens, then retryOnZooKeeperUserError() will just retry the operation from the start.
+                    String existing_entity_definition = zookeeper->get(entity_path);
 
-                AccessEntityPtr existing_entity = deserializeAccessEntity(existing_entity_definition, entity_path);
+                    AccessEntityPtr existing_entity = deserializeAccessEntity(existing_entity_definition, entity_path);
-                AccessEntityType existing_type = existing_entity->getType();
+                    AccessEntityType existing_type = existing_entity->getType();
-                String existing_name = existing_entity->getName();
+                    String existing_name = existing_entity->getName();
-                throwIDCollisionCannotInsert(id, type, name, existing_type, existing_name);
+                    throwIDCollisionCannotInsert(id, type, name, existing_type, existing_name);
+                }
+                else
+                {
+                    if (conflicting_id)
+                        *conflicting_id = id;
+                    return false;
+                }
+            }
+            else if (responses[1]->error == Coordination::Error::ZNODEEXISTS)
+            {
+                /// Couldn't insert the new entity because there is an existing entity with the same name.
+                if (throw_if_exists)
+                {
+                    throwNameCollisionCannotInsert(type, name);
+                }
+                else
+                {
+                    if (conflicting_id)
+                    {
+                        /// Get UUID of the existing entry with the same name.
+                        /// This itself can fail if the conflicting name disappears in the meantime.
+                        /// If that happens, then retryOnZooKeeperUserError() will just retry the operation from the start.
+                        *conflicting_id = parseUUID(zookeeper->get(name_path));
+                    }
+                    return false;
+                }
             }
             else
             {
-                /// Couldn't insert the new entity because there is an existing entity with such name.
+                zkutil::KeeperMultiException::check(res, ops, responses);
-                throwNameCollisionCannotInsert(type, name);
             }
         }
 
@@ -693,28 +718,10 @@ void ReplicatedAccessStorage::backup(BackupEntriesCollector & backup_entries_col
 }
 
 
-void ReplicatedAccessStorage::restoreFromBackup(RestorerFromBackup & restorer)
+bool ReplicatedAccessStorage::acquireReplicatedRestore(RestorerFromBackup & restorer) const
 {
-    if (!isRestoreAllowed())
-        throwRestoreNotAllowed();
-
     auto restore_coordination = restorer.getRestoreCoordination();
-    if (!restore_coordination->acquireReplicatedAccessStorage(zookeeper_path))
+    return restore_coordination->acquireReplicatedAccessStorage(zookeeper_path);
-        return;
-
-    auto entities = restorer.getAccessEntitiesToRestore();
-    if (entities.empty())
-        return;
-
-    auto create_access = restorer.getRestoreSettings().create_access;
-    bool replace_if_exists = (create_access == RestoreAccessCreationMode::kReplace);
-    bool throw_if_exists = (create_access == RestoreAccessCreationMode::kCreate);
-
-    restorer.addDataRestoreTask([this, my_entities = std::move(entities), replace_if_exists, throw_if_exists]
-    {
-        for (const auto & [id, entity] : my_entities)
-            insert(id, entity, replace_if_exists, throw_if_exists);
-    });
 }
 
 }

src/Access/ReplicatedAccessStorage.h

@@ -26,6 +26,7 @@ public:
     void shutdown() override;
 
     const char * getStorageType() const override { return STORAGE_TYPE; }
+    bool isReplicated() const override { return true; }
 
     void startPeriodicReloading() override { startWatchingThread(); }
     void stopPeriodicReloading() override { stopWatchingThread(); }
@@ -35,7 +36,6 @@ public:
 
     bool isBackupAllowed() const override { return backup_allowed; }
     void backup(BackupEntriesCollector & backup_entries_collector, const String & data_path_in_backup, AccessEntityType type) const override;
-    void restoreFromBackup(RestorerFromBackup & restorer) override;
 
 private:
     String zookeeper_path;
@@ -48,11 +48,11 @@ private:
     std::unique_ptr<ThreadFromGlobalPool> watching_thread;
     std::shared_ptr<ConcurrentBoundedQueue<UUID>> watched_queue;
 
-    bool insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists) override;
+    bool insertImpl(const UUID & id, const AccessEntityPtr & new_entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id) override;
     bool removeImpl(const UUID & id, bool throw_if_not_exists) override;
     bool updateImpl(const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists) override;
 
-    bool insertZooKeeper(const zkutil::ZooKeeperPtr & zookeeper, const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists);
+    bool insertZooKeeper(const zkutil::ZooKeeperPtr & zookeeper, const UUID & id, const AccessEntityPtr & entity, bool replace_if_exists, bool throw_if_exists, UUID * conflicting_id);
     bool removeZooKeeper(const zkutil::ZooKeeperPtr & zookeeper, const UUID & id, bool throw_if_not_exists);
     bool updateZooKeeper(const zkutil::ZooKeeperPtr & zookeeper, const UUID & id, const UpdateFunc & update_func, bool throw_if_not_exists);
 
@@ -80,6 +80,7 @@ private:
     std::optional<UUID> findImpl(AccessEntityType type, const String & name) const override;
     std::vector<UUID> findAllImpl(AccessEntityType type) const override;
     AccessEntityPtr readImpl(const UUID & id, bool throw_if_not_exists) const override;
+    bool acquireReplicatedRestore(RestorerFromBackup & restorer) const override;
 
     mutable std::mutex mutex;
     MemoryAccessStorage memory_storage TSA_GUARDED_BY(mutex);

src/Access/Role.cpp

@@ -21,7 +21,7 @@ std::vector<UUID> Role::findDependencies() const
     return res;
 }
 
-void Role::replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
+void Role::doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
 {
     granted_roles.replaceDependencies(old_to_new_ids);
     settings.replaceDependencies(old_to_new_ids);

src/Access/Role.h

@@ -21,7 +21,7 @@ struct Role : public IAccessEntity
     AccessEntityType getType() const override { return TYPE; }
 
     std::vector<UUID> findDependencies() const override;
-    void replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
+    void doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
     bool isBackupAllowed() const override { return settings.isBackupAllowed(); }
 };
 

src/Access/RowPolicy.cpp

@@ -63,7 +63,7 @@ std::vector<UUID> RowPolicy::findDependencies() const
     return to_roles.findDependencies();
 }
 
-void RowPolicy::replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
+void RowPolicy::doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
 {
     to_roles.replaceDependencies(old_to_new_ids);
 }

src/Access/RowPolicy.h

@@ -50,7 +50,7 @@ struct RowPolicy : public IAccessEntity
     AccessEntityType getType() const override { return TYPE; }
 
     std::vector<UUID> findDependencies() const override;
-    void replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
+    void doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
     bool isBackupAllowed() const override { return true; }
 
     /// Which roles or users should use this row policy.

src/Access/SettingsProfile.cpp

@@ -21,7 +21,7 @@ std::vector<UUID> SettingsProfile::findDependencies() const
     return res;
 }
 
-void SettingsProfile::replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
+void SettingsProfile::doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
 {
     elements.replaceDependencies(old_to_new_ids);
     to_roles.replaceDependencies(old_to_new_ids);

src/Access/SettingsProfile.h

@@ -22,7 +22,7 @@ struct SettingsProfile : public IAccessEntity
     AccessEntityType getType() const override { return TYPE; }
 
     std::vector<UUID> findDependencies() const override;
-    void replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
+    void doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
     bool isBackupAllowed() const override { return elements.isBackupAllowed(); }
 };
 

src/Access/User.cpp

@@ -49,7 +49,7 @@ std::vector<UUID> User::findDependencies() const
     return res;
 }
 
-void User::replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
+void User::doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids)
 {
     default_roles.replaceDependencies(old_to_new_ids);
     granted_roles.replaceDependencies(old_to_new_ids);

src/Access/User.h

@@ -32,7 +32,7 @@ struct User : public IAccessEntity
     void setName(const String & name_) override;
 
     std::vector<UUID> findDependencies() const override;
-    void replaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
+    void doReplaceDependencies(const std::unordered_map<UUID, UUID> & old_to_new_ids) override;
     bool isBackupAllowed() const override { return settings.isBackupAllowed(); }
 };
 

src/Analyzer/FunctionSecretArgumentsFinderTreeNode.h

@@ -3,370 +3,89 @@
 #include <Parsers/FunctionSecretArgumentsFinder.h>
 #include <Analyzer/ConstantNode.h>
 #include <Analyzer/FunctionNode.h>
-#include <Analyzer/IQueryTreeNode.h>
 #include <Analyzer/IdentifierNode.h>
-#include <Analyzer/ListNode.h>
-#include <Common/KnownObjectNames.h>
-#include <Core/QualifiedTableName.h>
-
-#include <boost/algorithm/string/predicate.hpp>
 
 
 namespace DB
 {
 
+class FunctionTreeNode : public AbstractFunction
+{
+public:
+    class ArgumentTreeNode : public Argument
+    {
+    public:
+        explicit ArgumentTreeNode(const IQueryTreeNode * argument_) : argument(argument_) {}
+        std::unique_ptr<AbstractFunction> getFunction() const override
+        {
+            if (const auto * f = argument->as<FunctionNode>())
+                return std::make_unique<FunctionTreeNode>(*f);
+            return nullptr;
+        }
+        bool isIdentifier() const override { return argument->as<IdentifierNode>(); }
+        bool tryGetString(String * res, bool allow_identifier) const override
+        {
+            if (const auto * literal = argument->as<ConstantNode>())
+            {
+                if (literal->getValue().getType() != Field::Types::String)
+                    return false;
+                if (res)
+                    *res = literal->getValue().safeGet<String>();
+                return true;
+            }
+
+            if (allow_identifier)
+            {
+                if (const auto * id = argument->as<IdentifierNode>())
+                {
+                    if (res)
+                        *res = id->getIdentifier().getFullName();
+                    return true;
+                }
+            }
+
+            return false;
+        }
+    private:
+        const IQueryTreeNode * argument = nullptr;
+    };
+
+    class ArgumentsTreeNode : public Arguments
+    {
+    public:
+        explicit ArgumentsTreeNode(const QueryTreeNodes * arguments_) : arguments(arguments_) {}
+        size_t size() const override { return arguments ? arguments->size() : 0; }
+        std::unique_ptr<Argument> at(size_t n) const override { return std::make_unique<ArgumentTreeNode>(arguments->at(n).get()); }
+    private:
+        const QueryTreeNodes * arguments = nullptr;
+    };
+
+    explicit FunctionTreeNode(const FunctionNode & function_) : function(&function_)
+    {
+        if (const auto & nodes = function->getArguments().getNodes(); !nodes.empty())
+            arguments = std::make_unique<ArgumentsTreeNode>(&nodes);
+    }
+    String name() const override { return function->getFunctionName(); }
+private:
+    const FunctionNode * function = nullptr;
+};
+
 
 /// Finds arguments of a specified function which should not be displayed for most users for security reasons.
 /// That involves passwords and secret keys.
-class FunctionSecretArgumentsFinderTreeNode
+class FunctionSecretArgumentsFinderTreeNode : public FunctionSecretArgumentsFinder
 {
 public:
-    explicit FunctionSecretArgumentsFinderTreeNode(const FunctionNode & function_) : function(function_), arguments(function.getArguments())
+    explicit FunctionSecretArgumentsFinderTreeNode(const FunctionNode & function_)
+        : FunctionSecretArgumentsFinder(std::make_unique<FunctionTreeNode>(function_))
     {
-        if (arguments.getNodes().empty())
+        if (!function->hasArguments())
             return;
 
-        findFunctionSecretArguments();
+        findOrdinaryFunctionSecretArguments();
     }
 
-    struct Result
-    {
-        /// Result constructed by default means no arguments will be hidden.
-        size_t start = static_cast<size_t>(-1);
-        size_t count = 0; /// Mostly it's either 0 or 1. There are only a few cases where `count` can be greater than 1 (e.g. see `encrypt`).
-                            /// In all known cases secret arguments are consecutive
-        bool are_named = false; /// Arguments like `password = 'password'` are considered as named arguments.
-        /// E.g. "headers" in `url('..', headers('foo' = '[HIDDEN]'))`
-        std::vector<std::string> nested_maps;
-
-        bool hasSecrets() const
-        {
-            return count != 0 || !nested_maps.empty();
-        }
-    };
-
     FunctionSecretArgumentsFinder::Result getResult() const { return result; }
-
-private:
-    const FunctionNode & function;
-    const ListNode & arguments;
-    FunctionSecretArgumentsFinder::Result result;
-
-    void markSecretArgument(size_t index, bool argument_is_named = false)
-    {
-        if (index >= arguments.getNodes().size())
-            return;
-        if (!result.count)
-        {
-            result.start = index;
-            result.are_named = argument_is_named;
-        }
-        chassert(index >= result.start); /// We always check arguments consecutively
-        result.count = index + 1 - result.start;
-        if (!argument_is_named)
-            result.are_named = false;
-    }
-
-    void findFunctionSecretArguments()
-    {
-        const auto & name = function.getFunctionName();
-
-        if ((name == "mysql") || (name == "postgresql") || (name == "mongodb"))
-        {
-            /// mysql('host:port', 'database', 'table', 'user', 'password', ...)
-            /// postgresql('host:port', 'database', 'table', 'user', 'password', ...)
-            /// mongodb('host:port', 'database', 'collection', 'user', 'password', ...)
-            findMySQLFunctionSecretArguments();
-        }
-        else if ((name == "s3") || (name == "cosn") || (name == "oss") ||
-                    (name == "deltaLake") || (name == "hudi") || (name == "iceberg"))
-        {
-            /// s3('url', 'aws_access_key_id', 'aws_secret_access_key', ...)
-            findS3FunctionSecretArguments(/* is_cluster_function= */ false);
-        }
-        else if (name == "s3Cluster")
-        {
-            /// s3Cluster('cluster_name', 'url', 'aws_access_key_id', 'aws_secret_access_key', ...)
-            findS3FunctionSecretArguments(/* is_cluster_function= */ true);
-        }
-        else if ((name == "remote") || (name == "remoteSecure"))
-        {
-            /// remote('addresses_expr', 'db', 'table', 'user', 'password', ...)
-            findRemoteFunctionSecretArguments();
-        }
-        else if ((name == "encrypt") || (name == "decrypt") ||
-                    (name == "aes_encrypt_mysql") || (name == "aes_decrypt_mysql") ||
-                    (name == "tryDecrypt"))
-        {
-            /// encrypt('mode', 'plaintext', 'key' [, iv, aad])
-            findEncryptionFunctionSecretArguments();
-        }
-        else if (name == "url")
-        {
-            findURLSecretArguments();
-        }
-    }
-
-    void findMySQLFunctionSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// mysql(named_collection, ..., password = 'password', ...)
-            findSecretNamedArgument("password", 1);
-        }
-        else
-        {
-            /// mysql('host:port', 'database', 'table', 'user', 'password', ...)
-            markSecretArgument(4);
-        }
-    }
-
-    /// Returns the number of arguments excluding "headers" and "extra_credentials" (which should
-    /// always be at the end). Marks "headers" as secret, if found.
-    size_t excludeS3OrURLNestedMaps()
-    {
-        const auto & nodes = arguments.getNodes();
-        size_t count = nodes.size();
-        while (count > 0)
-        {
-            const FunctionNode * f = nodes.at(count - 1)->as<FunctionNode>();
-            if (!f)
-                break;
-            if (f->getFunctionName() == "headers")
-                result.nested_maps.push_back(f->getFunctionName());
-            else if (f->getFunctionName() != "extra_credentials")
-                break;
-            count -= 1;
-        }
-        return count;
-    }
-
-    void findS3FunctionSecretArguments(bool is_cluster_function)
-    {
-        /// s3Cluster('cluster_name', 'url', ...) has 'url' as its second argument.
-        size_t url_arg_idx = is_cluster_function ? 1 : 0;
-
-        if (!is_cluster_function && isNamedCollectionName(0))
-        {
-            /// s3(named_collection, ..., secret_access_key = 'secret_access_key', ...)
-            findSecretNamedArgument("secret_access_key", 1);
-            return;
-        }
-
-        /// We should check other arguments first because we don't need to do any replacement in case of
-        /// s3('url', NOSIGN, 'format' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
-        /// s3('url', 'format', 'structure' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
-        size_t count = excludeS3OrURLNestedMaps();
-        if ((url_arg_idx + 3 <= count) && (count <= url_arg_idx + 4))
-        {
-            String second_arg;
-            if (tryGetStringFromArgument(url_arg_idx + 1, &second_arg))
-            {
-                if (boost::iequals(second_arg, "NOSIGN"))
-                    return; /// The argument after 'url' is "NOSIGN".
-
-                if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
-                    return; /// The argument after 'url' is a format: s3('url', 'format', ...)
-            }
-        }
-
-        /// We're going to replace 'aws_secret_access_key' with '[HIDDEN]' for the following signatures:
-        /// s3('url', 'aws_access_key_id', 'aws_secret_access_key', ...)
-        /// s3Cluster('cluster_name', 'url', 'aws_access_key_id', 'aws_secret_access_key', 'format', 'compression')
-        if (url_arg_idx + 2 < count)
-            markSecretArgument(url_arg_idx + 2);
-    }
-
-    void findURLSecretArguments()
-    {
-        if (!isNamedCollectionName(0))
-            excludeS3OrURLNestedMaps();
-    }
-
-    bool tryGetStringFromArgument(size_t arg_idx, String * res, bool allow_identifier = true) const
-    {
-        if (arg_idx >= arguments.getNodes().size())
-            return false;
-
-        return tryGetStringFromArgument(arguments.getNodes()[arg_idx], res, allow_identifier);
-    }
-
-    static bool tryGetStringFromArgument(const QueryTreeNodePtr argument, String * res, bool allow_identifier = true)
-    {
-        if (const auto * literal = argument->as<ConstantNode>())
-        {
-            if (literal->getValue().getType() != Field::Types::String)
-                return false;
-            if (res)
-                *res = literal->getValue().safeGet<String>();
-            return true;
-        }
-
-        if (allow_identifier)
-        {
-            if (const auto * id = argument->as<IdentifierNode>())
-            {
-                if (res)
-                    *res = id->getIdentifier().getFullName();
-                return true;
-            }
-        }
-
-        return false;
-    }
-
-    void findRemoteFunctionSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// remote(named_collection, ..., password = 'password', ...)
-            findSecretNamedArgument("password", 1);
-            return;
-        }
-
-        /// We're going to replace 'password' with '[HIDDEN'] for the following signatures:
-        /// remote('addresses_expr', db.table, 'user' [, 'password'] [, sharding_key])
-        /// remote('addresses_expr', 'db', 'table', 'user' [, 'password'] [, sharding_key])
-        /// remote('addresses_expr', table_function(), 'user' [, 'password'] [, sharding_key])
-
-        /// But we should check the number of arguments first because we don't need to do any replacements in case of
-        /// remote('addresses_expr', db.table)
-        if (arguments.getNodes().size() < 3)
-            return;
-
-        size_t arg_num = 1;
-
-        /// Skip 1 or 2 arguments with table_function() or db.table or 'db', 'table'.
-        const auto * table_function = arguments.getNodes()[arg_num]->as<FunctionNode>();
-        if (table_function && KnownTableFunctionNames::instance().exists(table_function->getFunctionName()))
-        {
-            ++arg_num;
-        }
-        else
-        {
-            std::optional<String> database;
-            std::optional<QualifiedTableName> qualified_table_name;
-            if (!tryGetDatabaseNameOrQualifiedTableName(arg_num, database, qualified_table_name))
-            {
-                /// We couldn't evaluate the argument so we don't know whether it is 'db.table' or just 'db'.
-                /// Hence we can't figure out whether we should skip one argument 'user' or two arguments 'table', 'user'
-                /// before the argument 'password'. So it's safer to wipe two arguments just in case.
-                /// The last argument can be also a `sharding_key`, so we need to check that argument is a literal string
-                /// before wiping it (because the `password` argument is always a literal string).
-                if (tryGetStringFromArgument(arg_num + 2, nullptr, /* allow_identifier= */ false))
-                {
-                    /// Wipe either `password` or `user`.
-                    markSecretArgument(arg_num + 2);
-                }
-                if (tryGetStringFromArgument(arg_num + 3, nullptr, /* allow_identifier= */ false))
-                {
-                    /// Wipe either `password` or `sharding_key`.
-                    markSecretArgument(arg_num + 3);
-                }
-                return;
-            }
-
-            /// Skip the current argument (which is either a database name or a qualified table name).
-            ++arg_num;
-            if (database)
-            {
-                /// Skip the 'table' argument if the previous argument was a database name.
-                ++arg_num;
-            }
-        }
-
-        /// Skip username.
-        ++arg_num;
-
-        /// Do our replacement:
-        /// remote('addresses_expr', db.table, 'user', 'password', ...) -> remote('addresses_expr', db.table, 'user', '[HIDDEN]', ...)
-        /// The last argument can be also a `sharding_key`, so we need to check that argument is a literal string
-        /// before wiping it (because the `password` argument is always a literal string).
-        bool can_be_password = tryGetStringFromArgument(arg_num, nullptr, /* allow_identifier= */ false);
-        if (can_be_password)
-            markSecretArgument(arg_num);
-    }
-
-    /// Tries to get either a database name or a qualified table name from an argument.
-    /// Empty string is also allowed (it means the default database).
-    /// The function is used by findRemoteFunctionSecretArguments() to determine how many arguments to skip before a password.
-    bool tryGetDatabaseNameOrQualifiedTableName(
-        size_t arg_idx,
-        std::optional<String> & res_database,
-        std::optional<QualifiedTableName> & res_qualified_table_name) const
-    {
-        res_database.reset();
-        res_qualified_table_name.reset();
-
-        String str;
-        if (!tryGetStringFromArgument(arg_idx, &str, /* allow_identifier= */ true))
-            return false;
-
-        if (str.empty())
-        {
-            res_database = "";
-            return true;
-        }
-
-        auto qualified_table_name = QualifiedTableName::tryParseFromString(str);
-        if (!qualified_table_name)
-            return false;
-
-        if (qualified_table_name->database.empty())
-            res_database = std::move(qualified_table_name->table);
-        else
-            res_qualified_table_name = std::move(qualified_table_name);
-        return true;
-    }
-
-    void findEncryptionFunctionSecretArguments()
-    {
-        if (arguments.getNodes().empty())
-            return;
-
-        /// We replace all arguments after 'mode' with '[HIDDEN]':
-        /// encrypt('mode', 'plaintext', 'key' [, iv, aad]) -> encrypt('mode', '[HIDDEN]')
-        result.start = 1;
-        result.count = arguments.getNodes().size() - 1;
-    }
-
-
-    /// Whether a specified argument can be the name of a named collection?
-    bool isNamedCollectionName(size_t arg_idx) const
-    {
-        if (arguments.getNodes().size() <= arg_idx)
-            return false;
-
-        const auto * identifier = arguments.getNodes()[arg_idx]->as<IdentifierNode>();
-        return identifier != nullptr;
-    }
-
-    /// Looks for a secret argument with a specified name. This function looks for arguments in format `key=value` where the key is specified.
-    void findSecretNamedArgument(const std::string_view & key, size_t start = 0)
-    {
-        for (size_t i = start; i < arguments.getNodes().size(); ++i)
-        {
-            const auto & argument = arguments.getNodes()[i];
-            const auto * equals_func = argument->as<FunctionNode>();
-            if (!equals_func || (equals_func->getFunctionName() != "equals"))
-                continue;
-
-            const auto * expr_list = equals_func->getArguments().as<ListNode>();
-            if (!expr_list)
-                continue;
-
-            const auto & equal_args = expr_list->getNodes();
-            if (equal_args.size() != 2)
-                continue;
-
-            String found_key;
-            if (!tryGetStringFromArgument(equal_args[0], &found_key))
-                continue;
-
-            if (found_key == key)
-                markSecretArgument(i, /* argument_is_named= */ true);
-        }
-    }
 };
 
 }

src/Analyzer/Resolve/QueryAnalyzer.cpp

@@ -2564,8 +2564,8 @@ void checkFunctionNodeHasEmptyNullsAction(FunctionNode const & node)
     if (node.getNullsAction() != NullsAction::EMPTY)
         throw Exception(
             ErrorCodes::SYNTAX_ERROR,
-            "Function with name '{}' cannot use {} NULLS",
+            "Function with name {} cannot use {} NULLS",
-            node.getFunctionName(),
+            backQuote(node.getFunctionName()),
             node.getNullsAction() == NullsAction::IGNORE_NULLS ? "IGNORE" : "RESPECT");
 }
 }
@@ -3228,16 +3228,16 @@ ProjectionNames QueryAnalyzer::resolveFunction(QueryTreeNodePtr & node, Identifi
             auto hints = NamePrompter<2>::getHints(function_name, possible_function_names);
 
             throw Exception(ErrorCodes::UNKNOWN_FUNCTION,
-                "Function with name '{}' does not exist. In scope {}{}",
+                "Function with name {} does not exist. In scope {}{}",
-                function_name,
+                backQuote(function_name),
                 scope.scope_node->formatASTForErrorMessage(),
                 getHintsErrorMessageSuffix(hints));
         }
 
         if (!function_lambda_arguments_indexes.empty())
             throw Exception(ErrorCodes::UNSUPPORTED_METHOD,
-                "Aggregate function '{}' does not support lambda arguments",
+                "Aggregate function {} does not support lambda arguments",
-                function_name);
+                backQuote(function_name));
 
         auto action = function_node_ptr->getNullsAction();
         std::string aggregate_function_name = rewriteAggregateFunctionNameIfNeeded(function_name, action, scope.context);
@@ -3674,10 +3674,10 @@ ProjectionNames QueryAnalyzer::resolveExpressionNode(
 
                 auto hints = IdentifierResolver::collectIdentifierTypoHints(unresolved_identifier, valid_identifiers);
 
-                throw Exception(ErrorCodes::UNKNOWN_IDENTIFIER, "Unknown {}{} identifier '{}' in scope {}{}",
+                throw Exception(ErrorCodes::UNKNOWN_IDENTIFIER, "Unknown {}{} identifier {} in scope {}{}",
                     toStringLowercase(IdentifierLookupContext::EXPRESSION),
                     message_clarification,
-                    unresolved_identifier.getFullName(),
+                    backQuote(unresolved_identifier.getFullName()),
                     scope.scope_node->formatASTForErrorMessage(),
                     getHintsErrorMessageSuffix(hints));
             }

src/Client/Connection.cpp

@@ -456,6 +456,9 @@ void Connection::sendAddendum()
         writeStringBinary(proto_recv_chunked, *out);
     }
 
+    if (server_revision >= DBMS_MIN_REVISION_WITH_VERSIONED_PARALLEL_REPLICAS_PROTOCOL)
+        writeVarUInt(DBMS_PARALLEL_REPLICAS_PROTOCOL_VERSION, *out);
+
     out->next();
 }
 
@@ -526,6 +529,8 @@ void Connection::receiveHello(const Poco::Timespan & handshake_timeout)
         readVarUInt(server_version_major, *in);
         readVarUInt(server_version_minor, *in);
         readVarUInt(server_revision, *in);
+        if (server_revision >= DBMS_MIN_REVISION_WITH_VERSIONED_PARALLEL_REPLICAS_PROTOCOL)
+            readVarUInt(server_parallel_replicas_protocol_version, *in);
         if (server_revision >= DBMS_MIN_REVISION_WITH_SERVER_TIMEZONE)
             readStringBinary(server_timezone, *in);
         if (server_revision >= DBMS_MIN_REVISION_WITH_SERVER_DISPLAY_NAME)
@@ -966,7 +971,7 @@ void Connection::sendReadTaskResponse(const String & response)
 void Connection::sendMergeTreeReadTaskResponse(const ParallelReadResponse & response)
 {
     writeVarUInt(Protocol::Client::MergeTreeReadTaskResponse, *out);
-    response.serialize(*out);
+    response.serialize(*out, server_parallel_replicas_protocol_version);
     out->finishChunk();
     out->next();
 }
@@ -1420,7 +1425,7 @@ ParallelReadRequest Connection::receiveParallelReadRequest() const
 
 InitialAllRangesAnnouncement Connection::receiveInitialParallelReadAnnouncement() const
 {
-    return InitialAllRangesAnnouncement::deserialize(*in);
+    return InitialAllRangesAnnouncement::deserialize(*in, server_parallel_replicas_protocol_version);
 }
 
 

src/Client/Connection.h

@@ -212,6 +212,7 @@ private:
     UInt64 server_version_minor = 0;
     UInt64 server_version_patch = 0;
     UInt64 server_revision = 0;
+    UInt64 server_parallel_replicas_protocol_version = 0;
     String server_timezone;
     String server_display_name;
 

src/Client/MultiplexedConnections.cpp

@@ -200,7 +200,6 @@ void MultiplexedConnections::sendQuery(
 }
 
 
-
 void MultiplexedConnections::sendIgnoredPartUUIDs(const std::vector<UUID> & uuids)
 {
     std::lock_guard lock(cancel_mutex);

src/Common/FailPoint.cpp

@@ -64,6 +64,7 @@ static struct InitFiu
     REGULAR(lazy_pipe_fds_fail_close) \
     PAUSEABLE(infinite_sleep) \
     PAUSEABLE(stop_moving_part_before_swap_with_active) \
+    REGULAR(slowdown_index_analysis) \
 
 
 namespace FailPoints

src/Common/ProfileEvents.cpp

@@ -376,6 +376,7 @@ The server successfully detected this situation and will download merged part fr
     M(ParallelReplicasReadAssignedMarks, "Sum across all replicas of how many of scheduled marks were assigned by consistent hash") \
     M(ParallelReplicasReadUnassignedMarks, "Sum across all replicas of how many unassigned marks were scheduled") \
     M(ParallelReplicasReadAssignedForStealingMarks, "Sum across all replicas of how many of scheduled marks were assigned for stealing by consistent hash") \
+    M(ParallelReplicasReadMarks, "How many marks were read by the given replica") \
     \
     M(ParallelReplicasStealingByHashMicroseconds, "Time spent collecting segments meant for stealing by hash") \
     M(ParallelReplicasProcessingPartsMicroseconds, "Time spent processing data parts") \
@@ -529,6 +530,7 @@ The server successfully detected this situation and will download merged part fr
     M(CachedReadBufferReadFromCacheMicroseconds, "Time reading from filesystem cache") \
     M(CachedReadBufferReadFromSourceBytes, "Bytes read from filesystem cache source (from remote fs, etc)") \
     M(CachedReadBufferReadFromCacheBytes, "Bytes read from filesystem cache") \
+    M(CachedReadBufferPredownloadedBytes, "Bytes read from filesystem cache source. Cache segments are read from left to right as a whole, it might be that we need to predownload some part of the segment irrelevant for the current task just to get to the needed data") \
     M(CachedReadBufferCacheWriteBytes, "Bytes written from source (remote fs, etc) to filesystem cache") \
     M(CachedReadBufferCacheWriteMicroseconds, "Time spent writing data into filesystem cache") \
     M(CachedReadBufferCreateBufferMicroseconds, "Prepare buffer time") \

src/Common/ZooKeeper/IKeeper.cpp

@@ -181,12 +181,6 @@ void SetACLRequest::addRootPath(const String & root_path) { Coordination::addRoo
 void GetACLRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
 void SyncRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
 
-void MultiRequest::addRootPath(const String & root_path)
-{
-    for (auto & request : requests)
-        request->addRootPath(root_path);
-}
-
 void CreateResponse::removeRootPath(const String & root_path) { Coordination::removeRootPath(path_created, root_path); }
 void WatchResponse::removeRootPath(const String & root_path) { Coordination::removeRootPath(path, root_path); }
 

src/Common/ZooKeeper/IKeeper.h

@@ -408,11 +408,17 @@ struct ReconfigResponse : virtual Response
     size_t bytesSize() const override { return value.size() + sizeof(stat); }
 };
 
+template <typename T>
 struct MultiRequest : virtual Request
 {
-    Requests requests;
+    std::vector<T> requests;
+
+    void addRootPath(const String & root_path) override
+    {
+        for (auto & request : requests)
+            request->addRootPath(root_path);
+    }
 
