Fix

A better fix for #67476

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

@@ -226,6 +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)");
         }
 
@@ -287,6 +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)");
         }
 

contrib/icu-cmake/CMakeLists.txt

@@ -481,6 +481,11 @@ if (ARCH_S390X)
 else()
     set(ICUDATA_SOURCE_FILE "${ICUDATA_SOURCE_DIR}/icudt75l_dat.S" )
 endif()
+# ^^ you might be confused how for different little endian platforms (x86, ARM) the same assembly files can be used.
+# These files are indeed assembly but they only contain data ('.long' directive), which makes them portable accross CPUs.
+# Only the endianness and the character set (ASCII, EBCDIC) makes a difference, also see
+# https://unicode-org.github.io/icu/userguide/icu_data/#sharing-icu-data-between-platforms, 'Sharing ICU Data Between Platforms')
+# (and as an experiment, try re-generating the data files on x86 vs. ARM, ... you'll get exactly the same files)
 
 set(ICUDATA_SOURCES
     "${ICUDATA_SOURCE_FILE}"

contrib/libarchive

@@ -1 +1 @@
-Subproject commit 0c21691b177fac5f4cceca2a1ff2ddfa5d60f51c
+Subproject commit 313aa1fa10b657de791e3202c168a6c833bc3543

contrib/libarchive-cmake/CMakeLists.txt

@@ -157,7 +157,7 @@ if (TARGET ch_contrib::zlib)
 endif()
 
 if (TARGET ch_contrib::zstd)
-    target_compile_definitions(_libarchive PUBLIC HAVE_ZSTD_H=1 HAVE_LIBZSTD=1 HAVE_LIBZSTD_COMPRESSOR=1)
+    target_compile_definitions(_libarchive PUBLIC HAVE_ZSTD_H=1 HAVE_LIBZSTD=1 HAVE_ZSTD_compressStream=1)
     target_link_libraries(_libarchive PRIVATE ch_contrib::zstd)
 endif()
 

contrib/libarchive-cmake/config.h

@@ -334,13 +334,16 @@ typedef uint64_t uintmax_t;
 /* #undef ARCHIVE_XATTR_LINUX */
 
 /* Version number of bsdcpio */
-#define BSDCPIO_VERSION_STRING "3.7.0"
+#define BSDCPIO_VERSION_STRING "3.7.4"
 
 /* Version number of bsdtar */
-#define BSDTAR_VERSION_STRING "3.7.0"
+#define BSDTAR_VERSION_STRING "3.7.4"
 
 /* Version number of bsdcat */
-#define BSDCAT_VERSION_STRING "3.7.0"
+#define BSDCAT_VERSION_STRING "3.7.4"
+
+/* Version number of bsdunzip */
+#define BSDUNZIP_VERSION_STRING "3.7.4"
 
 /* Define to 1 if you have the `acl_create_entry' function. */
 /* #undef HAVE_ACL_CREATE_ENTRY */
@@ -642,8 +645,8 @@ typedef uint64_t uintmax_t;
 /* Define to 1 if you have the `getgrnam_r' function. */
 #define HAVE_GETGRNAM_R 1
 
-/* Define to 1 if platform uses `optreset` to reset `getopt` */
-#define HAVE_GETOPT_OPTRESET 1
+/* Define to 1 if you have the `getline' function. */
+#define HAVE_GETLINE 1
 
 /* Define to 1 if you have the `getpid' function. */
 #define HAVE_GETPID 1
@@ -750,6 +753,12 @@ typedef uint64_t uintmax_t;
 /* Define to 1 if you have the `pcreposix' library (-lpcreposix). */
 /* #undef HAVE_LIBPCREPOSIX */
 
+/* Define to 1 if you have the `pcre2-8' library (-lpcre2-8). */
+/* #undef HAVE_LIBPCRE2 */
+
+/* Define to 1 if you have the `pcreposix' library (-lpcre2posix). */
+/* #undef HAVE_LIBPCRE2POSIX */
+
 /* Define to 1 if you have the `xml2' library (-lxml2). */
 #define HAVE_LIBXML2 1
 