-    void addRootPath(const String & root_path) override;
     String getPath() const override { return {}; }
 
     size_t bytesSize() const override

src/Common/ZooKeeper/TestKeeper.cpp

@@ -184,7 +184,7 @@ struct TestKeeperReconfigRequest final : ReconfigRequest, TestKeeperRequest
     std::pair<ResponsePtr, Undo> process(TestKeeper::Container & container, int64_t zxid) const override;
 };
 
-struct TestKeeperMultiRequest final : MultiRequest, TestKeeperRequest
+struct TestKeeperMultiRequest final : MultiRequest<RequestPtr>, TestKeeperRequest
 {
     explicit TestKeeperMultiRequest(const Requests & generic_requests)
         : TestKeeperMultiRequest(std::span(generic_requests))

src/Common/ZooKeeper/ZooKeeperCommon.cpp

@@ -18,14 +18,16 @@ using namespace DB;
 
 void ZooKeeperResponse::write(WriteBuffer & out) const
 {
-    /// Excessive copy to calculate length.
+    auto response_size = Coordination::size(xid) + Coordination::size(zxid) + Coordination::size(error);
-    WriteBufferFromOwnString buf;
-    Coordination::write(xid, buf);
-    Coordination::write(zxid, buf);
-    Coordination::write(error, buf);
     if (error == Error::ZOK)
-        writeImpl(buf);
+        response_size += sizeImpl();
-    Coordination::write(buf.str(), out);
+
+    Coordination::write(static_cast<int32_t>(response_size), out);
+    Coordination::write(xid, out);
+    Coordination::write(zxid, out);
+    Coordination::write(error, out);
+    if (error == Error::ZOK)
+        writeImpl(out);
 }
 
 std::string ZooKeeperRequest::toString(bool short_format) const
@@ -41,12 +43,12 @@ std::string ZooKeeperRequest::toString(bool short_format) const
 
 void ZooKeeperRequest::write(WriteBuffer & out) const
 {
-    /// Excessive copy to calculate length.
+    auto request_size = Coordination::size(xid) + Coordination::size(getOpNum()) + sizeImpl();
-    WriteBufferFromOwnString buf;
+
-    Coordination::write(xid, buf);
+    Coordination::write(static_cast<int32_t>(request_size), out);
-    Coordination::write(getOpNum(), buf);
+    Coordination::write(xid, out);
-    writeImpl(buf);
+    Coordination::write(getOpNum(), out);
-    Coordination::write(buf.str(), out);
+    writeImpl(out);
 }
 
 void ZooKeeperSyncRequest::writeImpl(WriteBuffer & out) const
@@ -54,6 +56,11 @@ void ZooKeeperSyncRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(path, out);
 }
 
+size_t ZooKeeperSyncRequest::sizeImpl() const
+{
+    return Coordination::size(path);
+}
+
 void ZooKeeperSyncRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -74,6 +81,11 @@ void ZooKeeperSyncResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(path, out);
 }
 
+size_t ZooKeeperSyncResponse::sizeImpl() const
+{
+    return Coordination::size(path);
+}
+
 void ZooKeeperReconfigRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(joining, out);
@@ -82,6 +94,11 @@ void ZooKeeperReconfigRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(version, out);
 }
 
+size_t ZooKeeperReconfigRequest::sizeImpl() const
+{
+    return Coordination::size(joining) + Coordination::size(leaving) + Coordination::size(new_members) + Coordination::size(version);
+}
+
 void ZooKeeperReconfigRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(joining, in);
@@ -109,6 +126,11 @@ void ZooKeeperReconfigResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperReconfigResponse::sizeImpl() const
+{
+    return Coordination::size(value) + Coordination::size(stat);
+}
+
 void ZooKeeperWatchResponse::readImpl(ReadBuffer & in)
 {
     Coordination::read(type, in);
@@ -123,6 +145,11 @@ void ZooKeeperWatchResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(path, out);
 }
 
+size_t ZooKeeperWatchResponse::sizeImpl() const
+{
+    return Coordination::size(type) + Coordination::size(state) + Coordination::size(path);
+}
+
 void ZooKeeperWatchResponse::write(WriteBuffer & out) const
 {
     if (error == Error::ZOK)
@@ -137,6 +164,11 @@ void ZooKeeperAuthRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(data, out);
 }
 
+size_t ZooKeeperAuthRequest::sizeImpl() const
+{
+    return Coordination::size(type) + Coordination::size(scheme) + Coordination::size(data);
+}
+
 void ZooKeeperAuthRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(type, in);
@@ -175,6 +207,12 @@ void ZooKeeperCreateRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(flags, out);
 }
 
+size_t ZooKeeperCreateRequest::sizeImpl() const
+{
+    int32_t flags = 0;
+    return Coordination::size(path) + Coordination::size(data) + Coordination::size(acls) + Coordination::size(flags);
+}
+
 void ZooKeeperCreateRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -211,12 +249,22 @@ void ZooKeeperCreateResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(path_created, out);
 }
 
+size_t ZooKeeperCreateResponse::sizeImpl() const
+{
+    return Coordination::size(path_created);
+}
+
 void ZooKeeperRemoveRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(path, out);
     Coordination::write(version, out);
 }
 
+size_t ZooKeeperRemoveRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(version);
+}
+
 std::string ZooKeeperRemoveRequest::toStringImpl(bool /*short_format*/) const
 {
     return fmt::format(
@@ -244,6 +292,11 @@ void ZooKeeperRemoveRecursiveRequest::readImpl(ReadBuffer & in)
     Coordination::read(remove_nodes_limit, in);
 }
 
+size_t ZooKeeperRemoveRecursiveRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(remove_nodes_limit);
+}
+
 std::string ZooKeeperRemoveRecursiveRequest::toStringImpl(bool /*short_format*/) const
 {
     return fmt::format(
@@ -259,6 +312,11 @@ void ZooKeeperExistsRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(has_watch, out);
 }
 
+size_t ZooKeeperExistsRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(has_watch);
+}
+
 void ZooKeeperExistsRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -280,12 +338,22 @@ void ZooKeeperExistsResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperExistsResponse::sizeImpl() const
+{
+    return Coordination::size(stat);
+}
+
 void ZooKeeperGetRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(path, out);
     Coordination::write(has_watch, out);
 }
 
+size_t ZooKeeperGetRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(has_watch);
+}
+
 void ZooKeeperGetRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -309,6 +377,11 @@ void ZooKeeperGetResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperGetResponse::sizeImpl() const
+{
+    return Coordination::size(data) + Coordination::size(stat);
+}
+
 void ZooKeeperSetRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(path, out);
@@ -316,6 +389,11 @@ void ZooKeeperSetRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(version, out);
 }
 
+size_t ZooKeeperSetRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(data) + Coordination::size(version);
+}
+
 void ZooKeeperSetRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -342,12 +420,22 @@ void ZooKeeperSetResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperSetResponse::sizeImpl() const
+{
+    return Coordination::size(stat);
+}
+
 void ZooKeeperListRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(path, out);
     Coordination::write(has_watch, out);
 }
 
+size_t ZooKeeperListRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(has_watch);
+}
+
 void ZooKeeperListRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -366,6 +454,11 @@ void ZooKeeperFilteredListRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(static_cast<uint8_t>(list_request_type), out);
 }
 
+size_t ZooKeeperFilteredListRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(has_watch) + Coordination::size(static_cast<uint8_t>(list_request_type));
+}
+
 void ZooKeeperFilteredListRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -397,6 +490,11 @@ void ZooKeeperListResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperListResponse::sizeImpl() const
+{
+    return Coordination::size(names) + Coordination::size(stat);
+}
+
 void ZooKeeperSimpleListResponse::readImpl(ReadBuffer & in)
 {
     Coordination::read(names, in);
@@ -407,6 +505,11 @@ void ZooKeeperSimpleListResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(names, out);
 }
 
+size_t ZooKeeperSimpleListResponse::sizeImpl() const
+{
+    return Coordination::size(names);
+}
+
 void ZooKeeperSetACLRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(path, out);
@@ -414,6 +517,11 @@ void ZooKeeperSetACLRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(version, out);
 }
 
+size_t ZooKeeperSetACLRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(acls) + Coordination::size(version);
+}
+
 void ZooKeeperSetACLRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -431,6 +539,11 @@ void ZooKeeperSetACLResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperSetACLResponse::sizeImpl() const
+{
+    return Coordination::size(stat);
+}
+
 void ZooKeeperSetACLResponse::readImpl(ReadBuffer & in)
 {
     Coordination::read(stat, in);
@@ -446,6 +559,11 @@ void ZooKeeperGetACLRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(path, out);
 }
 
+size_t ZooKeeperGetACLRequest::sizeImpl() const
+{
+    return Coordination::size(path);
+}
+
 std::string ZooKeeperGetACLRequest::toStringImpl(bool /*short_format*/) const
 {
     return fmt::format("path = {}", path);
@@ -457,6 +575,11 @@ void ZooKeeperGetACLResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(stat, out);
 }
 
+size_t ZooKeeperGetACLResponse::sizeImpl() const
+{
+    return Coordination::size(acl) + Coordination::size(stat);
+}
+
 void ZooKeeperGetACLResponse::readImpl(ReadBuffer & in)
 {
     Coordination::read(acl, in);
@@ -469,6 +592,11 @@ void ZooKeeperCheckRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(version, out);
 }
 
+size_t ZooKeeperCheckRequest::sizeImpl() const
+{
+    return Coordination::size(path) + Coordination::size(version);
+}
+
 void ZooKeeperCheckRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(path, in);
@@ -494,6 +622,11 @@ void ZooKeeperErrorResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(error, out);
 }
 
+size_t ZooKeeperErrorResponse::sizeImpl() const
+{
+    return Coordination::size(error);
+}
+
 void ZooKeeperMultiRequest::checkOperationType(OperationType type)
 {
     chassert(!operation_type.has_value() || *operation_type == type);
@@ -596,6 +729,27 @@ void ZooKeeperMultiRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(error, out);
 }
 
+size_t ZooKeeperMultiRequest::sizeImpl() const
+{
+    size_t total_size = 0;
+    for (const auto & request : requests)
+    {
+        const auto & zk_request = dynamic_cast<const ZooKeeperRequest &>(*request);
+
+        bool done = false;
+        int32_t error = -1;
+
+        total_size
+            += Coordination::size(zk_request.getOpNum()) + Coordination::size(done) + Coordination::size(error) + zk_request.sizeImpl();
+    }
+
+    OpNum op_num = OpNum::Error;
+    bool done = true;
+    int32_t error = -1;
+
+    return total_size + Coordination::size(op_num) + Coordination::size(done) + Coordination::size(error);
+}
+
 void ZooKeeperMultiRequest::readImpl(ReadBuffer & in)
 {
     while (true)
@@ -729,31 +883,54 @@ void ZooKeeperMultiResponse::writeImpl(WriteBuffer & out) const
     }
 }
 
-ZooKeeperResponsePtr ZooKeeperHeartbeatRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperHeartbeatResponse>()); }
+size_t ZooKeeperMultiResponse::sizeImpl() const
-ZooKeeperResponsePtr ZooKeeperSyncRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperSyncResponse>()); }
+{
-ZooKeeperResponsePtr ZooKeeperAuthRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperAuthResponse>()); }
+    size_t total_size = 0;
-ZooKeeperResponsePtr ZooKeeperRemoveRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperRemoveResponse>()); }
+    for (const auto & response : responses)
-ZooKeeperResponsePtr ZooKeeperRemoveRecursiveRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperRemoveRecursiveResponse>()); }
+    {
-ZooKeeperResponsePtr ZooKeeperExistsRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperExistsResponse>()); }
+        const ZooKeeperResponse & zk_response = dynamic_cast<const ZooKeeperResponse &>(*response);
-ZooKeeperResponsePtr ZooKeeperGetRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperGetResponse>()); }
+        OpNum op_num = zk_response.getOpNum();
-ZooKeeperResponsePtr ZooKeeperSetRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperSetResponse>()); }
+        bool done = false;
-ZooKeeperResponsePtr ZooKeeperReconfigRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperReconfigResponse>()); }
+        Error op_error = zk_response.error;
-ZooKeeperResponsePtr ZooKeeperListRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperListResponse>()); }
+
-ZooKeeperResponsePtr ZooKeeperSimpleListRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperSimpleListResponse>()); }
+        total_size += Coordination::size(op_num) + Coordination::size(done) + Coordination::size(op_error);
+        if (op_error == Error::ZOK || op_num == OpNum::Error)
+            total_size += zk_response.sizeImpl();
+    }
+
+    /// Footer.
+    OpNum op_num = OpNum::Error;
+    bool done = true;
+    int32_t error_read = - 1;
+
+    return total_size + Coordination::size(op_num) + Coordination::size(done) + Coordination::size(error_read);
+}
+
+ZooKeeperResponsePtr ZooKeeperHeartbeatRequest::makeResponse() const { return std::make_shared<ZooKeeperHeartbeatResponse>(); }
+ZooKeeperResponsePtr ZooKeeperSyncRequest::makeResponse() const { return std::make_shared<ZooKeeperSyncResponse>(); }
+ZooKeeperResponsePtr ZooKeeperAuthRequest::makeResponse() const { return std::make_shared<ZooKeeperAuthResponse>(); }
+ZooKeeperResponsePtr ZooKeeperRemoveRequest::makeResponse() const { return std::make_shared<ZooKeeperRemoveResponse>(); }
+ZooKeeperResponsePtr ZooKeeperRemoveRecursiveRequest::makeResponse() const { return std::make_shared<ZooKeeperRemoveRecursiveResponse>(); }
+ZooKeeperResponsePtr ZooKeeperExistsRequest::makeResponse() const { return std::make_shared<ZooKeeperExistsResponse>(); }
+ZooKeeperResponsePtr ZooKeeperGetRequest::makeResponse() const { return std::make_shared<ZooKeeperGetResponse>(); }
+ZooKeeperResponsePtr ZooKeeperSetRequest::makeResponse() const { return std::make_shared<ZooKeeperSetResponse>(); }
+ZooKeeperResponsePtr ZooKeeperReconfigRequest::makeResponse() const { return std::make_shared<ZooKeeperReconfigResponse>(); }
+ZooKeeperResponsePtr ZooKeeperListRequest::makeResponse() const { return std::make_shared<ZooKeeperListResponse>(); }
+ZooKeeperResponsePtr ZooKeeperSimpleListRequest::makeResponse() const { return std::make_shared<ZooKeeperSimpleListResponse>(); }
 
 ZooKeeperResponsePtr ZooKeeperCreateRequest::makeResponse() const
 {
     if (not_exists)
-        return setTime(std::make_shared<ZooKeeperCreateIfNotExistsResponse>());
+        return std::make_shared<ZooKeeperCreateIfNotExistsResponse>();
-    return setTime(std::make_shared<ZooKeeperCreateResponse>());
+    return std::make_shared<ZooKeeperCreateResponse>();
 }
 
 ZooKeeperResponsePtr ZooKeeperCheckRequest::makeResponse() const
 {
     if (not_exists)
-        return setTime(std::make_shared<ZooKeeperCheckNotExistsResponse>());
+        return std::make_shared<ZooKeeperCheckNotExistsResponse>();
 
-    return setTime(std::make_shared<ZooKeeperCheckResponse>());
+    return std::make_shared<ZooKeeperCheckResponse>();
 }
 
 ZooKeeperResponsePtr ZooKeeperMultiRequest::makeResponse() const
@@ -764,11 +941,12 @@ ZooKeeperResponsePtr ZooKeeperMultiRequest::makeResponse() const
     else
        response = std::make_shared<ZooKeeperMultiReadResponse>(requests);
 
-    return setTime(std::move(response));
+    return std::move(response);
 }
-ZooKeeperResponsePtr ZooKeeperCloseRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperCloseResponse>()); }
+
-ZooKeeperResponsePtr ZooKeeperSetACLRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperSetACLResponse>()); }
+ZooKeeperResponsePtr ZooKeeperCloseRequest::makeResponse() const { return std::make_shared<ZooKeeperCloseResponse>(); }
-ZooKeeperResponsePtr ZooKeeperGetACLRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperGetACLResponse>()); }
+ZooKeeperResponsePtr ZooKeeperSetACLRequest::makeResponse() const { return std::make_shared<ZooKeeperSetACLResponse>(); }
+ZooKeeperResponsePtr ZooKeeperGetACLRequest::makeResponse() const { return std::make_shared<ZooKeeperGetACLResponse>(); }
 
 void ZooKeeperSessionIDRequest::writeImpl(WriteBuffer & out) const
 {
@@ -777,6 +955,11 @@ void ZooKeeperSessionIDRequest::writeImpl(WriteBuffer & out) const
     Coordination::write(server_id, out);
 }
 
+size_t ZooKeeperSessionIDRequest::sizeImpl() const
+{
+    return Coordination::size(internal_id) + Coordination::size(session_timeout_ms) + Coordination::size(server_id);
+}
+
 void ZooKeeperSessionIDRequest::readImpl(ReadBuffer & in)
 {
     Coordination::read(internal_id, in);
@@ -803,6 +986,11 @@ void ZooKeeperSessionIDResponse::writeImpl(WriteBuffer & out) const
     Coordination::write(server_id, out);
 }
 
+size_t ZooKeeperSessionIDResponse::sizeImpl() const
+{
+    return Coordination::size(internal_id) + Coordination::size(session_id) + Coordination::size(server_id);
+}
+
 
 void ZooKeeperRequest::createLogElements(LogElements & elems) const
 {
@@ -960,40 +1148,6 @@ std::shared_ptr<ZooKeeperRequest> ZooKeeperRequest::read(ReadBuffer & in)
     return request;
 }
 
-ZooKeeperRequest::~ZooKeeperRequest()
-{
-    if (!request_created_time_ns)
-        return;
-    UInt64 elapsed_ns = clock_gettime_ns() - request_created_time_ns;
-    constexpr UInt64 max_request_time_ns = 1000000000ULL; /// 1 sec
-    if (max_request_time_ns < elapsed_ns)
-    {
-        LOG_TEST(getLogger(__PRETTY_FUNCTION__), "Processing of request xid={} took {} ms", xid, elapsed_ns / 1000000UL);
-    }
-}
-
-ZooKeeperResponsePtr ZooKeeperRequest::setTime(ZooKeeperResponsePtr response) const
-{
-    if (request_created_time_ns)
-    {
-        response->response_created_time_ns = clock_gettime_ns();
-    }
-    return response;
-}
-
-ZooKeeperResponse::~ZooKeeperResponse()
-{
-    if (!response_created_time_ns)
-        return;
-    UInt64 elapsed_ns = clock_gettime_ns() - response_created_time_ns;
-    constexpr UInt64 max_request_time_ns = 1000000000ULL; /// 1 sec
-    if (max_request_time_ns < elapsed_ns)
-    {
-        LOG_TEST(getLogger(__PRETTY_FUNCTION__), "Processing of response xid={} took {} ms", xid, elapsed_ns / 1000000UL);
-    }
-}
-
-
 ZooKeeperRequestPtr ZooKeeperRequestFactory::get(OpNum op_num) const
 {
     auto it = op_num_to_request.find(op_num);
@@ -1015,7 +1169,6 @@ void registerZooKeeperRequest(ZooKeeperRequestFactory & factory)
     factory.registerRequest(num, []
     {
         auto res = std::make_shared<RequestT>();
-        res->request_created_time_ns = clock_gettime_ns();
 
         if constexpr (num == OpNum::MultiRead)
             res->operation_type = ZooKeeperMultiRequest::OperationType::Read;

src/Common/ZooKeeper/ZooKeeperCommon.h

@@ -7,13 +7,11 @@
 #include <boost/noncopyable.hpp>
 #include <IO/ReadBuffer.h>
 #include <IO/WriteBuffer.h>
-#include <unordered_map>
 #include <vector>
 #include <memory>
 #include <cstdint>
 #include <optional>
 #include <functional>
-#include <span>
 
 
 namespace Coordination
@@ -25,13 +23,11 @@ struct ZooKeeperResponse : virtual Response
 {
     XID xid = 0;
 
-    UInt64 response_created_time_ns = 0;
-
     ZooKeeperResponse() = default;
     ZooKeeperResponse(const ZooKeeperResponse &) = default;
-    ~ZooKeeperResponse() override;
     virtual void readImpl(ReadBuffer &) = 0;
     virtual void writeImpl(WriteBuffer &) const = 0;
+    virtual size_t sizeImpl() const = 0;
     virtual void write(WriteBuffer & out) const;
     virtual OpNum getOpNum() const = 0;
     virtual void fillLogElements(LogElements & elems, size_t idx) const;
@@ -51,13 +47,11 @@ struct ZooKeeperRequest : virtual Request
 
     bool restored_from_zookeeper_log = false;
 
-    UInt64 request_created_time_ns = 0;
     UInt64 thread_id = 0;
     String query_id;
 
     ZooKeeperRequest() = default;
     ZooKeeperRequest(const ZooKeeperRequest &) = default;
-    ~ZooKeeperRequest() override;
 
     virtual OpNum getOpNum() const = 0;
 
@@ -66,6 +60,7 @@ struct ZooKeeperRequest : virtual Request
     std::string toString(bool short_format = false) const;
 
     virtual void writeImpl(WriteBuffer &) const = 0;
+    virtual size_t sizeImpl() const = 0;
     virtual void readImpl(ReadBuffer &) = 0;
 
     virtual std::string toStringImpl(bool /*short_format*/) const { return ""; }
@@ -73,7 +68,6 @@ struct ZooKeeperRequest : virtual Request
     static std::shared_ptr<ZooKeeperRequest> read(ReadBuffer & in);
 
     virtual ZooKeeperResponsePtr makeResponse() const = 0;
-    ZooKeeperResponsePtr setTime(ZooKeeperResponsePtr response) const;
     virtual bool isReadRequest() const = 0;
 
     virtual void createLogElements(LogElements & elems) const;
@@ -86,6 +80,7 @@ struct ZooKeeperHeartbeatRequest final : ZooKeeperRequest
     String getPath() const override { return {}; }
     OpNum getOpNum() const override { return OpNum::Heartbeat; }
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
     void readImpl(ReadBuffer &) override {}
     ZooKeeperResponsePtr makeResponse() const override;
     bool isReadRequest() const override { return false; }
@@ -97,6 +92,7 @@ struct ZooKeeperSyncRequest final : ZooKeeperRequest
     String getPath() const override { return path; }
     OpNum getOpNum() const override { return OpNum::Sync; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
     ZooKeeperResponsePtr makeResponse() const override;
@@ -109,6 +105,7 @@ struct ZooKeeperSyncResponse final : SyncResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::Sync; }
 };
 
@@ -122,6 +119,7 @@ struct ZooKeeperReconfigRequest final : ZooKeeperRequest
     String getPath() const override { return keeper_config_path; }
     OpNum getOpNum() const override { return OpNum::Reconfig; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
     ZooKeeperResponsePtr makeResponse() const override;
@@ -138,6 +136,7 @@ struct ZooKeeperReconfigResponse final : ReconfigResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::Reconfig; }
 };
 
@@ -145,6 +144,7 @@ struct ZooKeeperHeartbeatResponse final : ZooKeeperResponse
 {
     void readImpl(ReadBuffer &) override {}
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
     OpNum getOpNum() const override { return OpNum::Heartbeat; }
 };
 
@@ -153,6 +153,7 @@ struct ZooKeeperWatchResponse final : WatchResponse, ZooKeeperResponse
     void readImpl(ReadBuffer & in) override;
 
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
 
     void write(WriteBuffer & out) const override;
 
@@ -175,6 +176,7 @@ struct ZooKeeperAuthRequest final : ZooKeeperRequest
     String getPath() const override { return {}; }
     OpNum getOpNum() const override { return OpNum::Auth; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -189,6 +191,7 @@ struct ZooKeeperAuthResponse final : ZooKeeperResponse
 {
     void readImpl(ReadBuffer &) override {}
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
 
     OpNum getOpNum() const override { return OpNum::Auth; }
 
@@ -200,6 +203,7 @@ struct ZooKeeperCloseRequest final : ZooKeeperRequest
     String getPath() const override { return {}; }
     OpNum getOpNum() const override { return OpNum::Close; }
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
     void readImpl(ReadBuffer &) override {}
 
     ZooKeeperResponsePtr makeResponse() const override;
@@ -214,6 +218,7 @@ struct ZooKeeperCloseResponse final : ZooKeeperResponse
     }
 
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
 
     OpNum getOpNum() const override { return OpNum::Close; }
 };
@@ -228,6 +233,7 @@ struct ZooKeeperCreateRequest final : public CreateRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return not_exists ? OpNum::CreateIfNotExists : OpNum::Create; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -244,6 +250,7 @@ struct ZooKeeperCreateResponse : CreateResponse, ZooKeeperResponse
     void readImpl(ReadBuffer & in) override;
 
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
 
     OpNum getOpNum() const override { return OpNum::Create; }
 
@@ -265,6 +272,7 @@ struct ZooKeeperRemoveRequest final : RemoveRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return OpNum::Remove; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -280,6 +288,7 @@ struct ZooKeeperRemoveResponse final : RemoveResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer &) override {}
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
     OpNum getOpNum() const override { return OpNum::Remove; }
 
     size_t bytesSize() const override { return RemoveResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -293,6 +302,7 @@ struct ZooKeeperRemoveRecursiveRequest final : RemoveRecursiveRequest, ZooKeeper
     OpNum getOpNum() const override { return OpNum::RemoveRecursive; }
     void writeImpl(WriteBuffer & out) const override;
     void readImpl(ReadBuffer & in) override;
+    size_t sizeImpl() const override;
     std::string toStringImpl(bool short_format) const override;
 
     ZooKeeperResponsePtr makeResponse() const override;
@@ -305,6 +315,7 @@ struct ZooKeeperRemoveRecursiveResponse : RemoveRecursiveResponse, ZooKeeperResp
 {
     void readImpl(ReadBuffer &) override {}
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
     OpNum getOpNum() const override { return OpNum::RemoveRecursive; }
 
     size_t bytesSize() const override { return RemoveRecursiveResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -317,6 +328,7 @@ struct ZooKeeperExistsRequest final : ExistsRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return OpNum::Exists; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -330,6 +342,7 @@ struct ZooKeeperExistsResponse final : ExistsResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::Exists; }
 
     size_t bytesSize() const override { return ExistsResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -344,6 +357,7 @@ struct ZooKeeperGetRequest final : GetRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return OpNum::Get; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -357,6 +371,7 @@ struct ZooKeeperGetResponse final : GetResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::Get; }
 
     size_t bytesSize() const override { return GetResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -371,6 +386,7 @@ struct ZooKeeperSetRequest final : SetRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return OpNum::Set; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
     ZooKeeperResponsePtr makeResponse() const override;
@@ -385,6 +401,7 @@ struct ZooKeeperSetResponse final : SetResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::Set; }
 
     size_t bytesSize() const override { return SetResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -399,6 +416,7 @@ struct ZooKeeperListRequest : ListRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return OpNum::List; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
     ZooKeeperResponsePtr makeResponse() const override;
@@ -419,6 +437,7 @@ struct ZooKeeperFilteredListRequest final : ZooKeeperListRequest
 
     OpNum getOpNum() const override { return OpNum::FilteredList; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -429,6 +448,7 @@ struct ZooKeeperListResponse : ListResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::List; }
 
     size_t bytesSize() const override { return ListResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -440,6 +460,7 @@ struct ZooKeeperSimpleListResponse final : ZooKeeperListResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::SimpleList; }
 
     size_t bytesSize() const override { return ZooKeeperListResponse::bytesSize() - sizeof(stat); }
@@ -452,6 +473,7 @@ struct ZooKeeperCheckRequest : CheckRequest, ZooKeeperRequest
 
     OpNum getOpNum() const override { return not_exists ? OpNum::CheckNotExists : OpNum::Check; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -467,6 +489,7 @@ struct ZooKeeperCheckResponse : CheckResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer &) override {}
     void writeImpl(WriteBuffer &) const override {}
+    size_t sizeImpl() const override { return 0; }
     OpNum getOpNum() const override { return OpNum::Check; }
 
     size_t bytesSize() const override { return CheckResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -483,6 +506,7 @@ struct ZooKeeperErrorResponse final : ErrorResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
 
     OpNum getOpNum() const override { return OpNum::Error; }
 
@@ -493,6 +517,7 @@ struct ZooKeeperSetACLRequest final : SetACLRequest, ZooKeeperRequest
 {
     OpNum getOpNum() const override { return OpNum::SetACL; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
     ZooKeeperResponsePtr makeResponse() const override;
@@ -505,6 +530,7 @@ struct ZooKeeperSetACLResponse final : SetACLResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::SetACL; }
 
     size_t bytesSize() const override { return SetACLResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
@@ -514,6 +540,7 @@ struct ZooKeeperGetACLRequest final : GetACLRequest, ZooKeeperRequest
 {
     OpNum getOpNum() const override { return OpNum::GetACL; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
     ZooKeeperResponsePtr makeResponse() const override;
@@ -526,12 +553,13 @@ struct ZooKeeperGetACLResponse final : GetACLResponse, ZooKeeperResponse
 {
     void readImpl(ReadBuffer & in) override;
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     OpNum getOpNum() const override { return OpNum::GetACL; }
 
     size_t bytesSize() const override { return GetACLResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
 };
 
-struct ZooKeeperMultiRequest final : MultiRequest, ZooKeeperRequest
+struct ZooKeeperMultiRequest final : MultiRequest<ZooKeeperRequestPtr>, ZooKeeperRequest
 {
     OpNum getOpNum() const override;
     ZooKeeperMultiRequest() = default;
@@ -540,6 +568,7 @@ struct ZooKeeperMultiRequest final : MultiRequest, ZooKeeperRequest
     ZooKeeperMultiRequest(std::span<const Coordination::RequestPtr> generic_requests, const ACLs & default_acls);
 
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
     std::string toStringImpl(bool short_format) const override;
 
@@ -563,12 +592,14 @@ private:
 
 struct ZooKeeperMultiResponse : MultiResponse, ZooKeeperResponse
 {
-    explicit ZooKeeperMultiResponse(const Requests & requests)
+    ZooKeeperMultiResponse() = default;
+
+    explicit ZooKeeperMultiResponse(const std::vector<ZooKeeperRequestPtr> & requests)
     {
         responses.reserve(requests.size());
 
         for (const auto & request : requests)
-            responses.emplace_back(dynamic_cast<const ZooKeeperRequest &>(*request).makeResponse());
+            responses.emplace_back(request->makeResponse());
     }
 
     explicit ZooKeeperMultiResponse(const Responses & responses_)
@@ -579,6 +610,7 @@ struct ZooKeeperMultiResponse : MultiResponse, ZooKeeperResponse
     void readImpl(ReadBuffer & in) override;
 
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
 
     size_t bytesSize() const override { return MultiResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
 
@@ -609,6 +641,7 @@ struct ZooKeeperSessionIDRequest final : ZooKeeperRequest
     Coordination::OpNum getOpNum() const override { return OpNum::SessionID; }
     String getPath() const override { return {}; }
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
     void readImpl(ReadBuffer & in) override;
 
     Coordination::ZooKeeperResponsePtr makeResponse() const override;
@@ -627,6 +660,7 @@ struct ZooKeeperSessionIDResponse final : ZooKeeperResponse
     void readImpl(ReadBuffer & in) override;
 
     void writeImpl(WriteBuffer & out) const override;
+    size_t sizeImpl() const override;
 
     Coordination::OpNum getOpNum() const override { return OpNum::SessionID; }
 };

src/Common/ZooKeeper/ZooKeeperIO.cpp

@@ -42,6 +42,32 @@ void write(const Error & x, WriteBuffer & out)
     write(static_cast<int32_t>(x), out);
 }
 
+size_t size(OpNum x)
+{
+    return size(static_cast<int32_t>(x));
+}
+
+size_t size(const std::string & s)
+{
+    return size(static_cast<int32_t>(s.size())) + s.size();
+}
+
+size_t size(const ACL & acl)
+{
+    return size(acl.permissions) + size(acl.scheme) + size(acl.id);
+}
+
+size_t size(const Stat & stat)
+{
+    return size(stat.czxid) + size(stat.mzxid) + size(stat.ctime) + size(stat.mtime) + size(stat.version) + size(stat.cversion)
+        + size(stat.aversion) + size(stat.ephemeralOwner) + size(stat.dataLength) + size(stat.numChildren) + size(stat.pzxid);
+}
+
+size_t size(const Error & x)
+{
+    return size(static_cast<int32_t>(x));
+}
+
 void read(OpNum & x, ReadBuffer & in)
 {
     int32_t raw_op_num;

src/Common/ZooKeeper/ZooKeeperIO.h

@@ -43,6 +43,36 @@ void write(const std::vector<T> & arr, WriteBuffer & out)
         write(elem, out);
 }
 
+template <typename T>
+requires is_arithmetic_v<T>
+size_t size(T x)
+{
+    return sizeof(x);
+}
+
+size_t size(OpNum x);
+size_t size(const std::string & s);
+size_t size(const ACL & acl);
+size_t size(const Stat & stat);
+size_t size(const Error & x);
+
+template <size_t N>
+size_t size(const std::array<char, N>)
+{
+    return size(static_cast<int32_t>(N)) + N;
+}
+
+template <typename T>
+size_t size(const std::vector<T> & arr)
+{
+    size_t total_size = size(static_cast<int32_t>(arr.size()));
+    for (const auto & elem : arr)
+        total_size += size(elem);
+
+    return total_size;
+}
+
+
 template <typename T>
 requires is_arithmetic_v<T>
 void read(T & x, ReadBuffer & in)

src/Common/benchmarks/CMakeLists.txt

@@ -1,4 +1,4 @@
-clickhouse_add_executable(integer_hash_tables_and_hashes integer_hash_tables_and_hashes.cpp)
+clickhouse_add_executable(integer_hash_tables_and_hashes integer_hash_tables_and_hashes.cpp orc_string_dictionary.cpp)
 target_link_libraries (integer_hash_tables_and_hashes PRIVATE
     ch_contrib::gbenchmark_all
     dbms
@@ -7,3 +7,8 @@ target_link_libraries (integer_hash_tables_and_hashes PRIVATE
     ch_contrib::wyhash
     ch_contrib::farmhash
     ch_contrib::xxHash)
+
+clickhouse_add_executable(orc_string_dictionary orc_string_dictionary.cpp)
+target_link_libraries (orc_string_dictionary PRIVATE
+    ch_contrib::gbenchmark_all
+    dbms)

src/Common/benchmarks/orc_string_dictionary.cpp

@@ -0,0 +1,311 @@
+#include <cstdlib>
+#include <base/defines.h>
+#include <benchmark/benchmark.h>
+
+class OldSortedStringDictionary
+{
+public:
+    struct DictEntry
+    {
+        DictEntry(const char * str, size_t len) : data(str), length(len) { }
+        const char * data;
+        size_t length;
+    };
+
+    OldSortedStringDictionary() : totalLength(0) { }
+
+    // insert a new string into dictionary, return its insertion order
+    size_t insert(const char * str, size_t len);
+
+    // reorder input index buffer from insertion order to dictionary order
+    void reorder(std::vector<int64_t> & idxBuffer) const;
+
+    // get dict entries in insertion order
+    void getEntriesInInsertionOrder(std::vector<const DictEntry *> &) const;
+
+    size_t size() const;
+
+    // return total length of strings in the dictionary
+    uint64_t length() const;
+
+    void clear();
+
+    // store indexes of insertion order in the dictionary for not-null rows
+    std::vector<int64_t> idxInDictBuffer;
+
+private:
+    struct LessThan
+    {
+        bool operator()(const DictEntry & left, const DictEntry & right) const
+        {
+            int ret = memcmp(left.data, right.data, std::min(left.length, right.length));
+            if (ret != 0)
+            {
+                return ret < 0;
+            }
+            return left.length < right.length;
+        }
+    };
+
+    std::map<DictEntry, size_t, LessThan> dict;
+    std::vector<std::vector<char>> data;
+    uint64_t totalLength;
+};
+
+// insert a new string into dictionary, return its insertion order
+size_t OldSortedStringDictionary::insert(const char * str, size_t len)
+{
+    auto ret = dict.insert({DictEntry(str, len), dict.size()});
+    if (ret.second)
+    {
+        // make a copy to internal storage
+        data.push_back(std::vector<char>(len));
+        memcpy(data.back().data(), str, len);
+        // update dictionary entry to link pointer to internal storage
+        DictEntry * entry = const_cast<DictEntry *>(&(ret.first->first));
+        entry->data = data.back().data();
+        totalLength += len;
+    }
+    return ret.first->second;
+}
+
+/**
+   * Reorder input index buffer from insertion order to dictionary order
+   *
+   * We require this function because string values are buffered by indexes
+   * in their insertion order. Until the entire dictionary is complete can
+   * we get their sorted indexes in the dictionary in that ORC specification
+   * demands dictionary should be ordered. Therefore this function transforms
+   * the indexes from insertion order to dictionary value order for final
+   * output.
+   */
+void OldSortedStringDictionary::reorder(std::vector<int64_t> & idxBuffer) const
+{
+    // iterate the dictionary to get mapping from insertion order to value order
+    std::vector<size_t> mapping(dict.size());
+    size_t dictIdx = 0;
+    for (auto it = dict.cbegin(); it != dict.cend(); ++it)
+    {
+        mapping[it->second] = dictIdx++;
+    }
+
+    // do the transformation
+    for (size_t i = 0; i != idxBuffer.size(); ++i)
+    {
+        idxBuffer[i] = static_cast<int64_t>(mapping[static_cast<size_t>(idxBuffer[i])]);
+    }
+}
+
+// get dict entries in insertion order
+void OldSortedStringDictionary::getEntriesInInsertionOrder(std::vector<const DictEntry *> & entries) const
+{
+    entries.resize(dict.size());
+    for (auto it = dict.cbegin(); it != dict.cend(); ++it)
+    {
+        entries[it->second] = &(it->first);
+    }
+}
+
+// return count of entries
+size_t OldSortedStringDictionary::size() const
+{
+    return dict.size();
+}
+
+// return total length of strings in the dictionary
+uint64_t OldSortedStringDictionary::length() const
+{
+    return totalLength;
+}
+
+void OldSortedStringDictionary::clear()
+{
+    totalLength = 0;
+    data.clear();
+    dict.clear();
+}
+
+
+/**
+   * Implementation of increasing sorted string dictionary
+   */
+class NewSortedStringDictionary
+{
+public:
+    struct DictEntry
+    {
+        DictEntry(const char * str, size_t len) : data(str), length(len) { }
+        const char * data;
+        size_t length;
+    };
+
+    struct DictEntryWithIndex
+    {
+        DictEntryWithIndex(const char * str, size_t len, size_t index_) : entry(str, len), index(index_) { }
+        DictEntry entry;
+        size_t index;
+    };
+
+    NewSortedStringDictionary() : totalLength_(0) { }
+
+    // insert a new string into dictionary, return its insertion order
+    size_t insert(const char * str, size_t len);
+
+    // reorder input index buffer from insertion order to dictionary order
+    void reorder(std::vector<int64_t> & idxBuffer) const;
+
+    // get dict entries in insertion order
+    void getEntriesInInsertionOrder(std::vector<const DictEntry *> &) const;
+
+    // return count of entries
+    size_t size() const;
+
+    // return total length of strings in the dictionary
+    uint64_t length() const;
+
+    void clear();
+
+    // store indexes of insertion order in the dictionary for not-null rows
+    std::vector<int64_t> idxInDictBuffer;
+
+private:
+    struct LessThan
+    {
+        bool operator()(const DictEntryWithIndex & l, const DictEntryWithIndex & r)
+        {
+            const auto & left = l.entry;
+            const auto & right = r.entry;
+            int ret = memcmp(left.data, right.data, std::min(left.length, right.length));
+            if (ret != 0)
+            {
+                return ret < 0;
+            }
+            return left.length < right.length;
+        }
+    };
+
+    mutable std::vector<DictEntryWithIndex> flatDict_;
+    std::unordered_map<std::string, size_t> keyToIndex;
+    uint64_t totalLength_;
+};
+
+// insert a new string into dictionary, return its insertion order
+size_t NewSortedStringDictionary::insert(const char * str, size_t len)
+{
+    size_t index = flatDict_.size();
+    auto ret = keyToIndex.emplace(std::string(str, len), index);
+    if (ret.second)
+    {
+        flatDict_.emplace_back(ret.first->first.data(), ret.first->first.size(), index);
+        totalLength_ += len;
+    }
+    return ret.first->second;
+}
+
+/**
+   * Reorder input index buffer from insertion order to dictionary order
+   *
+   * We require this function because string values are buffered by indexes
+   * in their insertion order. Until the entire dictionary is complete can
+   * we get their sorted indexes in the dictionary in that ORC specification
+   * demands dictionary should be ordered. Therefore this function transforms
+   * the indexes from insertion order to dictionary value order for final
+   * output.
+   */
+void NewSortedStringDictionary::reorder(std::vector<int64_t> & idxBuffer) const
+{
+    // iterate the dictionary to get mapping from insertion order to value order
+    std::vector<size_t> mapping(flatDict_.size());
+    for (size_t i = 0; i < flatDict_.size(); ++i)
+    {
+        mapping[flatDict_[i].index] = i;
+    }
+
+    // do the transformation
+    for (size_t i = 0; i != idxBuffer.size(); ++i)
+    {
+        idxBuffer[i] = static_cast<int64_t>(mapping[static_cast<size_t>(idxBuffer[i])]);
+    }
+}
+
+// get dict entries in insertion order
+void NewSortedStringDictionary::getEntriesInInsertionOrder(std::vector<const DictEntry *> & entries) const
+{
+    std::sort(
+        flatDict_.begin(),
+        flatDict_.end(),
+        [](const DictEntryWithIndex & left, const DictEntryWithIndex & right) { return left.index < right.index; });
+
+    entries.resize(flatDict_.size());
+    for (size_t i = 0; i < flatDict_.size(); ++i)
+    {
+        entries[i] = &(flatDict_[i].entry);
+    }
+}
+
+// return count of entries
+size_t NewSortedStringDictionary::size() const
+{
+    return flatDict_.size();
+}
+
+// return total length of strings in the dictionary
+uint64_t NewSortedStringDictionary::length() const
+{
+    return totalLength_;
+}
+
+void NewSortedStringDictionary::clear()
+{
+    totalLength_ = 0;
+    keyToIndex.clear();
+    flatDict_.clear();
+}
+
+template <size_t cardinality>
+static std::vector<std::string> mockStrings()
+{
+    std::vector<std::string> res(1000000);
+    for (auto & s : res)
+    {
+        s = "test string dictionary " + std::to_string(rand() % cardinality);
+    }
+    return res;
+}
+
+template <typename DictionaryImpl>
+static NO_INLINE std::unique_ptr<DictionaryImpl> createAndWriteStringDictionary(const std::vector<std::string> & strs)
+{
+    auto dict = std::make_unique<DictionaryImpl>();
+    for (const auto & str : strs)
+    {
+        auto index = dict->insert(str.data(), str.size());
+        dict->idxInDictBuffer.push_back(index);
+    }
+    dict->reorder(dict->idxInDictBuffer);
+
+    return dict;
+}
+
+template <typename DictionaryImpl, size_t cardinality>
+static void BM_writeStringDictionary(benchmark::State & state)
+{
+    auto strs = mockStrings<cardinality>();
+    for (auto _ : state)
+    {
+        auto dict = createAndWriteStringDictionary<DictionaryImpl>(strs);
+        benchmark::DoNotOptimize(dict);
+    }
+}
+
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, OldSortedStringDictionary, 10);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, NewSortedStringDictionary, 10);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, OldSortedStringDictionary, 100);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, NewSortedStringDictionary, 100);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, OldSortedStringDictionary, 1000);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, NewSortedStringDictionary, 1000);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, OldSortedStringDictionary, 10000);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, NewSortedStringDictionary, 10000);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, OldSortedStringDictionary, 100000);
+BENCHMARK_TEMPLATE(BM_writeStringDictionary, NewSortedStringDictionary, 100000);
+

src/Coordination/ACLMap.cpp

@@ -45,6 +45,7 @@ uint64_t ACLMap::convertACLs(const Coordination::ACLs & acls)
     if (acls.empty())
         return 0;
 
+    std::lock_guard lock(map_mutex);
     if (acl_to_num.contains(acls))
         return acl_to_num[acls];
 
@@ -62,6 +63,7 @@ Coordination::ACLs ACLMap::convertNumber(uint64_t acls_id) const
     if (acls_id == 0)
         return Coordination::ACLs{};
 
+    std::lock_guard lock(map_mutex);
     if (!num_to_acl.contains(acls_id))
         throw Exception(ErrorCodes::LOGICAL_ERROR, "Unknown ACL id {}. It's a bug", acls_id);
 
@@ -70,6 +72,7 @@ Coordination::ACLs ACLMap::convertNumber(uint64_t acls_id) const
 
 void ACLMap::addMapping(uint64_t acls_id, const Coordination::ACLs & acls)
 {
+    std::lock_guard lock(map_mutex);
     num_to_acl[acls_id] = acls;
     acl_to_num[acls] = acls_id;
     max_acl_id = std::max(acls_id + 1, max_acl_id); /// max_acl_id pointer next slot
@@ -77,11 +80,13 @@ void ACLMap::addMapping(uint64_t acls_id, const Coordination::ACLs & acls)
 
 void ACLMap::addUsage(uint64_t acl_id)
 {
+    std::lock_guard lock(map_mutex);
     usage_counter[acl_id]++;
 }
 
 void ACLMap::removeUsage(uint64_t acl_id)
 {
+    std::lock_guard lock(map_mutex);
     if (!usage_counter.contains(acl_id))
         return;
 

src/Coordination/ACLMap.h

@@ -32,6 +32,8 @@ private:
     NumToACLMap num_to_acl;
     UsageCounter usage_counter;
     uint64_t max_acl_id{1};
+
+    mutable std::mutex map_mutex;
 public:
 
     /// Convert ACL to number. If it's new ACL than adds it to map

src/Coordination/FourLetterCommand.cpp

@@ -301,11 +301,13 @@ String MonitorCommand::run()
 
     print(ret, "server_state", keeper_info.getRole());
 
-    print(ret, "znode_count", state_machine.getNodesCount());
+    const auto & storage_stats = state_machine.getStorageStats();
-    print(ret, "watch_count", state_machine.getTotalWatchesCount());
+
-    print(ret, "ephemerals_count", state_machine.getTotalEphemeralNodesCount());
+    print(ret, "znode_count", storage_stats.nodes_count.load(std::memory_order_relaxed));
-    print(ret, "approximate_data_size", state_machine.getApproximateDataSize());
+    print(ret, "watch_count", storage_stats.total_watches_count.load(std::memory_order_relaxed));
-    print(ret, "key_arena_size", state_machine.getKeyArenaSize());
+    print(ret, "ephemerals_count", storage_stats.total_emphemeral_nodes_count.load(std::memory_order_relaxed));
+    print(ret, "approximate_data_size", storage_stats.approximate_data_size.load(std::memory_order_relaxed));
+    print(ret, "key_arena_size", 0);
     print(ret, "latest_snapshot_size", state_machine.getLatestSnapshotSize());
 
 #if defined(OS_LINUX) || defined(OS_DARWIN)
@@ -387,6 +389,7 @@ String ServerStatCommand::run()
 
     auto & stats = keeper_dispatcher.getKeeperConnectionStats();
     Keeper4LWInfo keeper_info = keeper_dispatcher.getKeeper4LWInfo();
+    const auto & storage_stats = keeper_dispatcher.getStateMachine().getStorageStats();
 
     write("ClickHouse Keeper version", String(VERSION_DESCRIBE) + "-" + VERSION_GITHASH);
 
@@ -398,9 +401,9 @@ String ServerStatCommand::run()
     write("Sent", toString(stats.getPacketsSent()));
     write("Connections", toString(keeper_info.alive_connections_count));
     write("Outstanding", toString(keeper_info.outstanding_requests_count));
-    write("Zxid", formatZxid(keeper_info.last_zxid));
+    write("Zxid", formatZxid(storage_stats.last_zxid.load(std::memory_order_relaxed)));
     write("Mode", keeper_info.getRole());
-    write("Node count", toString(keeper_info.total_nodes_count));
+    write("Node count", toString(storage_stats.nodes_count.load(std::memory_order_relaxed)));
 
     return buf.str();
 }
@@ -416,6 +419,7 @@ String StatCommand::run()
 
     auto & stats = keeper_dispatcher.getKeeperConnectionStats();
     Keeper4LWInfo keeper_info = keeper_dispatcher.getKeeper4LWInfo();
+    const auto & storage_stats = keeper_dispatcher.getStateMachine().getStorageStats();
 
     write("ClickHouse Keeper version", String(VERSION_DESCRIBE) + "-" + VERSION_GITHASH);
 
@@ -431,9 +435,9 @@ String StatCommand::run()
     write("Sent", toString(stats.getPacketsSent()));
     write("Connections", toString(keeper_info.alive_connections_count));
     write("Outstanding", toString(keeper_info.outstanding_requests_count));
-    write("Zxid", formatZxid(keeper_info.last_zxid));
+    write("Zxid", formatZxid(storage_stats.last_zxid.load(std::memory_order_relaxed)));
     write("Mode", keeper_info.getRole());
-    write("Node count", toString(keeper_info.total_nodes_count));
+    write("Node count", toString(storage_stats.nodes_count.load(std::memory_order_relaxed)));
 
     return buf.str();
 }

src/Coordination/Keeper4LWInfo.h

@@ -1,7 +1,5 @@
 #pragma once
 
-#include <string>
-
 #include <base/types.h>
 #include <Common/Exception.h>
 
@@ -30,9 +28,6 @@ struct Keeper4LWInfo
     uint64_t follower_count;
     uint64_t synced_follower_count;
 
-    uint64_t total_nodes_count;
-    int64_t last_zxid;
-
     String getRole() const
     {
         if (is_standalone)

src/Coordination/KeeperAsynchronousMetrics.cpp

@@ -38,15 +38,16 @@ void updateKeeperInformation(KeeperDispatcher & keeper_dispatcher, AsynchronousM
         is_follower = static_cast<size_t>(keeper_info.is_follower);
         is_exceeding_mem_soft_limit = static_cast<size_t>(keeper_info.is_exceeding_mem_soft_limit);
 
-        zxid = keeper_info.last_zxid;
         const auto & state_machine = keeper_dispatcher.getStateMachine();
-        znode_count = state_machine.getNodesCount();
+        const auto & storage_stats = state_machine.getStorageStats();
-        watch_count = state_machine.getTotalWatchesCount();
+        zxid = storage_stats.last_zxid.load(std::memory_order_relaxed);
-        ephemerals_count = state_machine.getTotalEphemeralNodesCount();
+        znode_count = storage_stats.nodes_count.load(std::memory_order_relaxed);
-        approximate_data_size = state_machine.getApproximateDataSize();
+        watch_count = storage_stats.total_watches_count.load(std::memory_order_relaxed);
-        key_arena_size = state_machine.getKeyArenaSize();
+        ephemerals_count = storage_stats.total_emphemeral_nodes_count.load(std::memory_order_relaxed);
-        session_with_watches = state_machine.getSessionsWithWatchesCount();
+        approximate_data_size = storage_stats.approximate_data_size.load(std::memory_order_relaxed);
-        paths_watched = state_machine.getWatchedPathsCount();
+        key_arena_size = 0;
+        session_with_watches = storage_stats.sessions_with_watches_count.load(std::memory_order_relaxed);
+        paths_watched = storage_stats.watched_paths_count.load(std::memory_order_relaxed);
 
 #    if defined(__linux__) || defined(__APPLE__)
         open_file_descriptor_count = getCurrentProcessFDCount();

src/Coordination/KeeperDispatcher.cpp

@@ -305,7 +305,7 @@ void KeeperDispatcher::requestThread()
                 if (has_read_request)
                 {
                     if (server->isLeaderAlive())
-                        server->putLocalReadRequest(request);
+                        server->putLocalReadRequest({request});
                     else
                         addErrorResponses({request}, Coordination::Error::ZCONNECTIONLOSS);
                 }

src/Coordination/KeeperServer.cpp

@@ -28,6 +28,16 @@
 #include <Common/getMultipleKeysFromConfig.h>
 #include <Common/getNumberOfPhysicalCPUCores.h>
 
+#if USE_SSL
+#    include <Server/CertificateReloader.h>
+#    include <openssl/ssl.h>
+#    include <Poco/Crypto/EVPPKey.h>
+#    include <Poco/Net/Context.h>
+#    include <Poco/Net/SSLManager.h>
+#    include <Poco/Net/Utility.h>
+#    include <Poco/StringTokenizer.h>
+#endif
+
 #include <chrono>
 #include <mutex>
 #include <string>
@@ -48,6 +58,7 @@ namespace ErrorCodes
     extern const int SUPPORT_IS_DISABLED;
     extern const int LOGICAL_ERROR;
     extern const int INVALID_CONFIG_PARAMETER;
+    extern const int BAD_ARGUMENTS;
 }
 
 using namespace std::chrono_literals;
@@ -56,6 +67,16 @@ namespace
 {
 
 #if USE_SSL
+
+int callSetCertificate(SSL * ssl, void * arg)
+{
+    if (!arg)
+        return -1;
+
+    const CertificateReloader::Data * data = reinterpret_cast<CertificateReloader::Data *>(arg);
+    return setCertificateCallback(ssl, data, getLogger("SSLContext"));
+}
+
 void setSSLParams(nuraft::asio_service::options & asio_opts)
 {
     const Poco::Util::LayeredConfiguration & config = Poco::Util::Application::instance().config();
@@ -69,18 +90,55 @@ void setSSLParams(nuraft::asio_service::options & asio_opts)
     if (!config.has(private_key_file_property))
         throw Exception(ErrorCodes::NO_ELEMENTS_IN_CONFIG, "Server private key file is not set.");
 
-    asio_opts.enable_ssl_ = true;
+    Poco::Net::Context::Params params;
-    asio_opts.server_cert_file_ = config.getString(certificate_file_property);
+    params.certificateFile = config.getString(certificate_file_property);
-    asio_opts.server_key_file_ = config.getString(private_key_file_property);
+    if (params.certificateFile.empty())
+        throw Exception(ErrorCodes::BAD_ARGUMENTS, "Server certificate file in config '{}' is empty", certificate_file_property);
+
+    params.privateKeyFile = config.getString(private_key_file_property);
+    if (params.privateKeyFile.empty())
+        throw Exception(ErrorCodes::BAD_ARGUMENTS, "Server key file in config '{}' is empty", private_key_file_property);
+
+    auto pass_phrase = config.getString("openSSL.server.privateKeyPassphraseHandler.options.password", "");
+    auto certificate_data = std::make_shared<CertificateReloader::Data>(params.certificateFile, params.privateKeyFile, pass_phrase);
 
     if (config.has(root_ca_file_property))
-        asio_opts.root_cert_file_ = config.getString(root_ca_file_property);
+        params.caLocation = config.getString(root_ca_file_property);
 
-    if (config.getBool("openSSL.server.loadDefaultCAFile", false))
+    params.loadDefaultCAs = config.getBool("openSSL.server.loadDefaultCAFile", false);
-        asio_opts.load_default_ca_file_ = true;
+    params.verificationMode = Poco::Net::Utility::convertVerificationMode(config.getString("openSSL.server.verificationMode", "none"));
 
-    if (config.getString("openSSL.server.verificationMode", "none") == "none")
+    std::string disabled_protocols_list = config.getString("openSSL.server.disableProtocols", "");
-        asio_opts.skip_verification_ = true;
+    Poco::StringTokenizer dp_tok(disabled_protocols_list, ";,", Poco::StringTokenizer::TOK_TRIM | Poco::StringTokenizer::TOK_IGNORE_EMPTY);
+    int disabled_protocols = 0;
+    for (const auto & token : dp_tok)
+    {
+        if (token == "sslv2")
+            disabled_protocols |= Poco::Net::Context::PROTO_SSLV2;
+        else if (token == "sslv3")
+            disabled_protocols |= Poco::Net::Context::PROTO_SSLV3;
+        else if (token == "tlsv1")
+            disabled_protocols |= Poco::Net::Context::PROTO_TLSV1;
+        else if (token == "tlsv1_1")
+            disabled_protocols |= Poco::Net::Context::PROTO_TLSV1_1;
+        else if (token == "tlsv1_2")
+            disabled_protocols |= Poco::Net::Context::PROTO_TLSV1_2;
+    }
+
+    asio_opts.ssl_context_provider_server_ = [params, certificate_data, disabled_protocols]
+    {
+        Poco::Net::Context context(Poco::Net::Context::Usage::TLSV1_2_SERVER_USE, params);
+        context.disableProtocols(disabled_protocols);
+        SSL_CTX * ssl_ctx = context.takeSslContext();
+        SSL_CTX_set_cert_cb(ssl_ctx, callSetCertificate, reinterpret_cast<void *>(certificate_data.get()));
+        return ssl_ctx;
+    };
+
+    asio_opts.ssl_context_provider_client_ = [ctx_params = std::move(params)]
+    {
+        Poco::Net::Context context(Poco::Net::Context::Usage::TLSV1_2_CLIENT_USE, ctx_params);
+        return context.takeSslContext();
+    };
 }
 #endif
 
@@ -1149,8 +1207,6 @@ Keeper4LWInfo KeeperServer::getPartiallyFilled4LWInfo() const
         result.synced_follower_count = getSyncedFollowerCount();
     }
     result.is_exceeding_mem_soft_limit = isExceedingMemorySoftLimit();
-    result.total_nodes_count = getKeeperStateMachine()->getNodesCount();
-    result.last_zxid = getKeeperStateMachine()->getLastProcessedZxid();
     return result;
 }
 

src/Coordination/KeeperSnapshotManager.cpp

@@ -78,20 +78,20 @@ namespace
             writeBinary(false, out);
 
         /// Serialize stat
-        writeBinary(node.czxid, out);
+        writeBinary(node.stats.czxid, out);
-        writeBinary(node.mzxid, out);
+        writeBinary(node.stats.mzxid, out);
-        writeBinary(node.ctime(), out);
+        writeBinary(node.stats.ctime(), out);
-        writeBinary(node.mtime, out);
+        writeBinary(node.stats.mtime, out);
-        writeBinary(node.version, out);
+        writeBinary(node.stats.version, out);
-        writeBinary(node.cversion, out);
+        writeBinary(node.stats.cversion, out);
-        writeBinary(node.aversion, out);
+        writeBinary(node.stats.aversion, out);
-        writeBinary(node.ephemeralOwner(), out);
+        writeBinary(node.stats.ephemeralOwner(), out);
         if (version < SnapshotVersion::V6)
-            writeBinary(static_cast<int32_t>(node.getData().size()), out);
+            writeBinary(static_cast<int32_t>(node.stats.data_size), out);
-        writeBinary(node.numChildren(), out);
+        writeBinary(node.stats.numChildren(), out);
-        writeBinary(node.pzxid, out);
+        writeBinary(node.stats.pzxid, out);
 
-        writeBinary(node.seqNum(), out);
+        writeBinary(node.stats.seqNum(), out);
 
         if (version >= SnapshotVersion::V4 && version <= SnapshotVersion::V5)
             writeBinary(node.sizeInBytes(), out);
@@ -100,11 +100,11 @@ namespace
     template<typename Node>
     void readNode(Node & node, ReadBuffer & in, SnapshotVersion version, ACLMap & acl_map)
     {
-        readVarUInt(node.data_size, in);
+        readVarUInt(node.stats.data_size, in);
-        if (node.data_size != 0)
+        if (node.stats.data_size != 0)
         {
-            node.data = std::unique_ptr<char[]>(new char[node.data_size]);
+            node.data = std::unique_ptr<char[]>(new char[node.stats.data_size]);
-            in.readStrict(node.data.get(), node.data_size);
+            in.readStrict(node.data.get(), node.stats.data_size);
         }
 
         if (version >= SnapshotVersion::V1)
@@ -141,19 +141,19 @@ namespace
         }
 
         /// Deserialize stat
-        readBinary(node.czxid, in);
+        readBinary(node.stats.czxid, in);
-        readBinary(node.mzxid, in);
+        readBinary(node.stats.mzxid, in);
         int64_t ctime;
         readBinary(ctime, in);
-        node.setCtime(ctime);
+        node.stats.setCtime(ctime);
-        readBinary(node.mtime, in);
+        readBinary(node.stats.mtime, in);
-        readBinary(node.version, in);
+        readBinary(node.stats.version, in);
-        readBinary(node.cversion, in);
+        readBinary(node.stats.cversion, in);
-        readBinary(node.aversion, in);
+        readBinary(node.stats.aversion, in);
         int64_t ephemeral_owner = 0;
         readBinary(ephemeral_owner, in);
         if (ephemeral_owner != 0)
-            node.setEphemeralOwner(ephemeral_owner);
+            node.stats.setEphemeralOwner(ephemeral_owner);
 
         if (version < SnapshotVersion::V6)
         {
@@ -163,14 +163,14 @@ namespace
         int32_t num_children = 0;
         readBinary(num_children, in);
         if (ephemeral_owner == 0)
-            node.setNumChildren(num_children);
+            node.stats.setNumChildren(num_children);
 
-        readBinary(node.pzxid, in);
+        readBinary(node.stats.pzxid, in);
 
         int32_t seq_num = 0;
         readBinary(seq_num, in);
         if (ephemeral_owner == 0)
-            node.setSeqNum(seq_num);
+            node.stats.setSeqNum(seq_num);
 
         if (version >= SnapshotVersion::V4 && version <= SnapshotVersion::V5)
         {
@@ -256,7 +256,7 @@ void KeeperStorageSnapshot<Storage>::serialize(const KeeperStorageSnapshot<Stora
         /// Benign race condition possible while taking snapshot: NuRaft decide to create snapshot at some log id
         /// and only after some time we lock storage and enable snapshot mode. So snapshot_container_size can be
         /// slightly bigger than required.
-        if (node.mzxid > snapshot.zxid)
+        if (node.stats.mzxid > snapshot.zxid)
             break;
         writeBinary(path, out);
         writeNode(node, snapshot.version, out);
@@ -306,7 +306,7 @@ void KeeperStorageSnapshot<Storage>::serialize(const KeeperStorageSnapshot<Stora
 }
 
 template<typename Storage>
-void KeeperStorageSnapshot<Storage>::deserialize(SnapshotDeserializationResult<Storage> & deserialization_result, ReadBuffer & in, KeeperContextPtr keeper_context)
+void KeeperStorageSnapshot<Storage>::deserialize(SnapshotDeserializationResult<Storage> & deserialization_result, ReadBuffer & in, KeeperContextPtr keeper_context) TSA_NO_THREAD_SAFETY_ANALYSIS
 {
     uint8_t version;
     readBinary(version, in);
@@ -435,13 +435,13 @@ void KeeperStorageSnapshot<Storage>::deserialize(SnapshotDeserializationResult<S
             }
         }
 