@@ -765,9 +774,8 @@ typedef uint64_t uintmax_t;
 /* Define to 1 if you have the `zstd' library (-lzstd). */
 /* #undef HAVE_LIBZSTD */
 
-/* Define to 1 if you have the `zstd' library (-lzstd) with compression
-   support. */
-/* #undef HAVE_LIBZSTD_COMPRESSOR */
+/* Define to 1 if you have the ZSTD_compressStream function. */
+/* #undef HAVE_ZSTD_compressStream */
 
 /* Define to 1 if you have the <limits.h> header file. */
 #define HAVE_LIMITS_H 1
@@ -923,6 +931,9 @@ typedef uint64_t uintmax_t;
 /* Define to 1 if you have the <pcreposix.h> header file. */
 /* #undef HAVE_PCREPOSIX_H */
 
+/* Define to 1 if you have the <pcre2posix.h> header file. */
+/* #undef HAVE_PCRE2POSIX_H */
+
 /* Define to 1 if you have the `pipe' function. */
 #define HAVE_PIPE 1
 
@@ -1029,6 +1040,12 @@ typedef uint64_t uintmax_t;
 /* Define to 1 if you have the `strrchr' function. */
 #define HAVE_STRRCHR 1
 
+/* Define to 1 if the system has the type `struct statfs'. */
+/* #undef HAVE_STRUCT_STATFS */
+
+/* Define to 1 if `f_iosize' is a member of `struct statfs'. */
+/* #undef HAVE_STRUCT_STATFS_F_IOSIZE */
+
 /* Define to 1 if `f_namemax' is a member of `struct statfs'. */
 /* #undef HAVE_STRUCT_STATFS_F_NAMEMAX */
 
@@ -1077,6 +1094,9 @@ typedef uint64_t uintmax_t;
 /* Define to 1 if you have the `symlink' function. */
 #define HAVE_SYMLINK 1
 
+/* Define to 1 if you have the `sysconf' function. */
+#define HAVE_SYSCONF 1
+
 /* Define to 1 if you have the <sys/acl.h> header file. */
 /* #undef HAVE_SYS_ACL_H */
 
@@ -1276,10 +1296,10 @@ typedef uint64_t uintmax_t;
 #define ICONV_CONST
 
 /* Version number of libarchive as a single integer */
-#define LIBARCHIVE_VERSION_NUMBER "3007000"
+#define LIBARCHIVE_VERSION_NUMBER "3007004"
 
 /* Version number of libarchive */
-#define LIBARCHIVE_VERSION_STRING "3.7.0"
+#define LIBARCHIVE_VERSION_STRING "3.7.4"
 
 /* Define to 1 if `lstat' dereferences a symlink specified with a trailing
    slash. */
@@ -1333,7 +1353,7 @@ typedef uint64_t uintmax_t;
 #endif /* SAFE_TO_DEFINE_EXTENSIONS */
 
 /* Version number of package */
-#define VERSION "3.7.0"
+#define VERSION "3.7.4"
 
 /* Number of bits in a file offset, on hosts where this is settable. */
 /* #undef _FILE_OFFSET_BITS */

docker/server/entrypoint.sh

@@ -109,7 +109,7 @@ if [ -n "$CLICKHOUSE_USER" ] && [ "$CLICKHOUSE_USER" != "default" ] || [ -n "$CL
           <networks>
             <ip>::/0</ip>
           </networks>
-          <password>${CLICKHOUSE_PASSWORD}</password>
+          <password><![CDATA[${CLICKHOUSE_PASSWORD//]]>/]]]]><![CDATA[>}]]></password>
           <quota>default</quota>
           <access_management>${CLICKHOUSE_ACCESS_MANAGEMENT}</access_management>
         </${CLICKHOUSE_USER}>

docker/test/base/setup_export_logs.sh

@@ -124,6 +124,8 @@ function setup_logs_replication
     check_logs_credentials || return 0
     __set_connection_args
 