-        auto ephemeral_owner = node.ephemeralOwner();
+        auto ephemeral_owner = node.stats.ephemeralOwner();
         if constexpr (!use_rocksdb)
-            if (!node.isEphemeral() && node.numChildren() > 0)
+            if (!node.stats.isEphemeral() && node.stats.numChildren() > 0)
-                node.getChildren().reserve(node.numChildren());
+                node.getChildren().reserve(node.stats.numChildren());
 
         if (ephemeral_owner != 0)
-            storage.ephemerals[node.ephemeralOwner()].insert(std::string{path});
+            storage.committed_ephemerals[node.stats.ephemeralOwner()].insert(std::string{path});
 
         if (recalculate_digest)
             storage.nodes_digest += node.getDigest(path);
@@ -467,16 +467,25 @@ void KeeperStorageSnapshot<Storage>::deserialize(SnapshotDeserializationResult<S
         {
             if (itr.key != "/")
             {
-                if (itr.value.numChildren() != static_cast<int32_t>(itr.value.getChildren().size()))
+                if (itr.value.stats.numChildren() != static_cast<int32_t>(itr.value.getChildren().size()))
                 {
 #ifdef NDEBUG
                     /// TODO (alesapin) remove this, it should be always CORRUPTED_DATA.
-                    LOG_ERROR(getLogger("KeeperSnapshotManager"), "Children counter in stat.numChildren {}"
+                    LOG_ERROR(
-                                " is different from actual children size {} for node {}", itr.value.numChildren(), itr.value.getChildren().size(), itr.key);
+                        getLogger("KeeperSnapshotManager"),
+                        "Children counter in stat.numChildren {}"
+                        " is different from actual children size {} for node {}",
+                        itr.value.stats.numChildren(),
+                        itr.value.getChildren().size(),
+                        itr.key);
 #else
-                    throw Exception(ErrorCodes::LOGICAL_ERROR, "Children counter in stat.numChildren {}"
+                    throw Exception(
-                                    " is different from actual children size {} for node {}",
+                        ErrorCodes::LOGICAL_ERROR,
-                                    itr.value.numChildren(), itr.value.getChildren().size(), itr.key);
+                        "Children counter in stat.numChildren {}"
+                        " is different from actual children size {} for node {}",
+                        itr.value.stats.numChildren(),
+                        itr.value.getChildren().size(),
+                        itr.key);
 #endif
                 }
             }
@@ -511,7 +520,7 @@ void KeeperStorageSnapshot<Storage>::deserialize(SnapshotDeserializationResult<S
                 session_auth_counter++;
             }
             if (!ids.empty())
-                storage.session_and_auth[active_session_id] = ids;
+                storage.committed_session_and_auth[active_session_id] = ids;
         }
         current_session_size++;
     }
@@ -527,6 +536,8 @@ void KeeperStorageSnapshot<Storage>::deserialize(SnapshotDeserializationResult<S
         buffer->pos(0);
         deserialization_result.cluster_config = ClusterConfig::deserialize(*buffer);
     }
+
+    storage.updateStats();
 }
 
 template<typename Storage>
@@ -544,7 +555,7 @@ KeeperStorageSnapshot<Storage>::KeeperStorageSnapshot(Storage * storage_, uint64
     begin = storage->getSnapshotIteratorBegin();
     session_and_timeout = storage->getActiveSessions();
     acl_map = storage->acl_map.getMapping();
-    session_and_auth = storage->session_and_auth;
+    session_and_auth = storage->committed_session_and_auth;
 }
 
 template<typename Storage>
@@ -563,7 +574,7 @@ KeeperStorageSnapshot<Storage>::KeeperStorageSnapshot(
     begin = storage->getSnapshotIteratorBegin();
     session_and_timeout = storage->getActiveSessions();
     acl_map = storage->acl_map.getMapping();
-    session_and_auth = storage->session_and_auth;
+    session_and_auth = storage->committed_session_and_auth;
 }
 
 template<typename Storage>

src/Coordination/KeeperStateMachine.cpp

@@ -36,6 +36,11 @@ namespace ProfileEvents
     extern const Event KeeperStorageLockWaitMicroseconds;
 }
 
+namespace CurrentMetrics
+{
+    extern const Metric KeeperAliveConnections;
+}
+
 namespace DB
 {
 
@@ -56,6 +61,7 @@ IKeeperStateMachine::IKeeperStateMachine(
     , snapshots_queue(snapshots_queue_)
     , min_request_size_to_cache(keeper_context_->getCoordinationSettings()->min_request_size_for_cache)
     , log(getLogger("KeeperStateMachine"))
+    , read_pool(CurrentMetrics::KeeperAliveConnections, CurrentMetrics::KeeperAliveConnections, CurrentMetrics::KeeperAliveConnections, 100, 10000, 10000)
     , superdigest(superdigest_)
     , keeper_context(keeper_context_)
     , snapshot_manager_s3(snapshot_manager_s3_)
@@ -175,18 +181,20 @@ void assertDigest(
     }
 }
 
-struct TSA_SCOPED_LOCKABLE LockGuardWithStats final
+template <bool shared = false>
+struct LockGuardWithStats final
 {
-    std::unique_lock<std::mutex> lock;
+    using LockType = std::conditional_t<shared, std::shared_lock<SharedMutex>, std::unique_lock<SharedMutex>>;
-    explicit LockGuardWithStats(std::mutex & mutex) TSA_ACQUIRE(mutex)
+    LockType lock;
+    explicit LockGuardWithStats(SharedMutex & mutex)
     {
         Stopwatch watch;
-        std::unique_lock l(mutex);
+        LockType l(mutex);
         ProfileEvents::increment(ProfileEvents::KeeperStorageLockWaitMicroseconds, watch.elapsedMicroseconds());
         lock = std::move(l);
     }
 
-    ~LockGuardWithStats() TSA_RELEASE() = default;
+    ~LockGuardWithStats() = default;
 };
 
 }
@@ -312,13 +320,12 @@ bool KeeperStateMachine<Storage>::preprocess(const KeeperStorageBase::RequestFor
     if (op_num == Coordination::OpNum::SessionID || op_num == Coordination::OpNum::Reconfig)
         return true;
 
-    LockGuardWithStats lock(storage_and_responses_lock);
-
     if (storage->isFinalized())
         return false;
 
     try
     {
+        LockGuardWithStats<true> lock(storage_mutex);
         storage->preprocessRequest(
             request_for_session.request,
             request_for_session.session_id,
@@ -335,7 +342,12 @@ bool KeeperStateMachine<Storage>::preprocess(const KeeperStorageBase::RequestFor
     }
 
     if (keeper_context->digestEnabled() && request_for_session.digest)
-        assertDigest(*request_for_session.digest, storage->getNodesDigest(false), *request_for_session.request, request_for_session.log_idx, false);
+        assertDigest(
+            *request_for_session.digest,
+            storage->getNodesDigest(false, /*lock_transaction_mutex=*/true),
+            *request_for_session.request,
+            request_for_session.log_idx,
+            false);
 
     return true;
 }
@@ -343,7 +355,7 @@ bool KeeperStateMachine<Storage>::preprocess(const KeeperStorageBase::RequestFor
 template<typename Storage>
 void KeeperStateMachine<Storage>::reconfigure(const KeeperStorageBase::RequestForSession& request_for_session)
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     KeeperStorageBase::ResponseForSession response = processReconfiguration(request_for_session);
     if (!responses_queue.push(response))
     {
@@ -461,7 +473,7 @@ nuraft::ptr<nuraft::buffer> KeeperStateMachine<Storage>::commit(const uint64_t l
             response_for_session.response = response;
             response_for_session.request = request_for_session->request;
 
-            LockGuardWithStats lock(storage_and_responses_lock);
+            LockGuardWithStats lock(storage_mutex);
             session_id = storage->getSessionID(session_id_request.session_timeout_ms);
             LOG_DEBUG(log, "Session ID response {} with timeout {}", session_id, session_id_request.session_timeout_ms);
             response->session_id = session_id;
@@ -472,24 +484,31 @@ nuraft::ptr<nuraft::buffer> KeeperStateMachine<Storage>::commit(const uint64_t l
             if (op_num == Coordination::OpNum::Close)
 
             {
-                std::lock_guard lock(request_cache_mutex);
+                std::lock_guard cache_lock(request_cache_mutex);
                 parsed_request_cache.erase(request_for_session->session_id);
             }
 
-            LockGuardWithStats lock(storage_and_responses_lock);
-            KeeperStorageBase::ResponsesForSessions responses_for_sessions
-                = storage->processRequest(request_for_session->request, request_for_session->session_id, request_for_session->zxid);
-
-            for (auto & response_for_session : responses_for_sessions)
             {
-                if (response_for_session.response->xid != Coordination::WATCH_XID)
+                LockGuardWithStats<true> lock(storage_mutex);
-                    response_for_session.request = request_for_session->request;
+                std::lock_guard response_lock(process_and_responses_lock);
+                KeeperStorageBase::ResponsesForSessions responses_for_sessions
+                    = storage->processRequest(request_for_session->request, request_for_session->session_id, request_for_session->zxid);
+                for (auto & response_for_session : responses_for_sessions)
+                {
+                    if (response_for_session.response->xid != Coordination::WATCH_XID)
+                        response_for_session.request = request_for_session->request;
 
-                try_push(response_for_session);
+                    try_push(response_for_session);
+                }
             }
 
             if (keeper_context->digestEnabled() && request_for_session->digest)
-                assertDigest(*request_for_session->digest, storage->getNodesDigest(true), *request_for_session->request, request_for_session->log_idx, true);
+                assertDigest(
+                    *request_for_session->digest,
+                    storage->getNodesDigest(true, /*lock_transaction_mutex=*/true),
+                    *request_for_session->request,
+                    request_for_session->log_idx,
+                    true);
         }
 
         ProfileEvents::increment(ProfileEvents::KeeperCommits);
@@ -534,8 +553,6 @@ bool KeeperStateMachine<Storage>::apply_snapshot(nuraft::snapshot & s)
     }
 
     { /// deserialize and apply snapshot to storage
-        LockGuardWithStats lock(storage_and_responses_lock);
-
         SnapshotDeserializationResult<Storage> snapshot_deserialization_result;
         if (latest_snapshot_ptr)
             snapshot_deserialization_result = snapshot_manager.deserializeSnapshotFromBuffer(latest_snapshot_ptr);
@@ -543,6 +560,7 @@ bool KeeperStateMachine<Storage>::apply_snapshot(nuraft::snapshot & s)
             snapshot_deserialization_result
                 = snapshot_manager.deserializeSnapshotFromBuffer(snapshot_manager.deserializeSnapshotBufferFromDisk(s.get_last_log_idx()));
 
+        LockGuardWithStats storage_lock(storage_mutex);
         /// maybe some logs were preprocessed with log idx larger than the snapshot idx
         /// we have to apply them to the new storage
         storage->applyUncommittedState(*snapshot_deserialization_result.storage, snapshot_deserialization_result.snapshot_meta->get_last_log_idx());
@@ -587,16 +605,7 @@ void KeeperStateMachine<Storage>::rollbackRequest(const KeeperStorageBase::Reque
     if (request_for_session.request->getOpNum() == Coordination::OpNum::SessionID)
         return;
 
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
-    storage->rollbackRequest(request_for_session.zxid, allow_missing);
-}
-
-template<typename Storage>
-void KeeperStateMachine<Storage>::rollbackRequestNoLock(const KeeperStorageBase::RequestForSession & request_for_session, bool allow_missing)
-{
-    if (request_for_session.request->getOpNum() == Coordination::OpNum::SessionID)
-        return;
-
     storage->rollbackRequest(request_for_session.zxid, allow_missing);
 }
 
@@ -616,7 +625,7 @@ void KeeperStateMachine<Storage>::create_snapshot(nuraft::snapshot & s, nuraft::
     auto snapshot_meta_copy = nuraft::snapshot::deserialize(*snp_buf);
     CreateSnapshotTask snapshot_task;
     { /// lock storage for a short period time to turn on "snapshot mode". After that we can read consistent storage state without locking.
-        LockGuardWithStats lock(storage_and_responses_lock);
+        LockGuardWithStats lock(storage_mutex);
         snapshot_task.snapshot = std::make_shared<KeeperStorageSnapshot<Storage>>(storage.get(), snapshot_meta_copy, getClusterConfig());
     }
 
@@ -681,7 +690,7 @@ void KeeperStateMachine<Storage>::create_snapshot(nuraft::snapshot & s, nuraft::
         }
         {
             /// Destroy snapshot with lock
-            LockGuardWithStats lock(storage_and_responses_lock);
+            LockGuardWithStats lock(storage_mutex);
             LOG_TRACE(log, "Clearing garbage after snapshot");
             /// Turn off "snapshot mode" and clear outdate part of storage state
             storage->clearGarbageAfterSnapshot();
@@ -824,10 +833,10 @@ template<typename Storage>
 void KeeperStateMachine<Storage>::processReadRequest(const KeeperStorageBase::RequestForSession & request_for_session)
 {
     /// Pure local request, just process it with storage
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats<true> storage_lock(storage_mutex);
+    std::lock_guard response_lock(process_and_responses_lock);
     auto responses = storage->processRequest(
         request_for_session.request, request_for_session.session_id, std::nullopt, true /*check_acl*/, true /*is_local*/);
-
     for (auto & response_for_session : responses)
     {
         if (response_for_session.response->xid != Coordination::WATCH_XID)
@@ -840,112 +849,116 @@ void KeeperStateMachine<Storage>::processReadRequest(const KeeperStorageBase::Re
 template<typename Storage>
 void KeeperStateMachine<Storage>::shutdownStorage()
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     storage->finalize();
 }
 
 template<typename Storage>
 std::vector<int64_t> KeeperStateMachine<Storage>::getDeadSessions()
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getDeadSessions();
 }
 
 template<typename Storage>
 int64_t KeeperStateMachine<Storage>::getNextZxid() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
     return storage->getNextZXID();
 }
 
 template<typename Storage>
 KeeperStorageBase::Digest KeeperStateMachine<Storage>::getNodesDigest() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
-    return storage->getNodesDigest(false);
+    return storage->getNodesDigest(false, /*lock_transaction_mutex=*/true);
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getLastProcessedZxid() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
     return storage->getZXID();
 }
 
+template<typename Storage>
+const KeeperStorageBase::Stats & KeeperStateMachine<Storage>::getStorageStats() const TSA_NO_THREAD_SAFETY_ANALYSIS
+{
+    return storage->getStorageStats();
+}
+
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getNodesCount() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getNodesCount();
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getTotalWatchesCount() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getTotalWatchesCount();
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getWatchedPathsCount() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getWatchedPathsCount();
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getSessionsWithWatchesCount() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getSessionsWithWatchesCount();
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getTotalEphemeralNodesCount() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getTotalEphemeralNodesCount();
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getSessionWithEphemeralNodesCount() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getSessionWithEphemeralNodesCount();
 }
 
 template<typename Storage>
 void KeeperStateMachine<Storage>::dumpWatches(WriteBufferFromOwnString & buf) const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     storage->dumpWatches(buf);
 }
 
 template<typename Storage>
 void KeeperStateMachine<Storage>::dumpWatchesByPath(WriteBufferFromOwnString & buf) const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     storage->dumpWatchesByPath(buf);
 }
 
 template<typename Storage>
 void KeeperStateMachine<Storage>::dumpSessionsAndEphemerals(WriteBufferFromOwnString & buf) const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     storage->dumpSessionsAndEphemerals(buf);
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getApproximateDataSize() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getApproximateDataSize();
 }
 
 template<typename Storage>
 uint64_t KeeperStateMachine<Storage>::getKeyArenaSize() const
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     return storage->getArenaDataSize();
 }
 
@@ -988,7 +1001,7 @@ ClusterConfigPtr IKeeperStateMachine::getClusterConfig() const
 template<typename Storage>
 void KeeperStateMachine<Storage>::recalculateStorageStats()
 {
-    LockGuardWithStats lock(storage_and_responses_lock);
+    LockGuardWithStats lock(storage_mutex);
     LOG_INFO(log, "Recalculating storage stats");
     storage->recalculateStats();
     LOG_INFO(log, "Done recalculating storage stats");

src/Coordination/KeeperStateMachine.h

@@ -85,6 +85,8 @@ public:
     /// Introspection functions for 4lw commands
     virtual uint64_t getLastProcessedZxid() const = 0;
 
+    virtual const KeeperStorageBase::Stats & getStorageStats() const = 0;
+
     virtual uint64_t getNodesCount() const = 0;
     virtual uint64_t getTotalWatchesCount() const = 0;
     virtual uint64_t getWatchedPathsCount() const = 0;
@@ -124,12 +126,16 @@ protected:
     /// Mutex for snapshots
     mutable std::mutex snapshots_lock;
 
-    /// Lock for storage and responses_queue. It's important to process requests
+    /// Lock for the storage
+    /// Storage works in thread-safe way ONLY for preprocessing/processing
+    /// In any other case, unique storage lock needs to be taken
+    mutable SharedMutex storage_mutex;
+    /// Lock for processing and responses_queue. It's important to process requests
     /// and push them to the responses queue while holding this lock. Otherwise
     /// we can get strange cases when, for example client send read request with
     /// watch and after that receive watch response and only receive response
     /// for request.
-    mutable std::mutex storage_and_responses_lock;
+    mutable std::mutex process_and_responses_lock;
 
     std::unordered_map<int64_t, std::unordered_map<Coordination::XID, std::shared_ptr<KeeperStorageBase::RequestForSession>>> parsed_request_cache;
     uint64_t min_request_size_to_cache{0};
@@ -146,6 +152,7 @@ protected:
     mutable std::mutex cluster_config_lock;
     ClusterConfigPtr cluster_config;
 
+    ThreadPool read_pool;
     /// Special part of ACL system -- superdigest specified in server config.
     const std::string superdigest;
 
@@ -153,10 +160,8 @@ protected:
 
     KeeperSnapshotManagerS3 * snapshot_manager_s3;
 
-    virtual KeeperStorageBase::ResponseForSession processReconfiguration(
+    virtual KeeperStorageBase::ResponseForSession processReconfiguration(const KeeperStorageBase::RequestForSession & request_for_session)
-        const KeeperStorageBase::RequestForSession& request_for_session)
+        = 0;
-        TSA_REQUIRES(storage_and_responses_lock) = 0;
-
 };
 
 /// ClickHouse Keeper state machine. Wrapper for KeeperStorage.
@@ -189,10 +194,6 @@ public:
     // (can happen in case of exception during preprocessing)
     void rollbackRequest(const KeeperStorageBase::RequestForSession & request_for_session, bool allow_missing) override;
 
-    void rollbackRequestNoLock(
-        const KeeperStorageBase::RequestForSession & request_for_session,
-        bool allow_missing) TSA_NO_THREAD_SAFETY_ANALYSIS;
-
     /// Apply preliminarily saved (save_logical_snp_obj) snapshot to our state.
     bool apply_snapshot(nuraft::snapshot & s) override;
 
@@ -205,7 +206,7 @@ public:
     // This should be used only for tests or keeper-data-dumper because it violates
     // TSA -- we can't acquire the lock outside of this class or return a storage under lock
     // in a reasonable way.
-    Storage & getStorageUnsafe() TSA_NO_THREAD_SAFETY_ANALYSIS
+    Storage & getStorageUnsafe()
     {
         return *storage;
     }
@@ -224,6 +225,8 @@ public:
     /// Introspection functions for 4lw commands
     uint64_t getLastProcessedZxid() const override;
 
+    const KeeperStorageBase::Stats & getStorageStats() const override;
+
     uint64_t getNodesCount() const override;
     uint64_t getTotalWatchesCount() const override;
     uint64_t getWatchedPathsCount() const override;
@@ -245,12 +248,12 @@ public:
 
 private:
     /// Main state machine logic
-    std::unique_ptr<Storage> storage; //TSA_PT_GUARDED_BY(storage_and_responses_lock);
+    std::unique_ptr<Storage> storage;
 
     /// Save/Load and Serialize/Deserialize logic for snapshots.
     KeeperSnapshotManager<Storage> snapshot_manager;
 
-    KeeperStorageBase::ResponseForSession processReconfiguration(const KeeperStorageBase::RequestForSession & request_for_session)
+    KeeperStorageBase::ResponseForSession processReconfiguration(const KeeperStorageBase::RequestForSession & request_for_session) override;
-        TSA_REQUIRES(storage_and_responses_lock) override;
 };
+
 }

src/Coordination/KeeperStorage.h

@@ -1,10 +1,16 @@
 #pragma once
 
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 #include <Coordination/ACLMap.h>
 #include <Coordination/SessionExpiryQueue.h>
 #include <Coordination/SnapshotableHashTable.h>
+#include "Common/StringHashForHeterogeneousLookup.h"
+#include <Common/SharedMutex.h>
+#include <Common/Concepts.h>
+
+#include <base/defines.h>
 
 #include <absl/container/flat_hash_set.h>
 
@@ -23,14 +29,11 @@ using ResponseCallback = std::function<void(const Coordination::ZooKeeperRespons
 using ChildrenSet = absl::flat_hash_set<StringRef, StringRefHash>;
 using SessionAndTimeout = std::unordered_map<int64_t, int64_t>;
 
-/// KeeperRocksNodeInfo is used in RocksDB keeper.
+struct NodeStats
-/// It is serialized directly as POD to RocksDB.
-struct KeeperRocksNodeInfo
 {
     int64_t czxid{0};
     int64_t mzxid{0};
     int64_t pzxid{0};
-    uint64_t acl_id = 0; /// 0 -- no ACL by default
 
     int64_t mtime{0};
 
@@ -38,225 +41,9 @@ struct KeeperRocksNodeInfo
     int32_t cversion{0};
     int32_t aversion{0};
 
-    int32_t seq_num = 0;
+    uint32_t data_size{0};
-    mutable UInt64 digest = 0; /// we cached digest for this node.
-
-    /// as ctime can't be negative because it stores the timestamp when the
-    /// node was created, we can use the MSB for a bool
-    struct
-    {
-        bool is_ephemeral : 1;
-        int64_t ctime : 63;
-    } is_ephemeral_and_ctime{false, 0};
-
-    /// ephemeral notes cannot have children so a node can set either
-    /// ephemeral_owner OR seq_num + num_children
-    union
-    {
-        int64_t ephemeral_owner;
-        struct
-        {
-            int32_t seq_num;
-            int32_t num_children;
-        } children_info;
-    } ephemeral_or_children_data{0};
-
-    bool isEphemeral() const
-    {
-        return is_ephemeral_and_ctime.is_ephemeral;
-    }
-
-    int64_t ephemeralOwner() const
-    {
-        if (isEphemeral())
-            return ephemeral_or_children_data.ephemeral_owner;
-
-        return 0;
-    }
-
-    void setEphemeralOwner(int64_t ephemeral_owner)
-    {
-        is_ephemeral_and_ctime.is_ephemeral = ephemeral_owner != 0;
-        ephemeral_or_children_data.ephemeral_owner = ephemeral_owner;
-    }
-
-    int32_t numChildren() const
-    {
-        if (isEphemeral())
-            return 0;
-
-        return ephemeral_or_children_data.children_info.num_children;
-    }
-
-    void setNumChildren(int32_t num_children)
-    {
-        ephemeral_or_children_data.children_info.num_children = num_children;
-    }
-
-    /// dummy interface for test
-    void addChild(StringRef) {}
-    auto getChildren() const
-    {
-        return std::vector<int>(numChildren());
-    }
-
-    void increaseNumChildren()
-    {
-        chassert(!isEphemeral());
-        ++ephemeral_or_children_data.children_info.num_children;
-    }
-
-    void decreaseNumChildren()
-    {
-        chassert(!isEphemeral());
-        --ephemeral_or_children_data.children_info.num_children;
-    }
-
-    int32_t seqNum() const
-    {
-        if (isEphemeral())
-            return 0;
-
-        return ephemeral_or_children_data.children_info.seq_num;
-    }
-
-    void setSeqNum(int32_t seq_num_)
-    {
-        ephemeral_or_children_data.children_info.seq_num = seq_num_;
-    }
-
-    void increaseSeqNum()
-    {
-        chassert(!isEphemeral());
-        ++ephemeral_or_children_data.children_info.seq_num;
-    }
-
-    int64_t ctime() const
-    {
-        return is_ephemeral_and_ctime.ctime;
-    }
-
-    void setCtime(uint64_t ctime)
-    {
-        is_ephemeral_and_ctime.ctime = ctime;
-    }
 
     void copyStats(const Coordination::Stat & stat);
-};
-
-/// KeeperRocksNode is the memory structure used by RocksDB
-struct KeeperRocksNode : public KeeperRocksNodeInfo
-{
-#if USE_ROCKSDB
-    friend struct RocksDBContainer<KeeperRocksNode>;
-#endif
-    using Meta = KeeperRocksNodeInfo;
-
-    uint64_t size_bytes = 0; // only for compatible, should be deprecated
-
-    uint64_t sizeInBytes() const { return data_size + sizeof(KeeperRocksNodeInfo); }
-    void setData(String new_data)
-    {
-        data_size = static_cast<uint32_t>(new_data.size());
-        if (data_size != 0)
-        {
-            data = std::unique_ptr<char[]>(new char[new_data.size()]);
-            memcpy(data.get(), new_data.data(), data_size);
-        }
-    }
-
-    void shallowCopy(const KeeperRocksNode & other)
-    {
-        czxid = other.czxid;
-        mzxid = other.mzxid;
-        pzxid = other.pzxid;
-        acl_id = other.acl_id; /// 0 -- no ACL by default
-
-        mtime = other.mtime;
-
-        is_ephemeral_and_ctime = other.is_ephemeral_and_ctime;
-
-        ephemeral_or_children_data = other.ephemeral_or_children_data;
-
-        data_size = other.data_size;
-        if (data_size != 0)
-        {
-            data = std::unique_ptr<char[]>(new char[data_size]);
-            memcpy(data.get(), other.data.get(), data_size);
-        }
-
-        version = other.version;
-        cversion = other.cversion;
-        aversion = other.aversion;
-
-        /// cached_digest = other.cached_digest;
-    }
-    void invalidateDigestCache() const;
-    UInt64 getDigest(std::string_view path) const;
-    String getEncodedString();
-    void decodeFromString(const String & buffer_str);
-    void recalculateSize() {}
-    std::string_view getData() const noexcept { return {data.get(), data_size}; }
-
-    void setResponseStat(Coordination::Stat & response_stat) const
-    {
-        response_stat.czxid = czxid;
-        response_stat.mzxid = mzxid;
-        response_stat.ctime = ctime();
-        response_stat.mtime = mtime;
-        response_stat.version = version;
-        response_stat.cversion = cversion;
-        response_stat.aversion = aversion;
-        response_stat.ephemeralOwner = ephemeralOwner();
-        response_stat.dataLength = static_cast<int32_t>(data_size);
-        response_stat.numChildren = numChildren();
-        response_stat.pzxid = pzxid;
-    }
-
-    void reset()
-    {
-        serialized = false;
-    }
-    bool empty() const
-    {
-        return data_size == 0 && mzxid == 0;
-    }
-    std::unique_ptr<char[]> data{nullptr};
-    uint32_t data_size{0};
-private:
-    bool serialized = false;
-};
-
-/// KeeperMemNode should have as minimal size as possible to reduce memory footprint
-/// of stored nodes
-/// New fields should be added to the struct only if it's really necessary
-struct KeeperMemNode
-{
-    int64_t czxid{0};
-    int64_t mzxid{0};
-    int64_t pzxid{0};
-    uint64_t acl_id = 0; /// 0 -- no ACL by default
-
-    int64_t mtime{0};
-
-    std::unique_ptr<char[]> data{nullptr};
-    uint32_t data_size{0};
-
-    int32_t version{0};
-    int32_t cversion{0};
-    int32_t aversion{0};
-
-    mutable uint64_t cached_digest = 0;
-
-    KeeperMemNode() = default;
-
-    KeeperMemNode & operator=(const KeeperMemNode & other);
-    KeeperMemNode(const KeeperMemNode & other);
-
-    KeeperMemNode & operator=(KeeperMemNode && other) noexcept;
-    KeeperMemNode(KeeperMemNode && other) noexcept;
-
-    bool empty() const;
 
     bool isEphemeral() const
     {
@@ -287,6 +74,7 @@ struct KeeperMemNode
 
     void setNumChildren(int32_t num_children)
     {
+        is_ephemeral_and_ctime.is_ephemeral = false;
         ephemeral_or_children_data.children_info.num_children = num_children;
     }
 
@@ -331,34 +119,6 @@ struct KeeperMemNode
         is_ephemeral_and_ctime.ctime = ctime;
     }
 
-    void copyStats(const Coordination::Stat & stat);
-
-    void setResponseStat(Coordination::Stat & response_stat) const;
-
-    /// Object memory size
-    uint64_t sizeInBytes() const;
-
-    void setData(const String & new_data);
-
-    std::string_view getData() const noexcept { return {data.get(), data_size}; }
-
-    void addChild(StringRef child_path);
-
-    void removeChild(StringRef child_path);
-
-    const auto & getChildren() const noexcept { return children; }
-    auto & getChildren() { return children; }
-
-    // Invalidate the calculated digest so it's recalculated again on the next
-    // getDigest call
-    void invalidateDigestCache() const;
-
-    // get the calculated digest of the node
-    UInt64 getDigest(std::string_view path) const;
-
-    // copy only necessary information for preprocessing and digest calculation
-    // (e.g. we don't need to copy list of children)
-    void shallowCopy(const KeeperMemNode & other);
 private:
     /// as ctime can't be negative because it stores the timestamp when the
     /// node was created, we can use the MSB for a bool
@@ -379,7 +139,132 @@ private:
             int32_t num_children;
         } children_info;
     } ephemeral_or_children_data{0};
+};
 
+/// KeeperRocksNodeInfo is used in RocksDB keeper.
+/// It is serialized directly as POD to RocksDB.
+struct KeeperRocksNodeInfo
+{
+    NodeStats stats;
+    uint64_t acl_id = 0; /// 0 -- no ACL by default
+
+    /// dummy interface for test
+    void addChild(StringRef) {}
+    auto getChildren() const
+    {
+        return std::vector<int>(stats.numChildren());
+    }
+
+    void copyStats(const Coordination::Stat & stat);
+};
+
+/// KeeperRocksNode is the memory structure used by RocksDB
+struct KeeperRocksNode : public KeeperRocksNodeInfo
+{
+#if USE_ROCKSDB
+    friend struct RocksDBContainer<KeeperRocksNode>;
+#endif
+    using Meta = KeeperRocksNodeInfo;
+
+    uint64_t size_bytes = 0; // only for compatible, should be deprecated
+
+    uint64_t sizeInBytes() const { return stats.data_size + sizeof(KeeperRocksNodeInfo); }
+
+    void setData(String new_data)
+    {
+        stats.data_size = static_cast<uint32_t>(new_data.size());
+        if (stats.data_size != 0)
+        {
+            data = std::unique_ptr<char[]>(new char[new_data.size()]);
+            memcpy(data.get(), new_data.data(), stats.data_size);
+        }
+    }
+
+    void shallowCopy(const KeeperRocksNode & other)
+    {
+        stats = other.stats;
+        acl_id = other.acl_id;
+        if (stats.data_size != 0)
+        {
+            data = std::unique_ptr<char[]>(new char[stats.data_size]);
+            memcpy(data.get(), other.data.get(), stats.data_size);
+        }
+
+        /// cached_digest = other.cached_digest;
+    }
+    void invalidateDigestCache() const;
+    UInt64 getDigest(std::string_view path) const;
+    String getEncodedString();
+    void decodeFromString(const String & buffer_str);
+    void recalculateSize() {}
+    std::string_view getData() const noexcept { return {data.get(), stats.data_size}; }
+
+    void setResponseStat(Coordination::Stat & response_stat) const;
+
+    void reset()
+    {
+        serialized = false;
+    }
+    bool empty() const
+    {
+        return stats.data_size == 0 && stats.mzxid == 0;
+    }
+    std::unique_ptr<char[]> data{nullptr};
+    mutable UInt64 cached_digest = 0; /// we cached digest for this node.
+private:
+    bool serialized = false;
+};
+
+/// KeeperMemNode should have as minimal size as possible to reduce memory footprint
+/// of stored nodes
+/// New fields should be added to the struct only if it's really necessary
+struct KeeperMemNode
+{
+    NodeStats stats;
+    std::unique_ptr<char[]> data{nullptr};
+    mutable uint64_t cached_digest = 0;
+
+    uint64_t acl_id = 0; /// 0 -- no ACL by default
+
+    KeeperMemNode() = default;
+
+    KeeperMemNode & operator=(const KeeperMemNode & other);
+    KeeperMemNode(const KeeperMemNode & other);
+
+    KeeperMemNode & operator=(KeeperMemNode && other) noexcept;
+    KeeperMemNode(KeeperMemNode && other) noexcept;
+
+    bool empty() const;
+
+    void copyStats(const Coordination::Stat & stat);
+
+    void setResponseStat(Coordination::Stat & response_stat) const;
+
+    /// Object memory size
+    uint64_t sizeInBytes() const;
+
+    void setData(const String & new_data);
+
+    std::string_view getData() const noexcept { return {data.get(), stats.data_size}; }
+
+    void addChild(StringRef child_path);
+
+    void removeChild(StringRef child_path);
+
+    const auto & getChildren() const noexcept { return children; }
+    auto & getChildren() { return children; }
+
+    // Invalidate the calculated digest so it's recalculated again on the next
+    // getDigest call
+    void invalidateDigestCache() const;
+
+    // get the calculated digest of the node
+    UInt64 getDigest(std::string_view path) const;
+
+    // copy only necessary information for preprocessing and digest calculation
+    // (e.g. we don't need to copy list of children)
+    void shallowCopy(const KeeperMemNode & other);
+private:
     ChildrenSet children{};
 };
 