+    echo "My hostname is ${HOSTNAME}"
+
     echo 'Create all configured system logs'
     clickhouse-client --query "SYSTEM FLUSH LOGS"
 
@@ -184,7 +186,17 @@ function setup_logs_replication
             /^TTL /d
             ')
 
-        echo -e "Creating remote destination table ${table}_${hash} with statement:\n${statement}" >&2
+        echo -e "Creating remote destination table ${table}_${hash} with statement:" >&2
+
+        echo "::group::${table}"
+        # there's the only way big "$statement" can be printed without causing EAGAIN error
+        # cat: write error: Resource temporarily unavailable
+        statement_print="${statement}"
+        if [ "${#statement_print}" -gt 4000 ]; then
+          statement_print="${statement::1999}\n…\n${statement:${#statement}-1999}"
+        fi
+        echo -e "$statement_print"
+        echo "::endgroup::"
 
         echo "$statement" | clickhouse-client --database_replicated_initial_query_timeout_sec=10 \
             --distributed_ddl_task_timeout=30 --distributed_ddl_output_mode=throw_only_active \

docs/en/engines/table-engines/integrations/azure-queue.md

@@ -0,0 +1,72 @@
+---
+slug: /en/engines/table-engines/integrations/azure-queue
+sidebar_position: 181
+sidebar_label: AzureQueue
+---
+
+# AzureQueue Table Engine
+
+This engine provides an integration with [Azure Blob Storage](https://azure.microsoft.com/en-us/products/storage/blobs) ecosystem, allowing streaming data import.
+
+## Create Table {#creating-a-table}
+
+``` sql
+CREATE TABLE test (name String, value UInt32)
+    ENGINE = AzureQueue(...)
+    [SETTINGS]
+    [mode = '',]
+    [after_processing = 'keep',]
+    [keeper_path = '',]
+    ...
+```
+
+**Engine parameters**
+
+`AzureQueue` parameters are the same as `AzureBlobStorage` table engine supports. See parameters section [here](../../../engines/table-engines/integrations/azureBlobStorage.md).
+
+**Example**
+
+```sql
+CREATE TABLE azure_queue_engine_table (name String, value UInt32)
+ENGINE=AzureQueue('DefaultEndpointsProtocol=http;AccountName=devstoreaccount1;AccountKey=Eby8vdM02xNOcqFlqUwJPLlmEtlCDXJ1OUzFT50uSRZ6IFsuFq2UVErCz4I6tq/K1SZFPTOtr/KBHBeksoGMGw==;BlobEndpoint=http://azurite1:10000/devstoreaccount1/data/')
+SETTINGS
+    mode = 'unordered'
+```
+
+## Settings {#settings}
+
+The set of supported settings is the same as for `S3Queue` table engine, but without `s3queue_` prefix. See [full list of settings settings](../../../engines/table-engines/integrations/s3queue.md#settings).
+
+## Description {#description}
+
+`SELECT` is not particularly useful for streaming import (except for debugging), because each file can be imported only once. It is more practical to create real-time threads using [materialized views](../../../sql-reference/statements/create/view.md). To do this:
+
+1.  Use the engine to create a table for consuming from specified path in S3 and consider it a data stream.
+2.  Create a table with the desired structure.
+3.  Create a materialized view that converts data from the engine and puts it into a previously created table.
+
+When the `MATERIALIZED VIEW` joins the engine, it starts collecting data in the background.
+
+Example:
+
+``` sql
+  CREATE TABLE azure_queue_engine_table (name String, value UInt32)
+    ENGINE=AzureQueue('<endpoint>', 'CSV', 'gzip')
+    SETTINGS
+        mode = 'unordered';
+
+  CREATE TABLE stats (name String, value UInt32)
+    ENGINE = MergeTree() ORDER BY name;
+
+  CREATE MATERIALIZED VIEW consumer TO stats
+    AS SELECT name, value FROM azure_queue_engine_table;
+
+  SELECT * FROM stats ORDER BY name;
+```
+
+## Virtual columns {#virtual-columns}
+
+- `_path` — Path to the file.
+- `_file` — Name of the file.
+
+For more information about virtual columns see [here](../../../engines/table-engines/index.md#table_engines-virtual_columns).