@@ -430,18 +315,187 @@ public:
     };
 
     using Ephemerals = std::unordered_map<int64_t, std::unordered_set<std::string>>;
-    using SessionAndWatcher = std::unordered_map<int64_t, std::unordered_set<std::string>>;
+    struct WatchInfo
+    {
+        std::string_view path;
+        bool is_list_watch;
+
+        bool operator==(const WatchInfo &) const = default;
+    };
+
+    struct WatchInfoHash
+    {
+        auto operator()(WatchInfo info) const
+        {
+            SipHash hash;
+            hash.update(info.path);
+            hash.update(info.is_list_watch);
+            return hash.get64();
+        }
+    };
+
+    using SessionAndWatcher = std::unordered_map<int64_t, std::unordered_set<WatchInfo, WatchInfoHash>>;
     using SessionIDs = std::unordered_set<int64_t>;
 
     /// Just vector of SHA1 from user:password
     using AuthIDs = std::vector<AuthID>;
     using SessionAndAuth = std::unordered_map<int64_t, AuthIDs>;
-    using Watches = std::unordered_map<String /* path, relative of root_path */, SessionIDs>;
+    using Watches = std::unordered_map<
+        String /* path, relative of root_path */,
+        SessionIDs,
+        StringHashForHeterogeneousLookup,
+        StringHashForHeterogeneousLookup::transparent_key_equal>;
+
+    // Applying ZooKeeper request to storage consists of two steps:
+    //  - preprocessing which, instead of applying the changes directly to storage,
+    //    generates deltas with those changes, denoted with the request ZXID
+    //  - processing which applies deltas with the correct ZXID to the storage
+    //
+    // Delta objects allow us two things:
+    //  - fetch the latest, uncommitted state of an object by getting the committed
+    //    state of that same object from the storage and applying the deltas
+    //    in the same order as they are defined
+    //  - quickly commit the changes to the storage
+    struct CreateNodeDelta
+    {
+        Coordination::Stat stat;
+        Coordination::ACLs acls;
+        String data;
+    };
+
+    struct RemoveNodeDelta
+    {
+        int32_t version{-1};
+        NodeStats stat;
+        Coordination::ACLs acls;
+        String data;
+    };
+
+    struct UpdateNodeStatDelta
+    {
+        template <is_any_of<KeeperMemNode, KeeperRocksNode> Node>
+        explicit UpdateNodeStatDelta(const Node & node)
+            : old_stats(node.stats)
+            , new_stats(node.stats)
+        {}
+
+        NodeStats old_stats;
+        NodeStats new_stats;
+        int32_t version{-1};
+    };
+
+    struct UpdateNodeDataDelta
+    {
+
+        std::string old_data;
+        std::string new_data;
+        int32_t version{-1};
+    };
+
+    struct SetACLDelta
+    {
+        Coordination::ACLs old_acls;
+        Coordination::ACLs new_acls;
+        int32_t version{-1};
+    };
+
+    struct ErrorDelta
+    {
+        Coordination::Error error;
+    };
+
+    struct FailedMultiDelta
+    {
+        std::vector<Coordination::Error> error_codes;
+        Coordination::Error global_error{Coordination::Error::ZOK};
+    };
+
+    // Denotes end of a subrequest in multi request
+    struct SubDeltaEnd
+    {
+    };
+
+    struct AddAuthDelta
+    {
+        int64_t session_id;
+        std::shared_ptr<AuthID> auth_id;
+    };
+
+    struct CloseSessionDelta
+    {
+        int64_t session_id;
+    };
+
+    using Operation = std::variant<
+        CreateNodeDelta,
+        RemoveNodeDelta,
+        UpdateNodeStatDelta,
+        UpdateNodeDataDelta,
+        SetACLDelta,
+        AddAuthDelta,
+        ErrorDelta,
+        SubDeltaEnd,
+        FailedMultiDelta,
+        CloseSessionDelta>;
+
+    struct Delta
+    {
+        Delta(String path_, int64_t zxid_, Operation operation_) : path(std::move(path_)), zxid(zxid_), operation(std::move(operation_)) { }
+
+        Delta(int64_t zxid_, Coordination::Error error) : Delta("", zxid_, ErrorDelta{error}) { }
+
+        Delta(int64_t zxid_, Operation subdelta) : Delta("", zxid_, subdelta) { }
+
+        String path;
+        int64_t zxid;
+        Operation operation;
+    };
+
+    using DeltaIterator = std::list<KeeperStorageBase::Delta>::const_iterator;
+    struct DeltaRange
+    {
+        DeltaIterator begin_it;
+        DeltaIterator end_it;
+
+        auto begin() const
+        {
+            return begin_it;
+        }
+
+        auto end() const
+        {
+            return end_it;
+        }
+
+        bool empty() const
+        {
+            return begin_it == end_it;
+        }
+
+        const auto & front() const
+        {
+            return *begin_it;
+        }
+    };
+
+    struct Stats
+    {
+        std::atomic<uint64_t> nodes_count = 0;
+        std::atomic<uint64_t> approximate_data_size = 0;
+        std::atomic<uint64_t> total_watches_count = 0;
+        std::atomic<uint64_t> watched_paths_count = 0;
+        std::atomic<uint64_t> sessions_with_watches_count = 0;
+        std::atomic<uint64_t> session_with_ephemeral_nodes_count = 0;
+        std::atomic<uint64_t> total_emphemeral_nodes_count = 0;
+        std::atomic<int64_t> last_zxid = 0;
+    };
+
+    Stats stats;
 
     static bool checkDigest(const Digest & first, const Digest & second);
-
 };
 
+
 /// Keeper state machine almost equal to the ZooKeeper's state machine.
 /// Implements all logic of operations, data changes, sessions allocation.
 /// In-memory and not thread safe.
@@ -472,160 +526,73 @@ public:
 
     int64_t session_id_counter{1};
 
-    SessionAndAuth session_and_auth;
+    mutable SharedMutex auth_mutex;
+    SessionAndAuth committed_session_and_auth;
 
+    mutable SharedMutex storage_mutex;
     /// Main hashtable with nodes. Contain all information about data.
     /// All other structures expect session_and_timeout can be restored from
     /// container.
     Container container;
 
-    // Applying ZooKeeper request to storage consists of two steps:
-    //  - preprocessing which, instead of applying the changes directly to storage,
-    //    generates deltas with those changes, denoted with the request ZXID
-    //  - processing which applies deltas with the correct ZXID to the storage
-    //
-    // Delta objects allow us two things:
-    //  - fetch the latest, uncommitted state of an object by getting the committed
-    //    state of that same object from the storage and applying the deltas
-    //    in the same order as they are defined
-    //  - quickly commit the changes to the storage
-    struct CreateNodeDelta
-    {
-        Coordination::Stat stat;
-        Coordination::ACLs acls;
-        String data;
-    };
-
-    struct RemoveNodeDelta
-    {
-        int32_t version{-1};
-        int64_t ephemeral_owner{0};
-    };
-
-    struct UpdateNodeDelta
-    {
-        std::function<void(Node &)> update_fn;
-        int32_t version{-1};
-    };
-
-    struct SetACLDelta
-    {
-        Coordination::ACLs acls;
-        int32_t version{-1};
-    };
-
-    struct ErrorDelta
-    {
-        Coordination::Error error;
-    };
-
-    struct FailedMultiDelta
-    {
-        std::vector<Coordination::Error> error_codes;
-        Coordination::Error global_error{Coordination::Error::ZOK};
-    };
-
-    // Denotes end of a subrequest in multi request
-    struct SubDeltaEnd
-    {
-    };
-
-    struct AddAuthDelta
-    {
-        int64_t session_id;
-        AuthID auth_id;
-    };
-
-    struct CloseSessionDelta
-    {
-        int64_t session_id;
-    };
-
-    using Operation = std::
-        variant<CreateNodeDelta, RemoveNodeDelta, UpdateNodeDelta, SetACLDelta, AddAuthDelta, ErrorDelta, SubDeltaEnd, FailedMultiDelta, CloseSessionDelta>;
-
-    struct Delta
-    {
-        Delta(String path_, int64_t zxid_, Operation operation_) : path(std::move(path_)), zxid(zxid_), operation(std::move(operation_)) { }
-
-        Delta(int64_t zxid_, Coordination::Error error) : Delta("", zxid_, ErrorDelta{error}) { }
-
-        Delta(int64_t zxid_, Operation subdelta) : Delta("", zxid_, subdelta) { }
-
-        String path;
-        int64_t zxid;
-        Operation operation;
-    };
-
     struct UncommittedState
     {
         explicit UncommittedState(KeeperStorage & storage_) : storage(storage_) { }
 
-        void addDelta(Delta new_delta);
+        void addDeltas(std::list<Delta> new_deltas);
-        void addDeltas(std::vector<Delta> new_deltas);
+        void cleanup(int64_t commit_zxid);
-        void commit(int64_t commit_zxid);
         void rollback(int64_t rollback_zxid);
+        void rollback(std::list<Delta> rollback_deltas);
 
-        std::shared_ptr<Node> getNode(StringRef path) const;
+        std::shared_ptr<Node> getNode(StringRef path, bool should_lock_storage = true) const;
         const Node * getActualNodeView(StringRef path, const Node & storage_node) const;
+
         Coordination::ACLs getACLs(StringRef path) const;
 
+        void applyDeltas(const std::list<Delta> & new_deltas);
         void applyDelta(const Delta & delta);
+        void rollbackDelta(const Delta & delta);
 
         bool hasACL(int64_t session_id, bool is_local, std::function<bool(const AuthID &)> predicate) const;
 
         void forEachAuthInSession(int64_t session_id, std::function<void(const AuthID &)> func) const;
 
-        std::shared_ptr<Node> tryGetNodeFromStorage(StringRef path) const;
+        std::shared_ptr<Node> tryGetNodeFromStorage(StringRef path, bool should_lock_storage = true) const;
 
-        std::unordered_map<int64_t, std::list<const AuthID *>> session_and_auth;
         std::unordered_set<int64_t> closed_sessions;
 
+        using ZxidToNodes = std::map<int64_t, std::unordered_set<std::string_view>>;
         struct UncommittedNode
         {
             std::shared_ptr<Node> node{nullptr};
-            Coordination::ACLs acls{};
+            std::optional<Coordination::ACLs> acls{};
-            int64_t zxid{0};
+            std::unordered_set<uint64_t> applied_zxids{};
-        };
 
-        struct Hash
+            void materializeACL(const ACLMap & current_acl_map);
-        {
-            auto operator()(const std::string_view view) const
-            {
-                SipHash hash;
-                hash.update(view);
-                return hash.get64();
-            }
-
-            using is_transparent = void; // required to make find() work with different type than key_type
-        };
-
-        struct Equal
-        {
-            auto operator()(const std::string_view a,
-                            const std::string_view b) const
-            {
-                return a == b;
-            }
-
-            using is_transparent = void; // required to make find() work with different type than key_type
         };
 
         struct PathCmp
         {
-            using is_transparent = std::true_type;
-
             auto operator()(const std::string_view a,
                             const std::string_view b) const
             {
-                return a.size() < b.size() || (a.size() == b.size() && a < b);
+                size_t level_a = std::count(a.begin(), a.end(), '/');
+                size_t level_b = std::count(b.begin(), b.end(), '/');
+                return level_a < level_b || (level_a == level_b && a < b);
             }
+
+            using is_transparent = void; // required to make find() work with different type than key_type
         };
 
-        mutable std::map<std::string, UncommittedNode, PathCmp> nodes;
+        Ephemerals ephemerals;
-        std::unordered_map<std::string, std::list<const Delta *>, Hash, Equal> deltas_for_path;
 
-        std::list<Delta> deltas;
+        std::unordered_map<int64_t, std::list<std::pair<int64_t, std::shared_ptr<AuthID>>>> session_and_auth;
+
+        mutable std::map<std::string, UncommittedNode, PathCmp> nodes;
+        mutable ZxidToNodes zxid_to_nodes;
+
+        mutable std::mutex deltas_mutex;
+        std::list<Delta> deltas TSA_GUARDED_BY(deltas_mutex);
         KeeperStorage<Container> & storage;
     };
 
@@ -635,7 +602,7 @@ public:
     // with zxid > last_zxid
     void applyUncommittedState(KeeperStorage & other, int64_t last_log_idx);
 
-    Coordination::Error commit(int64_t zxid);
+    Coordination::Error commit(DeltaRange deltas);
 
     // Create node in the storage
     // Returns false if it failed to create the node, true otherwise
@@ -653,12 +620,11 @@ public:
 
     bool checkACL(StringRef path, int32_t permissions, int64_t session_id, bool is_local);
 
-    void unregisterEphemeralPath(int64_t session_id, const std::string & path);
+    std::mutex ephemeral_mutex;
-
     /// Mapping session_id -> set of ephemeral nodes paths
-    Ephemerals ephemerals;
+    Ephemerals committed_ephemerals;
-    /// Mapping session_id -> set of watched nodes paths
+    size_t committed_ephemeral_nodes{0};
-    SessionAndWatcher sessions_and_watchers;
+
     /// Expiration queue for session, allows to get dead sessions at some point of time
     SessionExpiryQueue session_expiry_queue;
     /// All active sessions with timeout
@@ -667,8 +633,10 @@ public:
     /// ACLMap for more compact ACLs storage inside nodes.
     ACLMap acl_map;
 
+    mutable std::mutex transaction_mutex;
+
     /// Global id of all requests applied to storage
-    int64_t zxid{0};
+    int64_t zxid TSA_GUARDED_BY(transaction_mutex) = 0;
 
     // older Keeper node (pre V5 snapshots) can create snapshots and receive logs from newer Keeper nodes
     // this can lead to some inconsistencies, e.g. from snapshot it will use log_idx as zxid
@@ -685,11 +653,16 @@ public:
         int64_t log_idx = 0;
     };
 
-    std::deque<TransactionInfo> uncommitted_transactions;
+    std::list<TransactionInfo> uncommitted_transactions TSA_GUARDED_BY(transaction_mutex);
 
-    uint64_t nodes_digest{0};
+    uint64_t nodes_digest = 0;
 
-    bool finalized{false};
+    std::atomic<bool> finalized{false};
+
+
+    /// Mapping session_id -> set of watched nodes paths
+    SessionAndWatcher sessions_and_watchers;
+    size_t total_watches_count = 0;
 
     /// Currently active watches (node_path -> subscribed sessions)
     Watches watches;
@@ -698,45 +671,30 @@ public:
     void clearDeadWatches(int64_t session_id);
 
     /// Get current committed zxid
-    int64_t getZXID() const { return zxid; }
+    int64_t getZXID() const;
 
-    int64_t getNextZXID() const
+    int64_t getNextZXID() const;
-    {
+    int64_t getNextZXIDLocked() const TSA_REQUIRES(transaction_mutex);
-        if (uncommitted_transactions.empty())
-            return zxid + 1;
 
-        return uncommitted_transactions.back().zxid + 1;
+    Digest getNodesDigest(bool committed, bool lock_transaction_mutex) const;
-    }
-
-    Digest getNodesDigest(bool committed) const;
 
     KeeperContextPtr keeper_context;
 
     const String superdigest;
 
-    bool initialized{false};
+    std::atomic<bool> initialized{false};
 
     KeeperStorage(int64_t tick_time_ms, const String & superdigest_, const KeeperContextPtr & keeper_context_, bool initialize_system_nodes = true);
 
-    void initializeSystemNodes();
+    void initializeSystemNodes() TSA_NO_THREAD_SAFETY_ANALYSIS;
 
     /// Allocate new session id with the specified timeouts
-    int64_t getSessionID(int64_t session_timeout_ms)
+    int64_t getSessionID(int64_t session_timeout_ms);
-    {
-        auto result = session_id_counter++;
-        session_and_timeout.emplace(result, session_timeout_ms);
-        session_expiry_queue.addNewSessionOrUpdate(result, session_timeout_ms);
-        return result;
-    }
 
     /// Add session id. Used when restoring KeeperStorage from snapshot.
-    void addSessionID(int64_t session_id, int64_t session_timeout_ms)
+    void addSessionID(int64_t session_id, int64_t session_timeout_ms) TSA_NO_THREAD_SAFETY_ANALYSIS;
-    {
-        session_and_timeout.emplace(session_id, session_timeout_ms);
-        session_expiry_queue.addNewSessionOrUpdate(session_id, session_timeout_ms);
-    }
 
-    UInt64 calculateNodesDigest(UInt64 current_digest, const std::vector<Delta> & new_deltas) const;
+    UInt64 calculateNodesDigest(UInt64 current_digest, const std::list<Delta> & new_deltas) const;
 
     /// Process user request and return response.
     /// check_acl = false only when converting data from ZooKeeper.
@@ -763,42 +721,39 @@ public:
     /// Set of methods for creating snapshots
 
     /// Turn on snapshot mode, so data inside Container is not deleted, but replaced with new version.
-    void enableSnapshotMode(size_t up_to_version)
+    void enableSnapshotMode(size_t up_to_version);
-    {
-        container.enableSnapshotMode(up_to_version);
-    }
 
     /// Turn off snapshot mode.
-    void disableSnapshotMode()
+    void disableSnapshotMode();
-    {
-        container.disableSnapshotMode();
-    }
 
-    Container::const_iterator getSnapshotIteratorBegin() const { return container.begin(); }
+    Container::const_iterator getSnapshotIteratorBegin() const;
 
     /// Clear outdated data from internal container.
-    void clearGarbageAfterSnapshot() { container.clearOutdatedNodes(); }
+    void clearGarbageAfterSnapshot();
 
     /// Get all active sessions
-    const SessionAndTimeout & getActiveSessions() const { return session_and_timeout; }
+    SessionAndTimeout getActiveSessions() const;
 
     /// Get all dead sessions
-    std::vector<int64_t> getDeadSessions() const { return session_expiry_queue.getExpiredSessions(); }
+    std::vector<int64_t> getDeadSessions() const;
+
+    void updateStats();
+    const Stats & getStorageStats() const;
 
     /// Introspection functions mostly used in 4-letter commands
-    uint64_t getNodesCount() const { return container.size(); }
+    uint64_t getNodesCount() const;
 
-    uint64_t getApproximateDataSize() const { return container.getApproximateDataSize(); }
+    uint64_t getApproximateDataSize() const;
 
-    uint64_t getArenaDataSize() const { return container.keyArenaSize(); }
+    uint64_t getArenaDataSize() const;
 
     uint64_t getTotalWatchesCount() const;
 
-    uint64_t getWatchedPathsCount() const { return watches.size() + list_watches.size(); }
+    uint64_t getWatchedPathsCount() const;
 
     uint64_t getSessionsWithWatchesCount() const;
 
-    uint64_t getSessionWithEphemeralNodesCount() const { return ephemerals.size(); }
+    uint64_t getSessionWithEphemeralNodesCount() const;
     uint64_t getTotalEphemeralNodesCount() const;
 
     void dumpWatches(WriteBufferFromOwnString & buf) const;

src/Coordination/RocksDBContainer.h

@@ -155,11 +155,11 @@ public:
                 ReadBufferFromOwnString buffer(iter->value().ToStringView());
                 typename Node::Meta & meta = new_pair->value;
                 readPODBinary(meta, buffer);
-                readVarUInt(new_pair->value.data_size, buffer);
+                readVarUInt(new_pair->value.stats.data_size, buffer);
-                if (new_pair->value.data_size)
+                if (new_pair->value.stats.data_size)
                 {
-                    new_pair->value.data = std::unique_ptr<char[]>(new char[new_pair->value.data_size]);
+                    new_pair->value.data = std::unique_ptr<char[]>(new char[new_pair->value.stats.data_size]);
-                    buffer.readStrict(new_pair->value.data.get(), new_pair->value.data_size);
+                    buffer.readStrict(new_pair->value.data.get(), new_pair->value.stats.data_size);
                 }
                 pair = new_pair;
             }
@@ -211,7 +211,7 @@ public:
         }
     }
 
-    std::vector<std::pair<std::string, Node>> getChildren(const std::string & key_)
+    std::vector<std::pair<std::string, Node>> getChildren(const std::string & key_, bool read_data = false)
     {
         rocksdb::ReadOptions read_options;
         read_options.total_order_seek = true;
@@ -232,6 +232,15 @@ public:
             typename Node::Meta & meta = node;
             /// We do not read data here
             readPODBinary(meta, buffer);
+            if (read_data)
+            {
+                readVarUInt(meta.stats.data_size, buffer);
+                if (meta.stats.data_size)
+                {
+                    node.data = std::unique_ptr<char[]>(new char[meta.stats.data_size]);
+                    buffer.readStrict(node.data.get(), meta.stats.data_size);
+                }
+            }
             std::string real_key(iter->key().data() + len, iter->key().size() - len);
             // std::cout << "real key: " << real_key << std::endl;
             result.emplace_back(std::move(real_key), std::move(node));
@@ -268,11 +277,11 @@ public:
         typename Node::Meta & meta = kv->value;
         readPODBinary(meta, buffer);
         /// TODO: Sometimes we don't need to load data.
-        readVarUInt(kv->value.data_size, buffer);
+        readVarUInt(kv->value.stats.data_size, buffer);
-        if (kv->value.data_size)
+        if (kv->value.stats.data_size)
         {
-            kv->value.data = std::unique_ptr<char[]>(new char[kv->value.data_size]);
+            kv->value.data = std::unique_ptr<char[]>(new char[kv->value.stats.data_size]);
-            buffer.readStrict(kv->value.data.get(), kv->value.data_size);
+            buffer.readStrict(kv->value.data.get(), kv->value.stats.data_size);
         }
         return const_iterator(kv);
     }
@@ -281,7 +290,7 @@ public:
     {
         auto it = find(key);
         chassert(it != end());
-        return MockNode(it->value.numChildren(), it->value.getData());
+        return MockNode(it->value.stats.numChildren(), it->value.getData());
     }
 
     const_iterator updateValue(StringRef key_, ValueUpdater updater)

src/Coordination/ZooKeeperDataReader.cpp

@@ -93,7 +93,7 @@ void deserializeACLMap(Storage & storage, ReadBuffer & in)
 }
 
 template<typename Storage>
-int64_t deserializeStorageData(Storage & storage, ReadBuffer & in, LoggerPtr log)
+int64_t deserializeStorageData(Storage & storage, ReadBuffer & in, LoggerPtr log) TSA_NO_THREAD_SAFETY_ANALYSIS
 {
     int64_t max_zxid = 0;
     std::string path;
@@ -108,33 +108,33 @@ int64_t deserializeStorageData(Storage & storage, ReadBuffer & in, LoggerPtr log
         Coordination::read(node.acl_id, in);
 
         /// Deserialize stat
-        Coordination::read(node.czxid, in);
+        Coordination::read(node.stats.czxid, in);
-        Coordination::read(node.mzxid, in);
+        Coordination::read(node.stats.mzxid, in);
         /// For some reason ZXID specified in filename can be smaller
         /// then actual zxid from nodes. In this case we will use zxid from nodes.
-        max_zxid = std::max(max_zxid, node.mzxid);
+        max_zxid = std::max(max_zxid, node.stats.mzxid);
 
         int64_t ctime;
         Coordination::read(ctime, in);
-        node.setCtime(ctime);
+        node.stats.setCtime(ctime);
-        Coordination::read(node.mtime, in);
+        Coordination::read(node.stats.mtime, in);
-        Coordination::read(node.version, in);
+        Coordination::read(node.stats.version, in);
-        Coordination::read(node.cversion, in);
+        Coordination::read(node.stats.cversion, in);
-        Coordination::read(node.aversion, in);
+        Coordination::read(node.stats.aversion, in);
         int64_t ephemeral_owner;
         Coordination::read(ephemeral_owner, in);
         if (ephemeral_owner != 0)
-            node.setEphemeralOwner(ephemeral_owner);
+            node.stats.setEphemeralOwner(ephemeral_owner);
-        Coordination::read(node.pzxid, in);
+        Coordination::read(node.stats.pzxid, in);
         if (!path.empty())
         {
             if (ephemeral_owner == 0)
-                node.setSeqNum(node.cversion);
+                node.stats.setSeqNum(node.stats.cversion);
 
             storage.container.insertOrReplace(path, node);
 
             if (ephemeral_owner != 0)
-                storage.ephemerals[ephemeral_owner].insert(path);
+                storage.committed_ephemerals[ephemeral_owner].insert(path);
 
             storage.acl_map.addUsage(node.acl_id);
         }
@@ -149,7 +149,13 @@ int64_t deserializeStorageData(Storage & storage, ReadBuffer & in, LoggerPtr log
         if (itr.key != "/")
         {
             auto parent_path = parentNodePath(itr.key);
-            storage.container.updateValue(parent_path, [my_path = itr.key] (typename Storage::Node & value) { value.addChild(getBaseNodeName(my_path)); value.increaseNumChildren(); });
+            storage.container.updateValue(
+                parent_path,
+                [my_path = itr.key](typename Storage::Node & value)
+                {
+                    value.addChild(getBaseNodeName(my_path));
+                    value.stats.increaseNumChildren();
+                });
         }
     }
 
@@ -157,7 +163,7 @@ int64_t deserializeStorageData(Storage & storage, ReadBuffer & in, LoggerPtr log
 }
 
 template<typename Storage>
-void deserializeKeeperStorageFromSnapshot(Storage & storage, const std::string & snapshot_path, LoggerPtr log)
+void deserializeKeeperStorageFromSnapshot(Storage & storage, const std::string & snapshot_path, LoggerPtr log) TSA_NO_THREAD_SAFETY_ANALYSIS
 {
     LOG_INFO(log, "Deserializing storage snapshot {}", snapshot_path);
     int64_t zxid = getZxidFromName(snapshot_path);
@@ -487,7 +493,7 @@ bool hasErrorsInMultiRequest(Coordination::ZooKeeperRequestPtr request)
 }
 
 template<typename Storage>
-bool deserializeTxn(Storage & storage, ReadBuffer & in, LoggerPtr /*log*/)
+bool deserializeTxn(Storage & storage, ReadBuffer & in, LoggerPtr /*log*/) TSA_NO_THREAD_SAFETY_ANALYSIS
 {
     int64_t checksum;
     Coordination::read(checksum, in);
@@ -568,7 +574,7 @@ void deserializeLogAndApplyToStorage(Storage & storage, const std::string & log_
 }
 
 template<typename Storage>
-void deserializeLogsAndApplyToStorage(Storage & storage, const std::string & path, LoggerPtr log)
+void deserializeLogsAndApplyToStorage(Storage & storage, const std::string & path, LoggerPtr log) TSA_NO_THREAD_SAFETY_ANALYSIS
 {
     std::map<int64_t, std::string> existing_logs;
     for (const auto & p : fs::directory_iterator(path))

src/Coordination/tests/gtest_coordination.cpp

@@ -1,6 +1,7 @@
 #include <chrono>
 #include <gtest/gtest.h>
 
+#include "base/defines.h"
 #include "config.h"
 
 #if USE_NURAFT
@@ -1540,7 +1541,7 @@ void addNode(Storage & storage, const std::string & path, const std::string & da
     using Node = typename Storage::Node;
     Node node{};
     node.setData(data);
-    node.setEphemeralOwner(ephemeral_owner);
+    node.stats.setEphemeralOwner(ephemeral_owner);
     storage.container.insertOrReplace(path, node);
     auto child_it = storage.container.find(path);
     auto child_path = DB::getBaseNodeName(child_it->key);
@@ -1549,7 +1550,7 @@ void addNode(Storage & storage, const std::string & path, const std::string & da
         [&](auto & parent)
         {
             parent.addChild(child_path);
-            parent.increaseNumChildren();
+            parent.stats.increaseNumChildren();
         });
 }
 
@@ -1570,9 +1571,9 @@ TYPED_TEST(CoordinationTest, TestStorageSnapshotSimple)
     addNode(storage, "/hello1", "world", 1);
     addNode(storage, "/hello2", "somedata", 3);
     storage.session_id_counter = 5;
-    storage.zxid = 2;
+    TSA_SUPPRESS_WARNING_FOR_WRITE(storage.zxid) = 2;
-    storage.ephemerals[3] = {"/hello2"};
+    storage.committed_ephemerals[3] = {"/hello2"};
-    storage.ephemerals[1] = {"/hello1"};
+    storage.committed_ephemerals[1] = {"/hello1"};
     storage.getSessionID(130);
     storage.getSessionID(130);
 
@@ -1601,10 +1602,10 @@ TYPED_TEST(CoordinationTest, TestStorageSnapshotSimple)
     EXPECT_EQ(restored_storage->container.getValue("/hello1").getData(), "world");
     EXPECT_EQ(restored_storage->container.getValue("/hello2").getData(), "somedata");
     EXPECT_EQ(restored_storage->session_id_counter, 7);
-    EXPECT_EQ(restored_storage->zxid, 2);
+    EXPECT_EQ(restored_storage->getZXID(), 2);
-    EXPECT_EQ(restored_storage->ephemerals.size(), 2);
+    EXPECT_EQ(restored_storage->committed_ephemerals.size(), 2);
-    EXPECT_EQ(restored_storage->ephemerals[3].size(), 1);
+    EXPECT_EQ(restored_storage->committed_ephemerals[3].size(), 1);
-    EXPECT_EQ(restored_storage->ephemerals[1].size(), 1);
+    EXPECT_EQ(restored_storage->committed_ephemerals[1].size(), 1);
     EXPECT_EQ(restored_storage->session_and_timeout.size(), 2);
 }
 
@@ -2027,7 +2028,7 @@ TYPED_TEST(CoordinationTest, TestEphemeralNodeRemove)
     state_machine->commit(1, entry_c->get_buf());
     const auto & storage = state_machine->getStorageUnsafe();
 
-    EXPECT_EQ(storage.ephemerals.size(), 1);
+    EXPECT_EQ(storage.committed_ephemerals.size(), 1);
     std::shared_ptr<ZooKeeperRemoveRequest> request_d = std::make_shared<ZooKeeperRemoveRequest>();
     request_d->path = "/hello";
     /// Delete from other session
@@ -2035,7 +2036,7 @@ TYPED_TEST(CoordinationTest, TestEphemeralNodeRemove)
     state_machine->pre_commit(2, entry_d->get_buf());
     state_machine->commit(2, entry_d->get_buf());
 
-    EXPECT_EQ(storage.ephemerals.size(), 0);
+    EXPECT_EQ(storage.committed_ephemerals.size(), 0);
 }
 
 
@@ -2590,9 +2591,9 @@ TYPED_TEST(CoordinationTest, TestStorageSnapshotDifferentCompressions)
     addNode(storage, "/hello1", "world", 1);
     addNode(storage, "/hello2", "somedata", 3);
     storage.session_id_counter = 5;
-    storage.zxid = 2;
+    TSA_SUPPRESS_WARNING_FOR_WRITE(storage.zxid) = 2;
-    storage.ephemerals[3] = {"/hello2"};
+    storage.committed_ephemerals[3] = {"/hello2"};
-    storage.ephemerals[1] = {"/hello1"};
+    storage.committed_ephemerals[1] = {"/hello1"};
     storage.getSessionID(130);
     storage.getSessionID(130);
 
@@ -2617,10 +2618,10 @@ TYPED_TEST(CoordinationTest, TestStorageSnapshotDifferentCompressions)
     EXPECT_EQ(restored_storage->container.getValue("/hello1").getData(), "world");
     EXPECT_EQ(restored_storage->container.getValue("/hello2").getData(), "somedata");
     EXPECT_EQ(restored_storage->session_id_counter, 7);
-    EXPECT_EQ(restored_storage->zxid, 2);
+    EXPECT_EQ(restored_storage->getZXID(), 2);
-    EXPECT_EQ(restored_storage->ephemerals.size(), 2);
+    EXPECT_EQ(restored_storage->committed_ephemerals.size(), 2);
-    EXPECT_EQ(restored_storage->ephemerals[3].size(), 1);
+    EXPECT_EQ(restored_storage->committed_ephemerals[3].size(), 1);
-    EXPECT_EQ(restored_storage->ephemerals[1].size(), 1);
+    EXPECT_EQ(restored_storage->committed_ephemerals[1].size(), 1);
     EXPECT_EQ(restored_storage->session_and_timeout.size(), 2);
 }
 
@@ -2805,13 +2806,13 @@ TYPED_TEST(CoordinationTest, TestStorageSnapshotEqual)
 
         storage.session_id_counter = 5;
 
-        storage.ephemerals[3] = {"/hello"};
+        storage.committed_ephemerals[3] = {"/hello"};
-        storage.ephemerals[1] = {"/hello/somepath"};
+        storage.committed_ephemerals[1] = {"/hello/somepath"};
 
         for (size_t j = 0; j < 3333; ++j)
             storage.getSessionID(130 * j);
 
-        DB::KeeperStorageSnapshot<Storage> snapshot(&storage, storage.zxid);
+        DB::KeeperStorageSnapshot<Storage> snapshot(&storage, storage.getZXID());
 
         auto buf = manager.serializeSnapshotToBuffer(snapshot);
 
@@ -3315,7 +3316,7 @@ TYPED_TEST(CoordinationTest, TestCheckNotExistsRequest)
     create_path("/test_node");
     auto node_it = storage.container.find("/test_node");
     ASSERT_NE(node_it, storage.container.end());
-    auto node_version = node_it->value.version;
+    auto node_version = node_it->value.stats.version;
 
     {
         SCOPED_TRACE("CheckNotExists returns ZNODEEXISTS");
@@ -3566,12 +3567,12 @@ TYPED_TEST(CoordinationTest, TestRemoveRecursiveRequest)
     {
         SCOPED_TRACE("Recursive Remove Ephemeral");
         create("/T7", zkutil::CreateMode::Ephemeral);
-        ASSERT_EQ(storage.ephemerals.size(), 1);
+        ASSERT_EQ(storage.committed_ephemerals.size(), 1);
 
         auto responses = remove_recursive("/T7", 100);
         ASSERT_EQ(responses.size(), 1);
         ASSERT_EQ(responses[0].response->error, Coordination::Error::ZOK);
-        ASSERT_EQ(storage.ephemerals.size(), 0);
+        ASSERT_EQ(storage.committed_ephemerals.size(), 0);
         ASSERT_FALSE(exists("/T7"));
     }
 
@@ -3581,12 +3582,12 @@ TYPED_TEST(CoordinationTest, TestRemoveRecursiveRequest)
         create("/T8/A", zkutil::CreateMode::Persistent);
         create("/T8/B", zkutil::CreateMode::Ephemeral);
         create("/T8/A/C", zkutil::CreateMode::Ephemeral);
-        ASSERT_EQ(storage.ephemerals.size(), 1);
+        ASSERT_EQ(storage.committed_ephemerals.size(), 1);
 
         auto responses = remove_recursive("/T8", 4);
         ASSERT_EQ(responses.size(), 1);
         ASSERT_EQ(responses[0].response->error, Coordination::Error::ZOK);
-        ASSERT_EQ(storage.ephemerals.size(), 0);
+        ASSERT_EQ(storage.committed_ephemerals.size(), 0);
         ASSERT_FALSE(exists("/T8"));
         ASSERT_FALSE(exists("/T8/A"));
         ASSERT_FALSE(exists("/T8/B"));
@@ -3738,6 +3739,72 @@ TYPED_TEST(CoordinationTest, TestRemoveRecursiveInMultiRequest)
         ASSERT_FALSE(exists("/A/B"));
         ASSERT_FALSE(exists("/A/B/D"));
     }
+
+    {
+        SCOPED_TRACE("Recursive Remove For Subtree With Updated Node");
+        int create_zxid = ++zxid;
+        auto ops = prepare_create_tree();
+
+        /// First create nodes
+        const auto create_request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+        storage.preprocessRequest(create_request, 1, 0, create_zxid);
+        auto create_responses = storage.processRequest(create_request, 1, create_zxid);
+        ASSERT_EQ(create_responses.size(), 1);
+        ASSERT_TRUE(is_multi_ok(create_responses[0].response));
+
+        /// Small limit
+        int remove_zxid = ++zxid;
+        ops = {
+            zkutil::makeSetRequest("/A/B", "", -1),
+            zkutil::makeRemoveRecursiveRequest("/A", 3),
+        };
+        auto remove_request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+        storage.preprocessRequest(remove_request, 1, 0, remove_zxid);
+        auto remove_responses = storage.processRequest(remove_request, 1, remove_zxid);
+
+        ASSERT_EQ(remove_responses.size(), 1);
+        ASSERT_FALSE(is_multi_ok(remove_responses[0].response));
+
+        /// Big limit
+        remove_zxid = ++zxid;
+        ops[1] = zkutil::makeRemoveRecursiveRequest("/A", 4);
+        remove_request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+        storage.preprocessRequest(remove_request, 1, 0, remove_zxid);
+        remove_responses = storage.processRequest(remove_request, 1, remove_zxid);
+
+        ASSERT_EQ(remove_responses.size(), 1);
+        ASSERT_TRUE(is_multi_ok(remove_responses[0].response));
+        ASSERT_FALSE(exists("/A"));
+        ASSERT_FALSE(exists("/A/C"));
+        ASSERT_FALSE(exists("/A/B"));
+        ASSERT_FALSE(exists("/A/B/D"));
+    }
+
+    {
+        SCOPED_TRACE("[BUG] Recursive Remove Level Sorting");
+        int new_zxid = ++zxid;
+
+        Coordination::Requests ops = {
+            zkutil::makeCreateRequest("/a", "", zkutil::CreateMode::Persistent),
+            zkutil::makeCreateRequest("/a/bbbbbb", "", zkutil::CreateMode::Persistent),
+            zkutil::makeCreateRequest("/A", "", zkutil::CreateMode::Persistent),
+            zkutil::makeCreateRequest("/A/B", "", zkutil::CreateMode::Persistent),
+            zkutil::makeCreateRequest("/A/CCCCCCCCCCCC", "", zkutil::CreateMode::Persistent),
+            zkutil::makeRemoveRecursiveRequest("/A", 3),
+        };
+        auto remove_request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+        storage.preprocessRequest(remove_request, 1, 0, new_zxid);
+        auto remove_responses = storage.processRequest(remove_request, 1, new_zxid);
+
+        ASSERT_EQ(remove_responses.size(), 1);
+        ASSERT_TRUE(is_multi_ok(remove_responses[0].response));
+        ASSERT_TRUE(exists("/a"));
+        ASSERT_TRUE(exists("/a/bbbbbb"));
+        ASSERT_FALSE(exists("/A"));
+        ASSERT_FALSE(exists("/A/B"));
+        ASSERT_FALSE(exists("/A/CCCCCCCCCCCC"));
+    }
+
 }
 
 TYPED_TEST(CoordinationTest, TestRemoveRecursiveWatches)
@@ -3823,14 +3890,26 @@ TYPED_TEST(CoordinationTest, TestRemoveRecursiveWatches)
     auto responses = storage.processRequest(remove_request, 1, new_zxid);
 
     ASSERT_EQ(responses.size(), 7);
+    /// request response is last
+    ASSERT_EQ(dynamic_cast<Coordination::ZooKeeperWatchResponse *>(responses.back().response.get()), nullptr);
 
-    for (size_t i = 0; i < 7; ++i)
+    std::unordered_map<std::string, std::vector<Coordination::Event>> expected_watch_responses
+    {
+        {"/A/B/D", {Coordination::Event::DELETED}},
+        {"/A/B", {Coordination::Event::CHILD, Coordination::Event::DELETED}},
+        {"/A/C", {Coordination::Event::DELETED}},
+        {"/A", {Coordination::Event::CHILD, Coordination::Event::DELETED}},
+    };
+
+    std::unordered_map<std::string, std::vector<Coordination::Event>> actual_watch_responses;
+    for (size_t i = 0; i < 6; ++i)
     {
         ASSERT_EQ(responses[i].response->error, Coordination::Error::ZOK);
 
-        if (const auto * watch_response = dynamic_cast<Coordination::ZooKeeperWatchResponse *>(responses[i].response.get()))
+        const auto & watch_response = dynamic_cast<Coordination::ZooKeeperWatchResponse &>(*responses[i].response);
-            ASSERT_EQ(watch_response->type, Coordination::Event::DELETED);
+        actual_watch_responses[watch_response.path].push_back(static_cast<Coordination::Event>(watch_response.type));
     }
+    ASSERT_EQ(expected_watch_responses, actual_watch_responses);
 
     ASSERT_EQ(storage.watches.size(), 0);
     ASSERT_EQ(storage.list_watches.size(), 0);

src/Core/ProtocolDefines.h

@@ -33,7 +33,8 @@ static constexpr auto DBMS_MIN_REVISION_WITH_AGGREGATE_FUNCTIONS_VERSIONING = 54
 
 static constexpr auto DBMS_CLUSTER_PROCESSING_PROTOCOL_VERSION = 1;
 
-static constexpr auto DBMS_PARALLEL_REPLICAS_PROTOCOL_VERSION = 3;
+static constexpr auto DBMS_MIN_SUPPORTED_PARALLEL_REPLICAS_PROTOCOL_VERSION = 3;
+static constexpr auto DBMS_PARALLEL_REPLICAS_PROTOCOL_VERSION = 4;
 static constexpr auto DBMS_MIN_REVISION_WITH_PARALLEL_REPLICAS = 54453;
 
 static constexpr auto DBMS_MERGE_TREE_PART_INFO_VERSION = 1;
@@ -86,6 +87,8 @@ static constexpr auto DBMS_MIN_REVISION_WITH_ROWS_BEFORE_AGGREGATION = 54469;
 /// Packets size header
 static constexpr auto DBMS_MIN_PROTOCOL_VERSION_WITH_CHUNKED_PACKETS = 54470;
 
+static constexpr auto DBMS_MIN_REVISION_WITH_VERSIONED_PARALLEL_REPLICAS_PROTOCOL = 54471;
+
 /// Version of ClickHouse TCP protocol.
 ///
 /// Should be incremented manually on protocol changes.
@@ -93,6 +96,6 @@ static constexpr auto DBMS_MIN_PROTOCOL_VERSION_WITH_CHUNKED_PACKETS = 54470;
 /// NOTE: DBMS_TCP_PROTOCOL_VERSION has nothing common with VERSION_REVISION,
 /// later is just a number for server version (one number instead of commit SHA)
 /// for simplicity (sometimes it may be more convenient in some use cases).
-static constexpr auto DBMS_TCP_PROTOCOL_VERSION = 54470;
+static constexpr auto DBMS_TCP_PROTOCOL_VERSION = 54471;
 
 }

src/Core/Settings.h

@@ -947,7 +947,7 @@ class IColumn;
     M(Bool, parallel_replicas_for_non_replicated_merge_tree, false, "If true, ClickHouse will use parallel replicas algorithm also for non-replicated MergeTree tables", 0) \
     M(UInt64, parallel_replicas_min_number_of_rows_per_replica, 0, "Limit the number of replicas used in a query to (estimated rows to read / min_number_of_rows_per_replica). The max is still limited by 'max_parallel_replicas'", 0) \
     M(Bool, parallel_replicas_prefer_local_join, true, "If true, and JOIN can be executed with parallel replicas algorithm, and all storages of right JOIN part are *MergeTree, local JOIN will be used instead of GLOBAL JOIN.", 0) \
-    M(UInt64, parallel_replicas_mark_segment_size, 128, "Parts virtually divided into segments to be distributed between replicas for parallel reading. This setting controls the size of these segments. Not recommended to change until you're absolutely sure in what you're doing", 0) \
+    M(UInt64, parallel_replicas_mark_segment_size, 0, "Parts virtually divided into segments to be distributed between replicas for parallel reading. This setting controls the size of these segments. Not recommended to change until you're absolutely sure in what you're doing. Value should be in range [128; 16384]", 0) \
     M(Bool, allow_archive_path_syntax, true, "File/S3 engines/table function will parse paths with '::' as '<archive> :: <file>' if archive has correct extension", 0) \
     M(Bool, parallel_replicas_local_plan, false, "Build local plan for local replica", 0) \
     \
@@ -1272,6 +1272,7 @@ class IColumn;
     M(Bool, output_format_orc_string_as_string, true, "Use ORC String type instead of Binary for String columns", 0) \
     M(ORCCompression, output_format_orc_compression_method, "zstd", "Compression method for ORC output format. Supported codecs: lz4, snappy, zlib, zstd, none (uncompressed)", 0) \
     M(UInt64, output_format_orc_row_index_stride, 10'000, "Target row index stride in ORC output format", 0) \
+    M(Double, output_format_orc_dictionary_key_size_threshold, 0.0, "For a string column in ORC output format, if the number of distinct values is greater than this fraction of the total number of non-null rows, turn off dictionary encoding. Otherwise dictionary encoding is enabled", 0) \
     \
     M(CapnProtoEnumComparingMode, format_capn_proto_enum_comparising_mode, FormatSettings::CapnProtoEnumComparingMode::BY_VALUES, "How to map ClickHouse Enum and CapnProto Enum", 0) \
     \

src/Core/SettingsChangesHistory.cpp

@@ -71,6 +71,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
     },
     {"24.9",
         {
+            {"output_format_orc_dictionary_key_size_threshold", 0.0, 0.0, "For a string column in ORC output format, if the number of distinct values is greater than this fraction of the total number of non-null rows, turn off dictionary encoding. Otherwise dictionary encoding is enabled"},
             {"input_format_json_empty_as_default", false, false, "Added new setting to allow to treat empty fields in JSON input as default values."},
             {"input_format_try_infer_variants", false, false, "Try to infer Variant type in text formats when there is more than one possible type for column/array elements"},
             {"join_output_by_rowlist_perkey_rows_threshold", 0, 5, "The lower limit of per-key average rows in the right table to determine whether to output by row list in hash join."},
@@ -78,6 +79,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
             {"allow_materialized_view_with_bad_select", true, true, "Support (but not enable yet) stricter validation in CREATE MATERIALIZED VIEW"},
             {"output_format_always_quote_identifiers", false, false, "New setting."},
             {"output_format_identifier_quoting_style", "Backticks", "Backticks", "New setting."},
+            {"parallel_replicas_mark_segment_size", 128, 0, "Value for this setting now determined automatically"},
             {"database_replicated_allow_replicated_engine_arguments", 1, 0, "Don't allow explicit arguments by default"},
             {"database_replicated_allow_explicit_uuid", 0, 0, "Added a new setting to disallow explicitly specifying table UUID"},
             {"parallel_replicas_local_plan", false, false, "Use local plan for local replica in a query with parallel replicas"},

src/Databases/DatabaseFactory.cpp

@@ -59,35 +59,27 @@ void cckMetadataPathForOrdinary(const ASTCreateQuery & create, const String & me
 
 }
 
-/// validate validates the database engine that's specified in the create query for
+void DatabaseFactory::validate(const ASTCreateQuery & create_query) const
-/// engine arguments, settings and table overrides.
-void validate(const ASTCreateQuery & create_query)
-
 {
     auto * storage = create_query.storage;
 
-    /// Check engine may have arguments
-    static const std::unordered_set<std::string_view> engines_with_arguments{"MySQL", "MaterializeMySQL", "MaterializedMySQL",
-        "Lazy", "Replicated", "PostgreSQL", "MaterializedPostgreSQL", "SQLite", "Filesystem", "S3", "HDFS"};
-
     const String & engine_name = storage->engine->name;
-    bool engine_may_have_arguments = engines_with_arguments.contains(engine_name);
+    const EngineFeatures & engine_features = database_engines.at(engine_name).features;
 
-    if (storage->engine->arguments && !engine_may_have_arguments)
+    /// Check engine may have arguments
+    if (storage->engine->arguments && !engine_features.supports_arguments)
         throw Exception(ErrorCodes::BAD_ARGUMENTS, "Database engine `{}` cannot have arguments", engine_name);
 
     /// Check engine may have settings
-    bool may_have_settings = endsWith(engine_name, "MySQL") || engine_name == "Replicated" || engine_name == "MaterializedPostgreSQL";
     bool has_unexpected_element = storage->engine->parameters || storage->partition_by ||
         storage->primary_key || storage->order_by ||
         storage->sample_by;
-    if (has_unexpected_element || (!may_have_settings && storage->settings))
+    if (has_unexpected_element || (!engine_features.supports_settings && storage->settings))
         throw Exception(ErrorCodes::UNKNOWN_ELEMENT_IN_AST,
                         "Database engine `{}` cannot have parameters, primary_key, order_by, sample_by, settings", engine_name);
 
     /// Check engine with table overrides
-    static const std::unordered_set<std::string_view> engines_with_table_overrides{"MaterializeMySQL", "MaterializedMySQL", "MaterializedPostgreSQL"};
+    if (create_query.table_overrides && !engine_features.supports_table_overrides)
-    if (create_query.table_overrides && !engines_with_table_overrides.contains(engine_name))
         throw Exception(ErrorCodes::BAD_ARGUMENTS, "Database engine `{}` cannot have table overrides", engine_name);
 }
 
@@ -121,9 +113,9 @@ DatabasePtr DatabaseFactory::get(const ASTCreateQuery & create, const String & m
     return impl;
 }
 
-void DatabaseFactory::registerDatabase(const std::string & name, CreatorFn creator_fn)
+void DatabaseFactory::registerDatabase(const std::string & name, CreatorFn creator_fn, EngineFeatures features)
 {
-    if (!database_engines.emplace(name, std::move(creator_fn)).second)
+    if (!database_engines.emplace(name, Creator{std::move(creator_fn), features}).second)
         throw Exception(ErrorCodes::LOGICAL_ERROR, "DatabaseFactory: the database engine name '{}' is not unique", name);
 }
 
@@ -154,7 +146,7 @@ DatabasePtr DatabaseFactory::getImpl(const ASTCreateQuery & create, const String
         .context = context};
 
     // creator_fn creates and returns a DatabasePtr with the supplied arguments
-    auto creator_fn = database_engines.at(engine_name);
+    auto creator_fn = database_engines.at(engine_name).creator_fn;
 
     return creator_fn(arguments);
 }

src/Databases/DatabaseFactory.h

@@ -43,13 +43,30 @@ public:
         ContextPtr & context;
     };
 
-    DatabasePtr get(const ASTCreateQuery & create, const String & metadata_path, ContextPtr context);
+    struct EngineFeatures
+    {
+        bool supports_arguments = false;
+        bool supports_settings = false;
+        bool supports_table_overrides = false;
+    };
 
     using CreatorFn = std::function<DatabasePtr(const Arguments & arguments)>;
 
-    using DatabaseEngines = std::unordered_map<std::string, CreatorFn>;
+    struct Creator
+    {
+        CreatorFn creator_fn;
+        EngineFeatures features;
+    };
 
-    void registerDatabase(const std::string & name, CreatorFn creator_fn);
+    DatabasePtr get(const ASTCreateQuery & create, const String & metadata_path, ContextPtr context);
+
+    using DatabaseEngines = std::unordered_map<std::string, Creator>;
+
+    void registerDatabase(const std::string & name, CreatorFn creator_fn, EngineFeatures features = EngineFeatures{
+        .supports_arguments = false,
+        .supports_settings = false,
+        .supports_table_overrides = false,
+    });
 
     const DatabaseEngines & getDatabaseEngines() const { return database_engines; }
 
@@ -65,6 +82,10 @@ private:
     DatabaseEngines database_engines;
 
     DatabasePtr getImpl(const ASTCreateQuery & create, const String & metadata_path, ContextPtr context);
+
+    /// validate validates the database engine that's specified in the create query for
+    /// engine arguments, settings and table overrides.
+    void validate(const ASTCreateQuery & create_query) const;
 };
 
 }

src/Databases/DatabaseFilesystem.cpp

@@ -257,6 +257,6 @@ void registerDatabaseFilesystem(DatabaseFactory & factory)
 
         return std::make_shared<DatabaseFilesystem>(args.database_name, init_path, args.context);
     };
-    factory.registerDatabase("Filesystem", create_fn);
+    factory.registerDatabase("Filesystem", create_fn, {.supports_arguments = true});
 }
 }

src/Databases/DatabaseHDFS.cpp

@@ -253,7 +253,7 @@ void registerDatabaseHDFS(DatabaseFactory & factory)
 
         return std::make_shared<DatabaseHDFS>(args.database_name, source_url, args.context);
     };
-    factory.registerDatabase("HDFS", create_fn);
+    factory.registerDatabase("HDFS", create_fn, {.supports_arguments = true});
 }
 } // DB
 

src/Databases/DatabaseLazy.cpp

@@ -398,6 +398,6 @@ void registerDatabaseLazy(DatabaseFactory & factory)
             cache_expiration_time_seconds,
             args.context);
     };
-    factory.registerDatabase("Lazy", create_fn);
+    factory.registerDatabase("Lazy", create_fn, {.supports_arguments = true});
 }
 }

src/Databases/DatabaseReplicated.cpp

@@ -2001,6 +2001,6 @@ void registerDatabaseReplicated(DatabaseFactory & factory)
             replica_name,
             std::move(database_replicated_settings), args.context);
     };
-    factory.registerDatabase("Replicated", create_fn);
+    factory.registerDatabase("Replicated", create_fn, {.supports_arguments = true, .supports_settings = true});
 }
 }

src/Databases/DatabaseS3.cpp

@@ -326,7 +326,7 @@ void registerDatabaseS3(DatabaseFactory & factory)
 
         return std::make_shared<DatabaseS3>(args.database_name, config, args.context);
     };
-    factory.registerDatabase("S3", create_fn);
+    factory.registerDatabase("S3", create_fn, {.supports_arguments = true});
 }
 }
 #endif

src/Databases/MySQL/DatabaseMaterializedMySQL.cpp

@@ -290,8 +290,14 @@ void registerDatabaseMaterializedMySQL(DatabaseFactory & factory)
             binlog_client,
             std::move(materialize_mode_settings));
     };
-    factory.registerDatabase("MaterializeMySQL", create_fn);
+
-    factory.registerDatabase("MaterializedMySQL", create_fn);
+    DatabaseFactory::EngineFeatures features{
+        .supports_arguments = true,
+        .supports_settings = true,
+        .supports_table_overrides = true,
+    };
+    factory.registerDatabase("MaterializeMySQL", create_fn, features);
+    factory.registerDatabase("MaterializedMySQL", create_fn, features);
 }
 
 }

src/Databases/MySQL/DatabaseMySQL.cpp

@@ -584,7 +584,7 @@ void registerDatabaseMySQL(DatabaseFactory & factory)
             throw Exception(ErrorCodes::CANNOT_CREATE_DATABASE, "Cannot create MySQL database, because {}", exception_message);
         }
     };
-    factory.registerDatabase("MySQL", create_fn);
+    factory.registerDatabase("MySQL", create_fn, {.supports_arguments = true, .supports_settings = true});
 }
 }
 

src/Databases/PostgreSQL/DatabaseMaterializedPostgreSQL.cpp

@@ -546,7 +546,11 @@ void registerDatabaseMaterializedPostgreSQL(DatabaseFactory & factory)
             args.database_name, configuration.database, connection_info,
             std::move(postgresql_replica_settings));
     };
-    factory.registerDatabase("MaterializedPostgreSQL", create_fn);
+    factory.registerDatabase("MaterializedPostgreSQL", create_fn, {
+        .supports_arguments = true,
+        .supports_settings = true,
+        .supports_table_overrides = true,
+    });
 }
 }
 

src/Databases/PostgreSQL/DatabasePostgreSQL.cpp

@@ -558,7 +558,7 @@ void registerDatabasePostgreSQL(DatabaseFactory & factory)
             pool,
             use_table_cache);
     };
-    factory.registerDatabase("PostgreSQL", create_fn);
+    factory.registerDatabase("PostgreSQL", create_fn, {.supports_arguments = true});
 }
 }
 

src/Databases/SQLite/DatabaseSQLite.cpp

@@ -220,7 +220,7 @@ void registerDatabaseSQLite(DatabaseFactory & factory)
 
         return std::make_shared<DatabaseSQLite>(args.context, engine_define, args.create_query.attach, database_path);
     };
-    factory.registerDatabase("SQLite", create_fn);
+    factory.registerDatabase("SQLite", create_fn, {.supports_arguments = true});
 }
 }
 

src/Disks/IO/CachedOnDiskReadBufferFromFile.cpp

@@ -28,6 +28,7 @@ extern const Event CachedReadBufferReadFromCacheMicroseconds;
 extern const Event CachedReadBufferCacheWriteMicroseconds;
 extern const Event CachedReadBufferReadFromSourceBytes;
 extern const Event CachedReadBufferReadFromCacheBytes;
+extern const Event CachedReadBufferPredownloadedBytes;
 extern const Event CachedReadBufferCacheWriteBytes;
 extern const Event CachedReadBufferCreateBufferMicroseconds;
 
@@ -644,6 +645,7 @@ void CachedOnDiskReadBufferFromFile::predownload(FileSegment & file_segment)
             size_t current_predownload_size = std::min(current_impl_buffer_size, bytes_to_predownload);
 
             ProfileEvents::increment(ProfileEvents::CachedReadBufferReadFromSourceBytes, current_impl_buffer_size);
+            ProfileEvents::increment(ProfileEvents::CachedReadBufferPredownloadedBytes, current_impl_buffer_size);
 
             std::string failure_reason;
             bool continue_predownload = file_segment.reserve(

src/Formats/FormatFactory.cpp

@@ -244,6 +244,7 @@ FormatSettings getFormatSettings(const ContextPtr & context, const Settings & se
     format_settings.orc.output_string_as_string = settings.output_format_orc_string_as_string;
     format_settings.orc.output_compression_method = settings.output_format_orc_compression_method;
     format_settings.orc.output_row_index_stride = settings.output_format_orc_row_index_stride;
+    format_settings.orc.output_dictionary_key_size_threshold = settings.output_format_orc_dictionary_key_size_threshold;
     format_settings.orc.use_fast_decoder = settings.input_format_orc_use_fast_decoder;
     format_settings.orc.filter_push_down = settings.input_format_orc_filter_push_down;
     format_settings.orc.reader_time_zone_name = settings.input_format_orc_reader_time_zone_name;

src/Formats/FormatSettings.h

@@ -415,6 +415,7 @@ struct FormatSettings
         bool filter_push_down = true;
         UInt64 output_row_index_stride = 10'000;
         String reader_time_zone_name = "GMT";
+        double output_dictionary_key_size_threshold = 0.0;
     } orc{};
 
     /// For capnProto format we should determine how to

src/Functions/array/arrayZip.cpp

@@ -15,7 +15,6 @@ namespace ErrorCodes
 {
 extern const int ILLEGAL_TYPE_OF_ARGUMENT;
 extern const int SIZES_OF_ARRAYS_DONT_MATCH;
-extern const int TOO_FEW_ARGUMENTS_FOR_FUNCTION;
 extern const int ILLEGAL_COLUMN;
 }
 
@@ -38,13 +37,6 @@ public:
 
     DataTypePtr getReturnTypeImpl(const ColumnsWithTypeAndName & arguments) const override
     {
-        if (arguments.empty())
-            throw Exception(
-                ErrorCodes::TOO_FEW_ARGUMENTS_FOR_FUNCTION,
-                "Function {} needs at least one argument; passed {}.",
-                getName(),
-                arguments.size());
-
         DataTypes arguments_types;
         for (size_t index = 0; index < arguments.size(); ++index)
         {
@@ -68,9 +60,16 @@ public:
     }
 
     ColumnPtr
-    executeImpl(const ColumnsWithTypeAndName & arguments, const DataTypePtr & /*result_type*/, size_t input_rows_count) const override
+    executeImpl(const ColumnsWithTypeAndName & arguments, const DataTypePtr & result_type, size_t input_rows_count) const override
     {
         size_t num_arguments = arguments.size();
+        if (num_arguments == 0)
+        {
+            auto res_col = result_type->createColumn();
+            res_col->insertDefault();
+            return ColumnConst::create(std::move(res_col), input_rows_count);
+        }
+
         Columns holders(num_arguments);
         Columns tuple_columns(num_arguments);
 

src/Interpreters/Cache/FileCache.cpp

@@ -718,7 +718,12 @@ FileCache::getOrSet(
         }
     }
 
-    chassert(file_segments_limit ? file_segments.back()->range().left <= result_range.right : file_segments.back()->range().contains(result_range.right));
+    chassert(file_segments_limit
+             ? file_segments.back()->range().left <= result_range.right
+             : file_segments.back()->range().contains(result_range.right),
+             fmt::format("Unexpected state. Back: {}, result range: {}, limit: {}",
+                         file_segments.back()->range().toString(), result_range.toString(), file_segments_limit));
+
     chassert(!file_segments_limit || file_segments.size() <= file_segments_limit);
 
     return std::make_unique<FileSegmentsHolder>(std::move(file_segments));

src/Interpreters/ClusterProxy/executeQuery.cpp

@@ -532,7 +532,7 @@ void executeQueryWithParallelReplicas(
         max_replicas_to_use = shard.getAllNodeCount();
     }
 
-    auto coordinator = std::make_shared<ParallelReplicasReadingCoordinator>(max_replicas_to_use, settings.parallel_replicas_mark_segment_size);
+    auto coordinator = std::make_shared<ParallelReplicasReadingCoordinator>(max_replicas_to_use);
 
     auto external_tables = new_context->getExternalTables();
 

src/Interpreters/HashJoin/HashJoin.cpp

@@ -338,11 +338,8 @@ size_t HashJoin::getTotalRowCount() const
     return res;
 }
 
-size_t HashJoin::getTotalByteCount() const
+void HashJoin::doDebugAsserts() const
 {
-    if (!data)
-        return 0;
-
 #ifndef NDEBUG
     size_t debug_blocks_allocated_size = 0;
     for (const auto & block : data->blocks)
@@ -360,6 +357,14 @@ size_t HashJoin::getTotalByteCount() const
         throw Exception(ErrorCodes::LOGICAL_ERROR, "data->blocks_nullmaps_allocated_size != debug_blocks_nullmaps_allocated_size ({} != {})",
                         data->blocks_nullmaps_allocated_size, debug_blocks_nullmaps_allocated_size);
 #endif
+}
+
+size_t HashJoin::getTotalByteCount() const
+{
+    if (!data)
+        return 0;
+
+    doDebugAsserts();
 
     size_t res = 0;
 
@@ -544,9 +549,11 @@ bool HashJoin::addBlockToJoin(const Block & source_block_, bool check_limits)
             have_compressed = true;
         }
 
+        doDebugAsserts();
         data->blocks_allocated_size += block_to_save.allocatedBytes();
         data->blocks.emplace_back(std::move(block_to_save));
         Block * stored_block = &data->blocks.back();
+        doDebugAsserts();
 
         if (rows)
             data->empty = false;
@@ -634,9 +641,11 @@ bool HashJoin::addBlockToJoin(const Block & source_block_, bool check_limits)
 
             if (!flag_per_row && !is_inserted)
             {
+                doDebugAsserts();
                 LOG_TRACE(log, "Skipping inserting block with {} rows", rows);
                 data->blocks_allocated_size -= stored_block->allocatedBytes();
                 data->blocks.pop_back();
+                doDebugAsserts();
             }
 
             if (!check_limits)
@@ -683,6 +692,8 @@ void HashJoin::shrinkStoredBlocksToFit(size_t & total_bytes_in_join, bool force_
 
     for (auto & stored_block : data->blocks)
     {
+        doDebugAsserts();
+
         size_t old_size = stored_block.allocatedBytes();
         stored_block = stored_block.shrinkToFit();
         size_t new_size = stored_block.allocatedBytes();
@@ -700,6 +711,8 @@ void HashJoin::shrinkStoredBlocksToFit(size_t & total_bytes_in_join, bool force_
         else
             /// Sometimes after clone resized block can be bigger than original
             data->blocks_allocated_size += new_size - old_size;
+
+        doDebugAsserts();
     }
 
     auto new_total_bytes_in_join = getTotalByteCount();
@@ -1416,7 +1429,13 @@ void HashJoin::tryRerangeRightTableDataImpl(Map & map [[maybe_unused]])
         };
         BlocksList sorted_blocks;
         visit_rows_map(sorted_blocks, map);
+        doDebugAsserts();
         data->blocks.swap(sorted_blocks);
+        size_t new_blocks_allocated_size = 0;
+        for (const auto & block : data->blocks)
+            new_blocks_allocated_size += block.allocatedBytes();
+        data->blocks_allocated_size = new_blocks_allocated_size;
+        doDebugAsserts();
     }
 }
 

src/Interpreters/HashJoin/HashJoin.h

@@ -470,6 +470,7 @@ private:
     void tryRerangeRightTableData() override;
     template <JoinKind KIND, typename Map, JoinStrictness STRICTNESS>
     void tryRerangeRightTableDataImpl(Map & map);
+    void doDebugAsserts() const;
 };
 