docs/en/engines/table-engines/integrations/s3.md

@@ -35,7 +35,7 @@ CREATE TABLE s3_engine_table (name String, value UInt32)
     [SETTINGS ...]
 ```
 
-### Engine parameters
+### Engine parameters {#parameters}
 
 - `path` — Bucket url with path to file. Supports following wildcards in readonly mode: `*`, `**`, `?`, `{abc,def}` and `{N..M}` where `N`, `M` — numbers, `'abc'`, `'def'` — strings. For more information see [below](#wildcards-in-path).
 - `NOSIGN` - If this keyword is provided in place of credentials, all the requests will not be signed.

docs/en/engines/table-engines/integrations/s3queue.md

@@ -5,6 +5,7 @@ sidebar_label: S3Queue
 ---
 
 # S3Queue Table Engine
+
 This engine provides integration with [Amazon S3](https://aws.amazon.com/s3/) ecosystem and allows streaming import. This engine is similar to the [Kafka](../../../engines/table-engines/integrations/kafka.md), [RabbitMQ](../../../engines/table-engines/integrations/rabbitmq.md) engines, but provides S3-specific features.
 
 ## Create Table {#creating-a-table}
@@ -16,27 +17,25 @@ CREATE TABLE s3_queue_engine_table (name String, value UInt32)
     [mode = '',]
     [after_processing = 'keep',]
     [keeper_path = '',]
-    [s3queue_loading_retries = 0,]
-    [s3queue_processing_threads_num = 1,]
-    [s3queue_enable_logging_to_s3queue_log = 0,]
-    [s3queue_polling_min_timeout_ms = 1000,]
-    [s3queue_polling_max_timeout_ms = 10000,]
-    [s3queue_polling_backoff_ms = 0,]
-    [s3queue_tracked_file_ttl_sec = 0,]
-    [s3queue_tracked_files_limit = 1000,]
-    [s3queue_cleanup_interval_min_ms = 10000,]
-    [s3queue_cleanup_interval_max_ms = 30000,]
+    [loading_retries = 0,]
+    [processing_threads_num = 1,]
+    [enable_logging_to_s3queue_log = 0,]
+    [polling_min_timeout_ms = 1000,]
+    [polling_max_timeout_ms = 10000,]
+    [polling_backoff_ms = 0,]
+    [tracked_file_ttl_sec = 0,]
+    [tracked_files_limit = 1000,]
+    [cleanup_interval_min_ms = 10000,]
+    [cleanup_interval_max_ms = 30000,]
 ```
 
-Starting with `24.7` settings without `s3queue_` prefix are also supported.
+:::warning
+Before `24.7`, it is required to use `s3queue_` prefix for all settings apart from `mode`, `after_processing` and `keeper_path`.
+:::
 
 **Engine parameters**
 
-- `path` — Bucket url with path to file. Supports following wildcards in readonly mode: `*`, `**`, `?`, `{abc,def}` and `{N..M}` where `N`, `M` — numbers, `'abc'`, `'def'` — strings. For more information see [below](#wildcards-in-path).
-- `NOSIGN` - If this keyword is provided in place of credentials, all the requests will not be signed.
-- `format` — The [format](../../../interfaces/formats.md#formats) of the file.
-- `aws_access_key_id`, `aws_secret_access_key` - Long-term credentials for the [AWS](https://aws.amazon.com/) account user.  You can use these to authenticate your requests. Parameter is optional. If credentials are not specified, they are used from the configuration file. For more information see [Using S3 for Data Storage](../mergetree-family/mergetree.md#table_engine-mergetree-s3).
-- `compression` — Compression type. Supported values: `none`, `gzip/gz`, `brotli/br`, `xz/LZMA`, `zstd/zst`. Parameter is optional. By default, it will autodetect compression by file extension.
+`S3Queue` parameters are the same as `S3` table engine supports. See parameters section [here](../../../engines/table-engines/integrations/s3.md#parameters).
 