 }

src/Interpreters/InterpreterDropQuery.cpp

@@ -380,100 +380,99 @@ BlockIO InterpreterDropQuery::executeToDatabase(const ASTDropQuery & query)
 
 BlockIO InterpreterDropQuery::executeToDatabaseImpl(const ASTDropQuery & query, DatabasePtr & database, std::vector<UUID> & uuids_to_wait)
 {
+    if (query.kind != ASTDropQuery::Kind::Detach && query.kind != ASTDropQuery::Kind::Drop && query.kind != ASTDropQuery::Kind::Truncate)
+        return {};
+
     const auto & database_name = query.getDatabase();
     auto ddl_guard = DatabaseCatalog::instance().getDDLGuard(database_name, "");
 
     database = tryGetDatabase(database_name, query.if_exists);
-    if (database)
+    if (!database)
+        return {};
+
+    bool drop = query.kind == ASTDropQuery::Kind::Drop;
+    bool truncate = query.kind == ASTDropQuery::Kind::Truncate;
+
+    getContext()->checkAccess(AccessType::DROP_DATABASE, database_name);
+
+    if (query.kind == ASTDropQuery::Kind::Detach && query.permanently)
+        throw Exception(ErrorCodes::NOT_IMPLEMENTED, "DETACH PERMANENTLY is not implemented for databases");
+
+    if (query.if_empty)
+        throw Exception(ErrorCodes::NOT_IMPLEMENTED, "DROP IF EMPTY is not implemented for databases");
+
+    if (!truncate && database->hasReplicationThread())
+        database->stopReplication();
+
+    if (database->shouldBeEmptyOnDetach())
     {
-        if (query.kind == ASTDropQuery::Kind::Detach || query.kind == ASTDropQuery::Kind::Drop
+        /// Cancel restarting replicas in that database, wait for remaining RESTART queries to finish.
-            || query.kind == ASTDropQuery::Kind::Truncate)
+        /// So it will not startup tables concurrently with the flushAndPrepareForShutdown call below.
+        auto restart_replica_lock = DatabaseCatalog::instance().getLockForDropDatabase(database_name);
+
+        ASTDropQuery query_for_table;
+        query_for_table.kind = query.kind;
+        // For truncate operation on database, drop the tables
+        if (truncate)
+            query_for_table.kind = query.has_all_tables ? ASTDropQuery::Kind::Truncate : ASTDropQuery::Kind::Drop;
+        query_for_table.if_exists = true;
+        query_for_table.if_empty = false;
+        query_for_table.setDatabase(database_name);
+        query_for_table.sync = query.sync;
+
+        /// Flush should not be done if shouldBeEmptyOnDetach() == false,
+        /// since in this case getTablesIterator() may do some additional work,
+        /// see DatabaseMaterializedMySQL::getTablesIterator()
+        auto table_context = Context::createCopy(getContext());
+        table_context->setInternalQuery(true);
+        /// Do not hold extra shared pointers to tables
+        std::vector<std::pair<StorageID, bool>> tables_to_drop;
+        // NOTE: This means we wait for all tables to be loaded inside getTablesIterator() call in case of `async_load_databases = true`.
+        for (auto iterator = database->getTablesIterator(table_context); iterator->isValid(); iterator->next())
         {
-            bool drop = query.kind == ASTDropQuery::Kind::Drop;
+            auto table_ptr = iterator->table();
-            bool truncate = query.kind == ASTDropQuery::Kind::Truncate;
+            tables_to_drop.push_back({table_ptr->getStorageID(), table_ptr->isDictionary()});
+        }
 
-            getContext()->checkAccess(AccessType::DROP_DATABASE, database_name);
+        /// Prepare tables for shutdown in parallel.
-
+        ThreadPoolCallbackRunnerLocal<void> runner(getDatabaseCatalogDropTablesThreadPool().get(), "DropTables");
-            if (query.kind == ASTDropQuery::Kind::Detach && query.permanently)
+        for (const auto & [name, _] : tables_to_drop)
-                throw Exception(ErrorCodes::NOT_IMPLEMENTED, "DETACH PERMANENTLY is not implemented for databases");
+        {
-
+            auto table_ptr = DatabaseCatalog::instance().getTable(name, table_context);
-            if (query.if_empty)
+            runner([my_table_ptr = std::move(table_ptr)]()
-                throw Exception(ErrorCodes::NOT_IMPLEMENTED, "DROP IF EMPTY is not implemented for databases");
-
-            if (!truncate && database->hasReplicationThread())
-                database->stopReplication();
-
-            if (database->shouldBeEmptyOnDetach())
             {
-                /// Cancel restarting replicas in that database, wait for remaining RESTART queries to finish.
+                my_table_ptr->flushAndPrepareForShutdown();
-                /// So it will not startup tables concurrently with the flushAndPrepareForShutdown call below.
+            });
-                auto restart_replica_lock = DatabaseCatalog::instance().getLockForDropDatabase(database_name);
+        }
+        runner.waitForAllToFinishAndRethrowFirstError();
 
-                ASTDropQuery query_for_table;
+        for (const auto & table : tables_to_drop)
-                query_for_table.kind = query.kind;
+        {
-                // For truncate operation on database, drop the tables
+            query_for_table.setTable(table.first.getTableName());
-                if (truncate)
+            query_for_table.is_dictionary = table.second;
-                    query_for_table.kind = query.has_all_tables ? ASTDropQuery::Kind::Truncate : ASTDropQuery::Kind::Drop;
+            DatabasePtr db;
-                query_for_table.if_exists = true;
+            UUID table_to_wait = UUIDHelpers::Nil;
-                query_for_table.if_empty = false;
+            executeToTableImpl(table_context, query_for_table, db, table_to_wait);
-                query_for_table.setDatabase(database_name);
+            uuids_to_wait.push_back(table_to_wait);
-                query_for_table.sync = query.sync;
-
-                /// Flush should not be done if shouldBeEmptyOnDetach() == false,
-                /// since in this case getTablesIterator() may do some additional work,
-                /// see DatabaseMaterializedMySQL::getTablesIterator()
-                auto table_context = Context::createCopy(getContext());
-                table_context->setInternalQuery(true);
-                /// Do not hold extra shared pointers to tables
-                std::vector<std::pair<StorageID, bool>> tables_to_drop;
-                // NOTE: This means we wait for all tables to be loaded inside getTablesIterator() call in case of `async_load_databases = true`.
-                for (auto iterator = database->getTablesIterator(table_context); iterator->isValid(); iterator->next())
-                {
-                    auto table_ptr = iterator->table();
-                    tables_to_drop.push_back({table_ptr->getStorageID(), table_ptr->isDictionary()});
-                }
-
-                /// Prepare tables for shutdown in parallel.
-                ThreadPoolCallbackRunnerLocal<void> runner(getDatabaseCatalogDropTablesThreadPool().get(), "DropTables");
-                for (const auto & [name, _] : tables_to_drop)
-                {
-                    auto table_ptr = DatabaseCatalog::instance().getTable(name, table_context);
-                    runner([my_table_ptr = std::move(table_ptr)]()
-                    {
-                        my_table_ptr->flushAndPrepareForShutdown();
-                    });
-                }
-                runner.waitForAllToFinishAndRethrowFirstError();
-
-                for (const auto & table : tables_to_drop)
-                {
-                    query_for_table.setTable(table.first.getTableName());
-                    query_for_table.is_dictionary = table.second;
-                    DatabasePtr db;
-                    UUID table_to_wait = UUIDHelpers::Nil;
-                    executeToTableImpl(table_context, query_for_table, db, table_to_wait);
-                    uuids_to_wait.push_back(table_to_wait);
-                }
-            }
-            // only if operation is DETACH
-            if ((!drop || !truncate) && query.sync)
-            {
-                /// Avoid "some tables are still in use" when sync mode is enabled
-                for (const auto & table_uuid : uuids_to_wait)
-                    database->waitDetachedTableNotInUse(table_uuid);
-            }
-
-            /// Protects from concurrent CREATE TABLE queries
-            auto db_guard = DatabaseCatalog::instance().getExclusiveDDLGuardForDatabase(database_name);
-            // only if operation is DETACH
-            if (!drop || !truncate)
-                database->assertCanBeDetached(true);
-
-            /// DETACH or DROP database itself. If TRUNCATE skip dropping/erasing the database.
-            if (!truncate)
-                DatabaseCatalog::instance().detachDatabase(getContext(), database_name, drop, database->shouldBeEmptyOnDetach());
         }
     }
+    // only if operation is DETACH
+    if ((!drop || !truncate) && query.sync)
+    {
+        /// Avoid "some tables are still in use" when sync mode is enabled
+        for (const auto & table_uuid : uuids_to_wait)
+            database->waitDetachedTableNotInUse(table_uuid);
+    }
+
+    /// Protects from concurrent CREATE TABLE queries
+    auto db_guard = DatabaseCatalog::instance().getExclusiveDDLGuardForDatabase(database_name);
+    // only if operation is DETACH
+    if (!drop || !truncate)
+        database->assertCanBeDetached(true);
+
+    /// DETACH or DROP database itself. If TRUNCATE skip dropping/erasing the database.
+    if (!truncate)
+        DatabaseCatalog::instance().detachDatabase(getContext(), database_name, drop, database->shouldBeEmptyOnDetach());
 
     return {};
 }

src/Parsers/FunctionSecretArgumentsFinder.h

@@ -1,10 +1,42 @@
 #pragma once
 
-#include <vector>
+#include <Common/KnownObjectNames.h>
+#include <Core/QualifiedTableName.h>
+#include <base/defines.h>
+#include <boost/algorithm/string/predicate.hpp>
+
 
 namespace DB
 {
 
+class AbstractFunction
+{
+    friend class FunctionSecretArgumentsFinder;
+public:
+    class Argument
+    {
+    public:
+        virtual ~Argument() = default;
+        virtual std::unique_ptr<AbstractFunction> getFunction() const = 0;
+        virtual bool isIdentifier() const = 0;
+        virtual bool tryGetString(String * res, bool allow_identifier) const = 0;
+    };
+    class Arguments
+    {
+    public:
+        virtual ~Arguments() = default;
+        virtual size_t size() const = 0;
+        virtual std::unique_ptr<Argument> at(size_t n) const = 0;
+    };
+
+    virtual ~AbstractFunction() = default;
+    virtual String name() const = 0;
+    bool hasArguments() const { return !!arguments; }
+
+protected:
+    std::unique_ptr<Arguments> arguments;
+};
+
 class FunctionSecretArgumentsFinder
 {
 public:
@@ -23,6 +55,485 @@ public:
             return count != 0 || !nested_maps.empty();
         }
     };
+
+    explicit FunctionSecretArgumentsFinder(std::unique_ptr<AbstractFunction> && function_) : function(std::move(function_)) {}
+
+    FunctionSecretArgumentsFinder::Result getResult() const { return result; }
+
+protected:
+    const std::unique_ptr<AbstractFunction> function;
+    Result result;
+
+    void markSecretArgument(size_t index, bool argument_is_named = false)
+    {
+        if (index >= function->arguments->size())
+            return;
+        if (!result.count)
+        {
+            result.start = index;
+            result.are_named = argument_is_named;
+        }
+        chassert(index >= result.start); /// We always check arguments consecutively
+        result.count = index + 1 - result.start;
+        if (!argument_is_named)
+            result.are_named = false;
+    }
+
+    void findOrdinaryFunctionSecretArguments()
+    {
+        if ((function->name() == "mysql") || (function->name() == "postgresql") || (function->name() == "mongodb"))
+        {
+            /// mysql('host:port', 'database', 'table', 'user', 'password', ...)
+            /// postgresql('host:port', 'database', 'table', 'user', 'password', ...)
+            /// mongodb('host:port', 'database', 'collection', 'user', 'password', ...)
+            findMySQLFunctionSecretArguments();
+        }
+        else if ((function->name() == "s3") || (function->name() == "cosn") || (function->name() == "oss") ||
+                 (function->name() == "deltaLake") || (function->name() == "hudi") || (function->name() == "iceberg") ||
+                 (function->name() == "gcs"))
+        {
+            /// s3('url', 'aws_access_key_id', 'aws_secret_access_key', ...)
+            findS3FunctionSecretArguments(/* is_cluster_function= */ false);
+        }
+        else if (function->name() == "s3Cluster")
+        {
+            /// s3Cluster('cluster_name', 'url', 'aws_access_key_id', 'aws_secret_access_key', ...)
+            findS3FunctionSecretArguments(/* is_cluster_function= */ true);
+        }
+        else if (function->name() == "azureBlobStorage")
+        {
+            /// azureBlobStorage(connection_string|storage_account_url, container_name, blobpath, account_name, account_key, format, compression, structure)
+            findAzureBlobStorageFunctionSecretArguments(/* is_cluster_function= */ false);
+        }
+        else if (function->name() == "azureBlobStorageCluster")
+        {
+            /// azureBlobStorageCluster(cluster, connection_string|storage_account_url, container_name, blobpath, [account_name, account_key, format, compression, structure])
+            findAzureBlobStorageFunctionSecretArguments(/* is_cluster_function= */ true);
+        }
+        else if ((function->name() == "remote") || (function->name() == "remoteSecure"))
+        {
+            /// remote('addresses_expr', 'db', 'table', 'user', 'password', ...)
+            findRemoteFunctionSecretArguments();
+        }
+        else if ((function->name() == "encrypt") || (function->name() == "decrypt") ||
+                    (function->name() == "aes_encrypt_mysql") || (function->name() == "aes_decrypt_mysql") ||
+                    (function->name() == "tryDecrypt"))
+        {
+            /// encrypt('mode', 'plaintext', 'key' [, iv, aad])
+            findEncryptionFunctionSecretArguments();
+        }
+        else if (function->name() == "url")
+        {
+            findURLSecretArguments();
+        }
+    }
+
+    void findMySQLFunctionSecretArguments()
+    {
+        if (isNamedCollectionName(0))
+        {
+            /// mysql(named_collection, ..., password = 'password', ...)
+            findSecretNamedArgument("password", 1);
+        }
+        else
+        {
+            /// mysql('host:port', 'database', 'table', 'user', 'password', ...)
+            markSecretArgument(4);
+        }
+    }
+
+    /// Returns the number of arguments excluding "headers" and "extra_credentials" (which should
+    /// always be at the end). Marks "headers" as secret, if found.
+    size_t excludeS3OrURLNestedMaps()
+    {
+        size_t count = function->arguments->size();
+        while (count > 0)
+        {
+            const auto f = function->arguments->at(count - 1)->getFunction();
+            if (!f)
+                break;
+            if (f->name() == "headers")
+                result.nested_maps.push_back(f->name());
+            else if (f->name() != "extra_credentials")
+                break;
+            count -= 1;
+        }
+        return count;
+    }
+
+    void findS3FunctionSecretArguments(bool is_cluster_function)
+    {
+        /// s3Cluster('cluster_name', 'url', ...) has 'url' as its second argument.
+        size_t url_arg_idx = is_cluster_function ? 1 : 0;
+
+        if (!is_cluster_function && isNamedCollectionName(0))
+        {
+            /// s3(named_collection, ..., secret_access_key = 'secret_access_key', ...)
+            findSecretNamedArgument("secret_access_key", 1);
+            return;
+        }
+
+        /// We should check other arguments first because we don't need to do any replacement in case of
+        /// s3('url', NOSIGN, 'format' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
+        /// s3('url', 'format', 'structure' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
+        size_t count = excludeS3OrURLNestedMaps();
+        if ((url_arg_idx + 3 <= count) && (count <= url_arg_idx + 4))
+        {
+            String second_arg;
+            if (tryGetStringFromArgument(url_arg_idx + 1, &second_arg))
+            {
+                if (boost::iequals(second_arg, "NOSIGN"))
+                    return; /// The argument after 'url' is "NOSIGN".
+
+                if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
+                    return; /// The argument after 'url' is a format: s3('url', 'format', ...)
+            }
+        }
+
+        /// We're going to replace 'aws_secret_access_key' with '[HIDDEN]' for the following signatures:
+        /// s3('url', 'aws_access_key_id', 'aws_secret_access_key', ...)
+        /// s3Cluster('cluster_name', 'url', 'aws_access_key_id', 'aws_secret_access_key', 'format', 'compression')
+        if (url_arg_idx + 2 < count)
+            markSecretArgument(url_arg_idx + 2);
+    }
+
+    void findAzureBlobStorageFunctionSecretArguments(bool is_cluster_function)
+    {
+        /// azureBlobStorage('cluster_name', 'conn_string/storage_account_url', ...) has 'conn_string/storage_account_url' as its second argument.
+        size_t url_arg_idx = is_cluster_function ? 1 : 0;
+
+        if (!is_cluster_function && isNamedCollectionName(0))
+        {
+            /// azureBlobStorage(named_collection, ..., account_key = 'account_key', ...)
+            findSecretNamedArgument("account_key", 1);
+            return;
+        }
+        else if (is_cluster_function && isNamedCollectionName(1))
+        {
+            /// azureBlobStorageCluster(cluster, named_collection, ..., account_key = 'account_key', ...)
+            findSecretNamedArgument("account_key", 2);
+            return;
+        }
+
+        /// We should check other arguments first because we don't need to do any replacement in case storage_account_url is not used
+        /// azureBlobStorage(connection_string|storage_account_url, container_name, blobpath, account_name, account_key, format, compression, structure)
+        /// azureBlobStorageCluster(cluster, connection_string|storage_account_url, container_name, blobpath, [account_name, account_key, format, compression, structure])
+        size_t count = function->arguments->size();
+        if ((url_arg_idx + 4 <= count) && (count <= url_arg_idx + 7))
+        {
+            String second_arg;
+            if (tryGetStringFromArgument(url_arg_idx + 3, &second_arg))
+            {
+                if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
+                    return; /// The argument after 'url' is a format: s3('url', 'format', ...)
+            }
+        }
+
+        /// We're going to replace 'account_key' with '[HIDDEN]' if account_key is used in the signature
+        if (url_arg_idx + 4 < count)
+            markSecretArgument(url_arg_idx + 4);
+    }
+
+    void findURLSecretArguments()
+    {
+        if (!isNamedCollectionName(0))
+            excludeS3OrURLNestedMaps();
+    }
+
+    bool tryGetStringFromArgument(size_t arg_idx, String * res, bool allow_identifier = true) const
+    {
+        if (arg_idx >= function->arguments->size())
+            return false;
+
+        return tryGetStringFromArgument(*function->arguments->at(arg_idx), res, allow_identifier);
+    }
+
+    static bool tryGetStringFromArgument(const AbstractFunction::Argument & argument, String * res, bool allow_identifier = true)
+    {
+        return argument.tryGetString(res, allow_identifier);
+    }
+
+    void findRemoteFunctionSecretArguments()
+    {
+        if (isNamedCollectionName(0))
+        {
+            /// remote(named_collection, ..., password = 'password', ...)
+            findSecretNamedArgument("password", 1);
+            return;
+        }
+
+        /// We're going to replace 'password' with '[HIDDEN'] for the following signatures:
+        /// remote('addresses_expr', db.table, 'user' [, 'password'] [, sharding_key])
+        /// remote('addresses_expr', 'db', 'table', 'user' [, 'password'] [, sharding_key])
+        /// remote('addresses_expr', table_function(), 'user' [, 'password'] [, sharding_key])
+
+        /// But we should check the number of arguments first because we don't need to do any replacements in case of
+        /// remote('addresses_expr', db.table)
+        if (function->arguments->size() < 3)
+            return;
+
+        size_t arg_num = 1;
+
+        /// Skip 1 or 2 arguments with table_function() or db.table or 'db', 'table'.
+        auto table_function = function->arguments->at(arg_num)->getFunction();
+        if (table_function && KnownTableFunctionNames::instance().exists(table_function->name()))
+        {
+            ++arg_num;
+        }
+        else
+        {
+            std::optional<String> database;
+            std::optional<QualifiedTableName> qualified_table_name;
+            if (!tryGetDatabaseNameOrQualifiedTableName(arg_num, database, qualified_table_name))
+            {
+                /// We couldn't evaluate the argument so we don't know whether it is 'db.table' or just 'db'.
+                /// Hence we can't figure out whether we should skip one argument 'user' or two arguments 'table', 'user'
+                /// before the argument 'password'. So it's safer to wipe two arguments just in case.
+                /// The last argument can be also a `sharding_key`, so we need to check that argument is a literal string
+                /// before wiping it (because the `password` argument is always a literal string).
+                if (tryGetStringFromArgument(arg_num + 2, nullptr, /* allow_identifier= */ false))
+                {
+                    /// Wipe either `password` or `user`.
+                    markSecretArgument(arg_num + 2);
+                }
+                if (tryGetStringFromArgument(arg_num + 3, nullptr, /* allow_identifier= */ false))
+                {
+                    /// Wipe either `password` or `sharding_key`.
+                    markSecretArgument(arg_num + 3);
+                }
+                return;
+            }
+
+            /// Skip the current argument (which is either a database name or a qualified table name).
+            ++arg_num;
+            if (database)
+            {
+                /// Skip the 'table' argument if the previous argument was a database name.
+                ++arg_num;
+            }
+        }
+
+        /// Skip username.
+        ++arg_num;
+
+        /// Do our replacement:
+        /// remote('addresses_expr', db.table, 'user', 'password', ...) -> remote('addresses_expr', db.table, 'user', '[HIDDEN]', ...)
+        /// The last argument can be also a `sharding_key`, so we need to check that argument is a literal string
+        /// before wiping it (because the `password` argument is always a literal string).
+        bool can_be_password = tryGetStringFromArgument(arg_num, nullptr, /* allow_identifier= */ false);
+        if (can_be_password)
+            markSecretArgument(arg_num);
+    }
+
+    /// Tries to get either a database name or a qualified table name from an argument.
+    /// Empty string is also allowed (it means the default database).
+    /// The function is used by findRemoteFunctionSecretArguments() to determine how many arguments to skip before a password.
+    bool tryGetDatabaseNameOrQualifiedTableName(
+        size_t arg_idx,
+        std::optional<String> & res_database,
+        std::optional<QualifiedTableName> & res_qualified_table_name) const
+    {
+        res_database.reset();
+        res_qualified_table_name.reset();
+
+        String str;
+        if (!tryGetStringFromArgument(arg_idx, &str, /* allow_identifier= */ true))
+            return false;
+
+        if (str.empty())
+        {
+            res_database = "";
+            return true;
+        }
+
+        auto qualified_table_name = QualifiedTableName::tryParseFromString(str);
+        if (!qualified_table_name)
+            return false;
+
+        if (qualified_table_name->database.empty())
+            res_database = std::move(qualified_table_name->table);
+        else
+            res_qualified_table_name = std::move(qualified_table_name);
+        return true;
+    }
+
+    void findEncryptionFunctionSecretArguments()
+    {
+        if (function->arguments->size() == 0)
+            return;
+
+        /// We replace all arguments after 'mode' with '[HIDDEN]':
+        /// encrypt('mode', 'plaintext', 'key' [, iv, aad]) -> encrypt('mode', '[HIDDEN]')
+        result.start = 1;
+        result.count = function->arguments->size() - 1;
+    }
+
+    void findTableEngineSecretArguments()
+    {
+        const String & engine_name = function->name();
+        if (engine_name == "ExternalDistributed")
+        {
+            /// ExternalDistributed('engine', 'host:port', 'database', 'table', 'user', 'password')
+            findExternalDistributedTableEngineSecretArguments();
+        }
+        else if ((engine_name == "MySQL") || (engine_name == "PostgreSQL") ||
+                    (engine_name == "MaterializedPostgreSQL") || (engine_name == "MongoDB"))
+        {
+            /// MySQL('host:port', 'database', 'table', 'user', 'password', ...)
+            /// PostgreSQL('host:port', 'database', 'table', 'user', 'password', ...)
+            /// MaterializedPostgreSQL('host:port', 'database', 'table', 'user', 'password', ...)
+            /// MongoDB('host:port', 'database', 'collection', 'user', 'password', ...)
+            findMySQLFunctionSecretArguments();
+        }
+        else if ((engine_name == "S3") || (engine_name == "COSN") || (engine_name == "OSS") ||
+                    (engine_name == "DeltaLake") || (engine_name == "Hudi") || (engine_name == "Iceberg") || (engine_name == "S3Queue"))
+        {
+            /// S3('url', ['aws_access_key_id', 'aws_secret_access_key',] ...)
+            findS3TableEngineSecretArguments();
+        }
+        else if (engine_name == "URL")
+        {
+            findURLSecretArguments();
+        }
+    }
+
+    void findExternalDistributedTableEngineSecretArguments()
+    {
+        if (isNamedCollectionName(1))
+        {
+            /// ExternalDistributed('engine', named_collection, ..., password = 'password', ...)
+            findSecretNamedArgument("password", 2);
+        }
+        else
+        {
+            /// ExternalDistributed('engine', 'host:port', 'database', 'table', 'user', 'password')
+            markSecretArgument(5);
+        }
+    }
+
+    void findS3TableEngineSecretArguments()
+    {
+        if (isNamedCollectionName(0))
+        {
+            /// S3(named_collection, ..., secret_access_key = 'secret_access_key')
+            findSecretNamedArgument("secret_access_key", 1);
+            return;
+        }
+
+        /// We should check other arguments first because we don't need to do any replacement in case of
+        /// S3('url', NOSIGN, 'format' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
+        /// S3('url', 'format', 'compression' [, extra_credentials(..)] [, headers(..)])
+        size_t count = excludeS3OrURLNestedMaps();
+        if ((3 <= count) && (count <= 4))
+        {
+            String second_arg;
+            if (tryGetStringFromArgument(1, &second_arg))
+            {
+                if (boost::iequals(second_arg, "NOSIGN"))
+                    return; /// The argument after 'url' is "NOSIGN".
+
+                if (count == 3)
+                {
+                    if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
+                        return; /// The argument after 'url' is a format: S3('url', 'format', ...)
+                }
+            }
+        }
+
+        /// We replace 'aws_secret_access_key' with '[HIDDEN]' for the following signatures:
+        /// S3('url', 'aws_access_key_id', 'aws_secret_access_key')
+        /// S3('url', 'aws_access_key_id', 'aws_secret_access_key', 'format')
+        /// S3('url', 'aws_access_key_id', 'aws_secret_access_key', 'format', 'compression')
+        if (2 < count)
+            markSecretArgument(2);
+    }
+
+    void findDatabaseEngineSecretArguments()
+    {
+        const String & engine_name = function->name();
+        if ((engine_name == "MySQL") || (engine_name == "MaterializeMySQL") ||
+            (engine_name == "MaterializedMySQL") || (engine_name == "PostgreSQL") ||
+            (engine_name == "MaterializedPostgreSQL"))
+        {
+            /// MySQL('host:port', 'database', 'user', 'password')
+            /// PostgreSQL('host:port', 'database', 'user', 'password')
+            findMySQLDatabaseSecretArguments();
+        }
+        else if (engine_name == "S3")
+        {
+            /// S3('url', 'access_key_id', 'secret_access_key')
+            findS3DatabaseSecretArguments();
+        }
+    }
+
+    void findMySQLDatabaseSecretArguments()
+    {
+        if (isNamedCollectionName(0))
+        {
+            /// MySQL(named_collection, ..., password = 'password', ...)
+            findSecretNamedArgument("password", 1);
+        }
+        else
+        {
+            /// MySQL('host:port', 'database', 'user', 'password')
+            markSecretArgument(3);
+        }
+    }
+
+    void findS3DatabaseSecretArguments()
+    {
+        if (isNamedCollectionName(0))
+        {
+            /// S3(named_collection, ..., secret_access_key = 'password', ...)
+            findSecretNamedArgument("secret_access_key", 1);
+        }
+        else
+        {
+            /// S3('url', 'access_key_id', 'secret_access_key')
+            markSecretArgument(2);
+        }
+    }
+
+    void findBackupNameSecretArguments()
+    {
+        const String & engine_name = function->name();
+        if (engine_name == "S3")
+        {
+            /// BACKUP ... TO S3(url, [aws_access_key_id, aws_secret_access_key])
+            markSecretArgument(2);
+        }
+    }
+
+    /// Whether a specified argument can be the name of a named collection?
+    bool isNamedCollectionName(size_t arg_idx) const
+    {
+        if (function->arguments->size() <= arg_idx)
+            return false;
+
+        return function->arguments->at(arg_idx)->isIdentifier();
+    }
+
+    /// Looks for a secret argument with a specified name. This function looks for arguments in format `key=value` where the key is specified.
+    void findSecretNamedArgument(const std::string_view & key, size_t start = 0)
+    {
+        for (size_t i = start; i < function->arguments->size(); ++i)
+        {
+            const auto & argument = function->arguments->at(i);
+            const auto equals_func = argument->getFunction();
+            if (!equals_func || (equals_func->name() != "equals"))
+                continue;
+
+            if (!equals_func->arguments || equals_func->arguments->size() != 2)
+                continue;
+
+            String found_key;
+            if (!tryGetStringFromArgument(*equals_func->arguments->at(0), &found_key))
+                continue;
+
+            if (found_key == key)
+                markSecretArgument(i, /* argument_is_named= */ true);
+        }
+    }
 };
 
 }

src/Parsers/FunctionSecretArgumentsFinderAST.h

@@ -1,35 +1,97 @@
 #pragma once
 
 #include <Parsers/FunctionSecretArgumentsFinder.h>
-#include <Core/QualifiedTableName.h>
 #include <Parsers/ASTFunction.h>
 #include <Parsers/ASTLiteral.h>
 #include <Parsers/ASTIdentifier.h>
-#include <Common/KnownObjectNames.h>
-
-#include <boost/algorithm/string/predicate.hpp>
 
 
 namespace DB
 {
 
-
+class FunctionAST : public AbstractFunction
-/// Finds arguments of a specified function which should not be displayed for most users for security reasons.
-/// That involves passwords and secret keys.
-class FunctionSecretArgumentsFinderAST
 {
 public:
-    explicit FunctionSecretArgumentsFinderAST(const ASTFunction & function_) : function(function_)
+    class ArgumentAST : public Argument
     {
-        if (!function.arguments)
+    public:
+        explicit ArgumentAST(const IAST * argument_) : argument(argument_) {}
+        std::unique_ptr<AbstractFunction> getFunction() const override
+        {
+            if (const auto * f = argument->as<ASTFunction>())
+                return std::make_unique<FunctionAST>(*f);
+            return nullptr;
+        }
+        bool isIdentifier() const override { return argument->as<ASTIdentifier>(); }
+        bool tryGetString(String * res, bool allow_identifier) const override
+        {
+            if (const auto * literal = argument->as<ASTLiteral>())
+            {
+                if (literal->value.getType() != Field::Types::String)
+                    return false;
+                if (res)
+                    *res = literal->value.safeGet<String>();
+                return true;
+            }
+
+            if (allow_identifier)
+            {
+                if (const auto * id = argument->as<ASTIdentifier>())
+                {
+                    if (res)
+                        *res = id->name();
+                    return true;
+                }
+            }
+
+            return false;
+        }
+    private:
+        const IAST * argument = nullptr;
+    };
+
+    class ArgumentsAST : public Arguments
+    {
+    public:
+        explicit ArgumentsAST(const ASTs * arguments_) : arguments(arguments_) {}
+        size_t size() const override { return arguments ? arguments->size() : 0; }
+        std::unique_ptr<Argument> at(size_t n) const override
+        {
+            return std::make_unique<ArgumentAST>(arguments->at(n).get());
+        }
+    private:
+        const ASTs * arguments = nullptr;
+    };
+
+    explicit FunctionAST(const ASTFunction & function_) : function(&function_)
+    {
+        if (!function->arguments)
             return;
 
-        const auto * expr_list = function.arguments->as<ASTExpressionList>();
+        const auto * expr_list = function->arguments->as<ASTExpressionList>();
         if (!expr_list)
             return;
 
-        arguments = &expr_list->children;
+        arguments = std::make_unique<ArgumentsAST>(&expr_list->children);
-        switch (function.kind)
+    }
+
+    String name() const override { return function->name; }
+private:
+    const ASTFunction * function = nullptr;
+};
+
+/// Finds arguments of a specified function which should not be displayed for most users for security reasons.
+/// That involves passwords and secret keys.
+class FunctionSecretArgumentsFinderAST : public FunctionSecretArgumentsFinder
+{
+public:
+    explicit FunctionSecretArgumentsFinderAST(const ASTFunction & function_)
+        : FunctionSecretArgumentsFinder(std::make_unique<FunctionAST>(function_))
+    {
+        if (!function->hasArguments())
+            return;
+
+        switch (function_.kind)
         {
             case ASTFunction::Kind::ORDINARY_FUNCTION: findOrdinaryFunctionSecretArguments(); break;
             case ASTFunction::Kind::WINDOW_FUNCTION: break;
@@ -43,507 +105,7 @@ public:
     }
 
     FunctionSecretArgumentsFinder::Result getResult() const { return result; }
-
-private:
-    const ASTFunction & function;
-    const ASTs * arguments = nullptr;
-    FunctionSecretArgumentsFinder::Result result;
-
-    void markSecretArgument(size_t index, bool argument_is_named = false)
-    {
-        if (index >= arguments->size())
-            return;
-        if (!result.count)
-        {
-            result.start = index;
-            result.are_named = argument_is_named;
-        }
-        chassert(index >= result.start); /// We always check arguments consecutively
-        result.count = index + 1 - result.start;
-        if (!argument_is_named)
-            result.are_named = false;
-    }
-
-    void findOrdinaryFunctionSecretArguments()
-    {
-        if ((function.name == "mysql") || (function.name == "postgresql") || (function.name == "mongodb"))
-        {
-            /// mysql('host:port', 'database', 'table', 'user', 'password', ...)
-            /// postgresql('host:port', 'database', 'table', 'user', 'password', ...)
-            /// mongodb('host:port', 'database', 'collection', 'user', 'password', ...)
-            findMySQLFunctionSecretArguments();
-        }
-        else if ((function.name == "s3") || (function.name == "cosn") || (function.name == "oss") ||
-                 (function.name == "deltaLake") || (function.name == "hudi") || (function.name == "iceberg") ||
-                 (function.name == "gcs"))
-        {
-            /// s3('url', 'aws_access_key_id', 'aws_secret_access_key', ...)
-            findS3FunctionSecretArguments(/* is_cluster_function= */ false);
-        }
-        else if (function.name == "s3Cluster")
-        {
-            /// s3Cluster('cluster_name', 'url', 'aws_access_key_id', 'aws_secret_access_key', ...)
-            findS3FunctionSecretArguments(/* is_cluster_function= */ true);
-        }
-        else if (function.name == "azureBlobStorage")
-        {
-            /// azureBlobStorage(connection_string|storage_account_url, container_name, blobpath, account_name, account_key, format, compression, structure)
-            findAzureBlobStorageFunctionSecretArguments(/* is_cluster_function= */ false);
-        }
-        else if (function.name == "azureBlobStorageCluster")
-        {
-            /// azureBlobStorageCluster(cluster, connection_string|storage_account_url, container_name, blobpath, [account_name, account_key, format, compression, structure])
-            findAzureBlobStorageFunctionSecretArguments(/* is_cluster_function= */ true);
-        }
-        else if ((function.name == "remote") || (function.name == "remoteSecure"))
-        {
-            /// remote('addresses_expr', 'db', 'table', 'user', 'password', ...)
-            findRemoteFunctionSecretArguments();
-        }
-        else if ((function.name == "encrypt") || (function.name == "decrypt") ||
-                    (function.name == "aes_encrypt_mysql") || (function.name == "aes_decrypt_mysql") ||
-                    (function.name == "tryDecrypt"))
-        {
-            /// encrypt('mode', 'plaintext', 'key' [, iv, aad])
-            findEncryptionFunctionSecretArguments();
-        }
-        else if (function.name == "url")
-        {
-            findURLSecretArguments();
-        }
-    }
-
-    void findMySQLFunctionSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// mysql(named_collection, ..., password = 'password', ...)
-            findSecretNamedArgument("password", 1);
-        }
-        else
-        {
-            /// mysql('host:port', 'database', 'table', 'user', 'password', ...)
-            markSecretArgument(4);
-        }
-    }
-
-    /// Returns the number of arguments excluding "headers" and "extra_credentials" (which should
-    /// always be at the end). Marks "headers" as secret, if found.
-    size_t excludeS3OrURLNestedMaps()
-    {
-        size_t count = arguments->size();
-        while (count > 0)
-        {
-            const ASTFunction * f = arguments->at(count - 1)->as<ASTFunction>();
-            if (!f)
-                break;
-            if (f->name == "headers")
-                result.nested_maps.push_back(f->name);
-            else if (f->name != "extra_credentials")
-                break;
-            count -= 1;
-        }
-        return count;
-    }
-
-    void findS3FunctionSecretArguments(bool is_cluster_function)
-    {
-        /// s3Cluster('cluster_name', 'url', ...) has 'url' as its second argument.
-        size_t url_arg_idx = is_cluster_function ? 1 : 0;
-
-        if (!is_cluster_function && isNamedCollectionName(0))
-        {
-            /// s3(named_collection, ..., secret_access_key = 'secret_access_key', ...)
-            findSecretNamedArgument("secret_access_key", 1);
-            return;
-        }
-
-        /// We should check other arguments first because we don't need to do any replacement in case of
-        /// s3('url', NOSIGN, 'format' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
-        /// s3('url', 'format', 'structure' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
-        size_t count = excludeS3OrURLNestedMaps();
-        if ((url_arg_idx + 3 <= count) && (count <= url_arg_idx + 4))
-        {
-            String second_arg;
-            if (tryGetStringFromArgument(url_arg_idx + 1, &second_arg))
-            {
-                if (boost::iequals(second_arg, "NOSIGN"))
-                    return; /// The argument after 'url' is "NOSIGN".
-
-                if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
-                    return; /// The argument after 'url' is a format: s3('url', 'format', ...)
-            }
-        }
-
-        /// We're going to replace 'aws_secret_access_key' with '[HIDDEN]' for the following signatures:
-        /// s3('url', 'aws_access_key_id', 'aws_secret_access_key', ...)
-        /// s3Cluster('cluster_name', 'url', 'aws_access_key_id', 'aws_secret_access_key', 'format', 'compression')
-        if (url_arg_idx + 2 < count)
-            markSecretArgument(url_arg_idx + 2);
-    }
-
-    void findAzureBlobStorageFunctionSecretArguments(bool is_cluster_function)
-    {
-        /// azureBlobStorage('cluster_name', 'conn_string/storage_account_url', ...) has 'conn_string/storage_account_url' as its second argument.
-        size_t url_arg_idx = is_cluster_function ? 1 : 0;
-
-        if (!is_cluster_function && isNamedCollectionName(0))
-        {
-            /// azureBlobStorage(named_collection, ..., account_key = 'account_key', ...)
-            findSecretNamedArgument("account_key", 1);
-            return;
-        }
-        else if (is_cluster_function && isNamedCollectionName(1))
-        {
-            /// azureBlobStorageCluster(cluster, named_collection, ..., account_key = 'account_key', ...)
-            findSecretNamedArgument("account_key", 2);
-            return;
-        }
-
-        /// We should check other arguments first because we don't need to do any replacement in case storage_account_url is not used
-        /// azureBlobStorage(connection_string|storage_account_url, container_name, blobpath, account_name, account_key, format, compression, structure)
-        /// azureBlobStorageCluster(cluster, connection_string|storage_account_url, container_name, blobpath, [account_name, account_key, format, compression, structure])
-        size_t count = arguments->size();
-        if ((url_arg_idx + 4 <= count) && (count <= url_arg_idx + 7))
-        {
-            String second_arg;
-            if (tryGetStringFromArgument(url_arg_idx + 3, &second_arg))
-            {
-                if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
-                    return; /// The argument after 'url' is a format: s3('url', 'format', ...)
-            }
-        }
-
-        /// We're going to replace 'account_key' with '[HIDDEN]' if account_key is used in the signature
-        if (url_arg_idx + 4 < count)
-            markSecretArgument(url_arg_idx + 4);
-    }
-
-    void findURLSecretArguments()
-    {
-        if (!isNamedCollectionName(0))
-            excludeS3OrURLNestedMaps();
-    }
-
-    bool tryGetStringFromArgument(size_t arg_idx, String * res, bool allow_identifier = true) const
-    {
-        if (arg_idx >= arguments->size())
-            return false;
-
-        return tryGetStringFromArgument(*(*arguments)[arg_idx], res, allow_identifier);
-    }
-
-    static bool tryGetStringFromArgument(const IAST & argument, String * res, bool allow_identifier = true)
-    {
-        if (const auto * literal = argument.as<ASTLiteral>())
-        {
-            if (literal->value.getType() != Field::Types::String)
-                return false;
-            if (res)
-                *res = literal->value.safeGet<String>();
-            return true;
-        }
-
-        if (allow_identifier)
-        {
-            if (const auto * id = argument.as<ASTIdentifier>())
-            {
-                if (res)
-                    *res = id->name();
-                return true;
-            }
-        }
-
-        return false;
-    }
-
-    void findRemoteFunctionSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// remote(named_collection, ..., password = 'password', ...)
-            findSecretNamedArgument("password", 1);
-            return;
-        }
-
-        /// We're going to replace 'password' with '[HIDDEN'] for the following signatures:
-        /// remote('addresses_expr', db.table, 'user' [, 'password'] [, sharding_key])
-        /// remote('addresses_expr', 'db', 'table', 'user' [, 'password'] [, sharding_key])
-        /// remote('addresses_expr', table_function(), 'user' [, 'password'] [, sharding_key])
-
-        /// But we should check the number of arguments first because we don't need to do any replacements in case of
-        /// remote('addresses_expr', db.table)
-        if (arguments->size() < 3)
-            return;
-
-        size_t arg_num = 1;
-
-        /// Skip 1 or 2 arguments with table_function() or db.table or 'db', 'table'.
-        const auto * table_function = (*arguments)[arg_num]->as<ASTFunction>();
-        if (table_function && KnownTableFunctionNames::instance().exists(table_function->name))
-        {
-            ++arg_num;
-        }
-        else
-        {
-            std::optional<String> database;
-            std::optional<QualifiedTableName> qualified_table_name;
-            if (!tryGetDatabaseNameOrQualifiedTableName(arg_num, database, qualified_table_name))
-            {
-                /// We couldn't evaluate the argument so we don't know whether it is 'db.table' or just 'db'.
-                /// Hence we can't figure out whether we should skip one argument 'user' or two arguments 'table', 'user'
-                /// before the argument 'password'. So it's safer to wipe two arguments just in case.
-                /// The last argument can be also a `sharding_key`, so we need to check that argument is a literal string
-                /// before wiping it (because the `password` argument is always a literal string).
-                if (tryGetStringFromArgument(arg_num + 2, nullptr, /* allow_identifier= */ false))
-                {
-                    /// Wipe either `password` or `user`.
-                    markSecretArgument(arg_num + 2);
-                }
-                if (tryGetStringFromArgument(arg_num + 3, nullptr, /* allow_identifier= */ false))
-                {
-                    /// Wipe either `password` or `sharding_key`.
-                    markSecretArgument(arg_num + 3);
-                }
-                return;
-            }
-
-            /// Skip the current argument (which is either a database name or a qualified table name).
-            ++arg_num;
-            if (database)
-            {
-                /// Skip the 'table' argument if the previous argument was a database name.
-                ++arg_num;
-            }
-        }
-
-        /// Skip username.
-        ++arg_num;
-
-        /// Do our replacement:
-        /// remote('addresses_expr', db.table, 'user', 'password', ...) -> remote('addresses_expr', db.table, 'user', '[HIDDEN]', ...)
-        /// The last argument can be also a `sharding_key`, so we need to check that argument is a literal string
-        /// before wiping it (because the `password` argument is always a literal string).
-        bool can_be_password = tryGetStringFromArgument(arg_num, nullptr, /* allow_identifier= */ false);
-        if (can_be_password)
-            markSecretArgument(arg_num);
-    }
-
-    /// Tries to get either a database name or a qualified table name from an argument.
-    /// Empty string is also allowed (it means the default database).
-    /// The function is used by findRemoteFunctionSecretArguments() to determine how many arguments to skip before a password.
-    bool tryGetDatabaseNameOrQualifiedTableName(
-        size_t arg_idx,
-        std::optional<String> & res_database,
-        std::optional<QualifiedTableName> & res_qualified_table_name) const
-    {
-        res_database.reset();
-        res_qualified_table_name.reset();
-
-        String str;
-        if (!tryGetStringFromArgument(arg_idx, &str, /* allow_identifier= */ true))
-            return false;
-
-        if (str.empty())
-        {
-            res_database = "";
-            return true;
-        }
-
-        auto qualified_table_name = QualifiedTableName::tryParseFromString(str);
-        if (!qualified_table_name)
-            return false;
-
-        if (qualified_table_name->database.empty())
-            res_database = std::move(qualified_table_name->table);
-        else
-            res_qualified_table_name = std::move(qualified_table_name);
-        return true;
-    }
-
-    void findEncryptionFunctionSecretArguments()
-    {
-        if (arguments->empty())
-            return;
-
-        /// We replace all arguments after 'mode' with '[HIDDEN]':
-        /// encrypt('mode', 'plaintext', 'key' [, iv, aad]) -> encrypt('mode', '[HIDDEN]')
-        result.start = 1;
-        result.count = arguments->size() - 1;
-    }
-
-    void findTableEngineSecretArguments()
-    {
-        const String & engine_name = function.name;
-        if (engine_name == "ExternalDistributed")
-        {
-            /// ExternalDistributed('engine', 'host:port', 'database', 'table', 'user', 'password')
-            findExternalDistributedTableEngineSecretArguments();
-        }
-        else if ((engine_name == "MySQL") || (engine_name == "PostgreSQL") ||
-                    (engine_name == "MaterializedPostgreSQL") || (engine_name == "MongoDB"))
-        {
-            /// MySQL('host:port', 'database', 'table', 'user', 'password', ...)
-            /// PostgreSQL('host:port', 'database', 'table', 'user', 'password', ...)
-            /// MaterializedPostgreSQL('host:port', 'database', 'table', 'user', 'password', ...)
-            /// MongoDB('host:port', 'database', 'collection', 'user', 'password', ...)
-            findMySQLFunctionSecretArguments();
-        }
-        else if ((engine_name == "S3") || (engine_name == "COSN") || (engine_name == "OSS") ||
-                    (engine_name == "DeltaLake") || (engine_name == "Hudi") || (engine_name == "Iceberg") || (engine_name == "S3Queue"))
-        {
-            /// S3('url', ['aws_access_key_id', 'aws_secret_access_key',] ...)
-            findS3TableEngineSecretArguments();
-        }
-        else if (engine_name == "URL")
-        {
-            findURLSecretArguments();
-        }
-    }
-
-    void findExternalDistributedTableEngineSecretArguments()
-    {
-        if (isNamedCollectionName(1))
-        {
-            /// ExternalDistributed('engine', named_collection, ..., password = 'password', ...)
-            findSecretNamedArgument("password", 2);
-        }
-        else
-        {
-            /// ExternalDistributed('engine', 'host:port', 'database', 'table', 'user', 'password')
-            markSecretArgument(5);
-        }
-    }
-
-    void findS3TableEngineSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// S3(named_collection, ..., secret_access_key = 'secret_access_key')
-            findSecretNamedArgument("secret_access_key", 1);
-            return;
-        }
-
-        /// We should check other arguments first because we don't need to do any replacement in case of
-        /// S3('url', NOSIGN, 'format' [, 'compression'] [, extra_credentials(..)] [, headers(..)])
-        /// S3('url', 'format', 'compression' [, extra_credentials(..)] [, headers(..)])
-        size_t count = excludeS3OrURLNestedMaps();
-        if ((3 <= count) && (count <= 4))
-        {
-            String second_arg;
-            if (tryGetStringFromArgument(1, &second_arg))
-            {
-                if (boost::iequals(second_arg, "NOSIGN"))
-                    return; /// The argument after 'url' is "NOSIGN".
-
-                if (count == 3)
-                {
-                    if (second_arg == "auto" || KnownFormatNames::instance().exists(second_arg))
-                        return; /// The argument after 'url' is a format: S3('url', 'format', ...)
-                }
-            }
-        }
-
-        /// We replace 'aws_secret_access_key' with '[HIDDEN]' for the following signatures:
-        /// S3('url', 'aws_access_key_id', 'aws_secret_access_key')
-        /// S3('url', 'aws_access_key_id', 'aws_secret_access_key', 'format')
-        /// S3('url', 'aws_access_key_id', 'aws_secret_access_key', 'format', 'compression')
-        if (2 < count)
-            markSecretArgument(2);
-    }
-
-    void findDatabaseEngineSecretArguments()
-    {
-        const String & engine_name = function.name;
-        if ((engine_name == "MySQL") || (engine_name == "MaterializeMySQL") ||
-            (engine_name == "MaterializedMySQL") || (engine_name == "PostgreSQL") ||
-            (engine_name == "MaterializedPostgreSQL"))
-        {
-            /// MySQL('host:port', 'database', 'user', 'password')
-            /// PostgreSQL('host:port', 'database', 'user', 'password')
-            findMySQLDatabaseSecretArguments();
-        }
-        else if (engine_name == "S3")
-        {
-            /// S3('url', 'access_key_id', 'secret_access_key')
-            findS3DatabaseSecretArguments();
-        }
-    }
-
-    void findMySQLDatabaseSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// MySQL(named_collection, ..., password = 'password', ...)
-            findSecretNamedArgument("password", 1);
-        }
-        else
-        {
-            /// MySQL('host:port', 'database', 'user', 'password')
-            markSecretArgument(3);
-        }
-    }
-
-    void findS3DatabaseSecretArguments()
-    {
-        if (isNamedCollectionName(0))
-        {
-            /// S3(named_collection, ..., secret_access_key = 'password', ...)
-            findSecretNamedArgument("secret_access_key", 1);
-        }
-        else
-        {
-            /// S3('url', 'access_key_id', 'secret_access_key')
-            markSecretArgument(2);
-        }
-    }
-
-    void findBackupNameSecretArguments()
-    {
-        const String & engine_name = function.name;
-        if (engine_name == "S3")
-        {
-            /// BACKUP ... TO S3(url, [aws_access_key_id, aws_secret_access_key])
-            markSecretArgument(2);
-        }
-    }
-
-    /// Whether a specified argument can be the name of a named collection?
-    bool isNamedCollectionName(size_t arg_idx) const
-    {
-        if (arguments->size() <= arg_idx)
-            return false;
-
-        const auto * identifier = (*arguments)[arg_idx]->as<ASTIdentifier>();
-        return identifier != nullptr;
-    }
-
-    /// Looks for a secret argument with a specified name. This function looks for arguments in format `key=value` where the key is specified.
-    void findSecretNamedArgument(const std::string_view & key, size_t start = 0)
-    {
-        for (size_t i = start; i < arguments->size(); ++i)
-        {
-            const auto & argument = (*arguments)[i];
-            const auto * equals_func = argument->as<ASTFunction>();
-            if (!equals_func || (equals_func->name != "equals"))
-                continue;
-
-            const auto * expr_list = equals_func->arguments->as<ASTExpressionList>();
-            if (!expr_list)
-                continue;
-
-            const auto & equal_args = expr_list->children;
-            if (equal_args.size() != 2)
-                continue;
-
-            String found_key;
-            if (!tryGetStringFromArgument(*equal_args[0], &found_key))
-                continue;
-
-            if (found_key == key)
-                markSecretArgument(i, /* argument_is_named= */ true);
-        }
-    }
 };
 
+
 }

src/Planner/PlannerJoinTree.cpp

@@ -778,7 +778,18 @@ JoinTreeQueryPlan buildQueryPlanForTableExpression(QueryTreeNodePtr table_expres
                 auto & prewhere_info = table_expression_query_info.prewhere_info;
                 const auto & prewhere_actions = table_expression_data.getPrewhereFilterActions();
 
-                if (prewhere_actions)
+                std::vector<std::pair<FilterDAGInfo, std::string>> where_filters;
+
+                if (prewhere_actions && select_query_options.build_logical_plan)
+                {
+                    where_filters.emplace_back(
+                        FilterDAGInfo{
+                            prewhere_actions->clone(),
+                            prewhere_actions->getOutputs().at(0)->result_name,
+                            true},
+                        "Prewhere");
+                }
+                else if (prewhere_actions)
                 {
                     prewhere_info = std::make_shared<PrewhereInfo>();
                     prewhere_info->prewhere_actions = prewhere_actions->clone();
@@ -791,7 +802,6 @@ JoinTreeQueryPlan buildQueryPlanForTableExpression(QueryTreeNodePtr table_expres
 
                 const auto & columns_names = table_expression_data.getColumnNames();
 
-                std::vector<std::pair<FilterDAGInfo, std::string>> where_filters;
                 const auto add_filter = [&](FilterDAGInfo & filter_info, std::string description)
                 {
                     bool is_final = table_expression_query_info.table_expression_modifiers

src/Processors/Formats/Impl/ORCBlockOutputFormat.cpp

@@ -78,7 +78,9 @@ void ORCOutputStream::write(const void* buf, size_t length)
 }
 
 ORCBlockOutputFormat::ORCBlockOutputFormat(WriteBuffer & out_, const Block & header_, const FormatSettings & format_settings_)
-    : IOutputFormat(header_, out_), format_settings{format_settings_}, output_stream(out_)
+    : IOutputFormat(header_, out_)
+    , format_settings{format_settings_}
+    , output_stream(out_)
 {
     for (const auto & type : header_.getDataTypes())
         data_types.push_back(recursiveRemoveLowCardinality(type));
@@ -565,6 +567,7 @@ void ORCBlockOutputFormat::prepareWriter()
     schema = orc::createStructType();
     options.setCompression(getORCCompression(format_settings.orc.output_compression_method));
     options.setRowIndexStride(format_settings.orc.output_row_index_stride);
+    options.setDictionaryKeySizeThreshold(format_settings.orc.output_dictionary_key_size_threshold);
     size_t columns_count = header.columns();
     for (size_t i = 0; i != columns_count; ++i)
         schema->addStructField(header.safeGetByPosition(i).name, getORCType(recursiveRemoveLowCardinality(data_types[i])));

src/Processors/QueryPlan/ReadFromMergeTree.cpp

@@ -1,6 +1,8 @@
 #include <Processors/QueryPlan/ReadFromMergeTree.h>
 
+#include <Core/Settings.h>
 #include <IO/Operators.h>
+#include <Interpreters/Cluster.h>
 #include <Interpreters/Context.h>
 #include <Interpreters/ExpressionAnalyzer.h>
 #include <Interpreters/InterpreterSelectQuery.h>
@@ -8,6 +10,8 @@
 #include <Parsers/ASTFunction.h>
 #include <Parsers/ASTIdentifier.h>
 #include <Parsers/ASTSelectQuery.h>
+#include <Parsers/ExpressionListParsers.h>
+#include <Parsers/parseIdentifierOrStringLiteral.h>
 #include <Processors/ConcatProcessor.h>
 #include <Processors/Merges/AggregatingSortedTransform.h>
 #include <Processors/Merges/CollapsingSortedTransform.h>
@@ -16,6 +20,7 @@
 #include <Processors/Merges/ReplacingSortedTransform.h>
 #include <Processors/Merges/SummingSortedTransform.h>
 #include <Processors/Merges/VersionedCollapsingTransform.h>
+#include <Processors/QueryPlan/IQueryPlanStep.h>
 #include <Processors/QueryPlan/PartsSplitter.h>
 #include <Processors/Sources/NullSource.h>
 #include <Processors/Transforms/ExpressionTransform.h>
@@ -24,10 +29,11 @@
 #include <Processors/Transforms/SelectByIndicesTransform.h>
 #include <QueryPipeline/QueryPipelineBuilder.h>
 #include <Storages/MergeTree/MergeTreeDataSelectExecutor.h>
-#include <Storages/MergeTree/MergeTreeIndexVectorSimilarity.h>
 #include <Storages/MergeTree/MergeTreeIndexLegacyVectorSimilarity.h>
-#include <Storages/MergeTree/MergeTreeReadPool.h>
+#include <Storages/MergeTree/MergeTreeIndexMinMax.h>
+#include <Storages/MergeTree/MergeTreeIndexVectorSimilarity.h>
 #include <Storages/MergeTree/MergeTreePrefetchedReadPool.h>
+#include <Storages/MergeTree/MergeTreeReadPool.h>
 #include <Storages/MergeTree/MergeTreeReadPoolInOrder.h>
 #include <Storages/MergeTree/MergeTreeReadPoolParallelReplicas.h>
 #include <Storages/MergeTree/MergeTreeReadPoolParallelReplicasInOrder.h>
@@ -41,11 +47,6 @@
 #include <Common/JSONBuilder.h>
 #include <Common/isLocalAddress.h>
 #include <Common/logger_useful.h>
-#include <Core/Settings.h>
-#include <Processors/QueryPlan/IQueryPlanStep.h>
-#include <Parsers/parseIdentifierOrStringLiteral.h>
-#include <Parsers/ExpressionListParsers.h>
-#include <Storages/MergeTree/MergeTreeIndexMinMax.h>
 
 #include <algorithm>
 #include <iterator>
@@ -381,6 +382,7 @@ Pipe ReadFromMergeTree::readFromPoolParallelReplicas(RangesInDataParts parts_wit
         .all_callback = all_ranges_callback.value(),
         .callback = read_task_callback.value(),
         .number_of_current_replica = number_of_current_replica.value_or(client_info.number_of_current_replica),
+        .total_nodes_count = context->getClusterForParallelReplicas()->getShardsInfo().at(0).getAllNodeCount(),
     };
 
     /// We have a special logic for local replica. It has to read less data, because in some cases it should
@@ -563,6 +565,7 @@ Pipe ReadFromMergeTree::readInOrder(
             .all_callback = all_ranges_callback.value(),
             .callback = read_task_callback.value(),
             .number_of_current_replica = number_of_current_replica.value_or(client_info.number_of_current_replica),
+            .total_nodes_count = context->getClusterForParallelReplicas()->getShardsInfo().at(0).getAllNodeCount(),
         };
 
         auto multiplier = context->getSettingsRef().parallel_replicas_single_task_marks_count_multiplier;

src/Processors/QueryPlan/Serialization.cpp

@@ -552,7 +552,7 @@ static QueryPlanResourceHolder replaceReadingFromTable(QueryPlan::Node & node, Q
     {
         SelectQueryOptions options(QueryProcessingStage::FetchColumns);
         options.ignore_rename_columns = true;
-        InterpreterSelectQueryAnalyzer interpreter(wrapWithUnion(std::move(query)), context, options, column_names);
+        InterpreterSelectQueryAnalyzer interpreter(wrapWithUnion(std::move(query)), context, options);
         reading_plan = std::move(interpreter).extractQueryPlan();
     }
     else

src/Server/CertificateReloader.cpp

@@ -34,8 +34,12 @@ int CertificateReloader::setCertificate(SSL * ssl, const CertificateReloader::Mu
     auto current = pdata->data.get();
     if (!current)
         return -1;
+    return setCertificateCallback(ssl, current.get(), log);
+}
 
-    if (current->certs_chain.empty())
+int setCertificateCallback(SSL * ssl, const CertificateReloader::Data * current_data, LoggerPtr log)
+{
+    if (current_data->certs_chain.empty())
         return -1;
 
     if (auto err = SSL_clear_chain_certs(ssl); err != 1)
@@ -43,12 +47,12 @@ int CertificateReloader::setCertificate(SSL * ssl, const CertificateReloader::Mu
         LOG_ERROR(log, "Clear certificates {}", Poco::Net::Utility::getLastError());
         return -1;
     }
-    if (auto err = SSL_use_certificate(ssl, const_cast<X509 *>(current->certs_chain[0].certificate())); err != 1)
+    if (auto err = SSL_use_certificate(ssl, const_cast<X509 *>(current_data->certs_chain[0].certificate())); err != 1)
     {
         LOG_ERROR(log, "Use certificate {}", Poco::Net::Utility::getLastError());
         return -1;
     }
-    for (auto cert = current->certs_chain.begin() + 1; cert != current->certs_chain.end(); cert++)
+    for (auto cert = current_data->certs_chain.begin() + 1; cert != current_data->certs_chain.end(); cert++)
     {
         if (auto err = SSL_add1_chain_cert(ssl, const_cast<X509 *>(cert->certificate())); err != 1)
         {
@@ -56,7 +60,7 @@ int CertificateReloader::setCertificate(SSL * ssl, const CertificateReloader::Mu
             return -1;
         }
     }
-    if (auto err = SSL_use_PrivateKey(ssl, const_cast<EVP_PKEY *>(static_cast<const EVP_PKEY *>(current->key))); err != 1)
+    if (auto err = SSL_use_PrivateKey(ssl, const_cast<EVP_PKEY *>(static_cast<const EVP_PKEY *>(current_data->key))); err != 1)
     {
         LOG_ERROR(log, "Use private key {}", Poco::Net::Utility::getLastError());
         return -1;

src/Server/CertificateReloader.h

@@ -104,6 +104,9 @@ private:
     mutable std::mutex data_mutex;
 };
 
+/// A callback for OpenSSL
+int setCertificateCallback(SSL * ssl, const CertificateReloader::Data * current_data, LoggerPtr log);
+
 }
 
 #endif

src/Server/TCPHandler.cpp

@@ -1271,7 +1271,7 @@ void TCPHandler::sendReadTaskRequestAssumeLocked()
 void TCPHandler::sendMergeTreeAllRangesAnnouncementAssumeLocked(InitialAllRangesAnnouncement announcement)
 {
     writeVarUInt(Protocol::Server::MergeTreeAllRangesAnnouncement, *out);
-    announcement.serialize(*out);
+    announcement.serialize(*out, client_parallel_replicas_protocol_version);
 
     out->finishChunk();
     out->next();
@@ -1281,7 +1281,7 @@ void TCPHandler::sendMergeTreeAllRangesAnnouncementAssumeLocked(InitialAllRanges
 void TCPHandler::sendMergeTreeReadTaskRequestAssumeLocked(ParallelReadRequest request)
 {
     writeVarUInt(Protocol::Server::MergeTreeReadTaskRequest, *out);
-    request.serialize(*out);
+    request.serialize(*out, client_parallel_replicas_protocol_version);
 
     out->finishChunk();
     out->next();
@@ -1663,6 +1663,9 @@ void TCPHandler::receiveAddendum()
         readStringBinary(proto_send_chunked_cl, *in);
         readStringBinary(proto_recv_chunked_cl, *in);
     }
+
+    if (client_tcp_protocol_version >= DBMS_MIN_REVISION_WITH_VERSIONED_PARALLEL_REPLICAS_PROTOCOL)
+        readVarUInt(client_parallel_replicas_protocol_version, *in);
 }
 
 
@@ -1690,6 +1693,8 @@ void TCPHandler::sendHello()
     writeVarUInt(VERSION_MAJOR, *out);
     writeVarUInt(VERSION_MINOR, *out);
     writeVarUInt(DBMS_TCP_PROTOCOL_VERSION, *out);
+    if (client_tcp_protocol_version >= DBMS_MIN_REVISION_WITH_VERSIONED_PARALLEL_REPLICAS_PROTOCOL)
+        writeVarUInt(DBMS_PARALLEL_REPLICAS_PROTOCOL_VERSION, *out);
     if (client_tcp_protocol_version >= DBMS_MIN_REVISION_WITH_SERVER_TIMEZONE)
         writeStringBinary(DateLUT::instance().getTimeZone(), *out);
     if (client_tcp_protocol_version >= DBMS_MIN_REVISION_WITH_SERVER_DISPLAY_NAME)

src/Server/TCPHandler.h

@@ -191,6 +191,7 @@ private:
     UInt64 client_version_minor = 0;
     UInt64 client_version_patch = 0;
     UInt32 client_tcp_protocol_version = 0;
+    UInt32 client_parallel_replicas_protocol_version = 0;
     String proto_send_chunked_cl = "notchunked";
     String proto_recv_chunked_cl = "notchunked";
     String quota_key;