 **Example**
 

docs/en/interfaces/formats.md

@@ -1396,6 +1396,7 @@ SELECT * FROM json_each_row_nested
 - [input_format_json_ignore_unknown_keys_in_named_tuple](/docs/en/operations/settings/settings-formats.md/#input_format_json_ignore_unknown_keys_in_named_tuple) - ignore unknown keys in json object for named tuples. Default value - `false`.
 - [input_format_json_compact_allow_variable_number_of_columns](/docs/en/operations/settings/settings-formats.md/#input_format_json_compact_allow_variable_number_of_columns) - allow variable number of columns in JSONCompact/JSONCompactEachRow format, ignore extra columns and use default values on missing columns. Default value - `false`.
 - [input_format_json_throw_on_bad_escape_sequence](/docs/en/operations/settings/settings-formats.md/#input_format_json_throw_on_bad_escape_sequence) - throw an exception if JSON string contains bad escape sequence. If disabled, bad escape sequences will remain as is in the data. Default value - `true`.
+- [input_format_json_empty_as_default](/docs/en/operations/settings/settings-formats.md/#input_format_json_empty_as_default) - treat empty fields in JSON input as default values. Default value - `false`. For complex default expressions [input_format_defaults_for_omitted_fields](/docs/en/operations/settings/settings-formats.md/#input_format_defaults_for_omitted_fields) must be enabled too.
 - [output_format_json_quote_64bit_integers](/docs/en/operations/settings/settings-formats.md/#output_format_json_quote_64bit_integers) - controls quoting of 64-bit integers in JSON output format. Default value - `true`.
 - [output_format_json_quote_64bit_floats](/docs/en/operations/settings/settings-formats.md/#output_format_json_quote_64bit_floats) - controls quoting of 64-bit floats in JSON output format. Default value - `false`.
 - [output_format_json_quote_denormals](/docs/en/operations/settings/settings-formats.md/#output_format_json_quote_denormals) - enables '+nan', '-nan', '+inf', '-inf' outputs in JSON output format. Default value - `false`.

docs/en/interfaces/third-party/gui.md

@@ -233,6 +233,16 @@ Features:
 - Useful tools: Zookeeper data exploration, query EXPLAIN, kill queries, etc.
 - Visualization metric charts: queries and resource usage, number of merges/mutation, merge performance, query performance, etc.
 
+### CKibana {#ckibana}
+
+[CKibana](https://github.com/TongchengOpenSource/ckibana) is a lightweight service that allows you to effortlessly search, explore, and visualize ClickHouse data using the native Kibana UI.
+
+Features:
+
+- Translates chart requests from the native Kibana UI into ClickHouse query syntax.
+- Supports advanced features such as sampling and caching to enhance query performance.
+- Minimizes the learning cost for users after migrating from ElasticSearch to ClickHouse.
+
 ## Commercial {#commercial}
 
 ### DataGrip {#datagrip}

docs/en/operations/external-authenticators/ssl-x509.md

@@ -6,7 +6,7 @@ import SelfManaged from '@site/docs/en/_snippets/_self_managed_only_no_roadmap.m
 
 <SelfManaged />
 
-[SSL 'strict' option](../server-configuration-parameters/settings.md#server_configuration_parameters-openssl) enables mandatory certificate validation for the incoming connections. In this case, only connections with trusted certificates can be established. Connections with untrusted certificates will be rejected. Thus, certificate validation allows to uniquely authenticate an incoming connection. `Common Name` or `subjectAltName extension` field of the certificate is used to identify the connected user. This allows to associate multiple certificates with the same user. Additionally, reissuing and revoking of the certificates does not affect the ClickHouse configuration.
+[SSL 'strict' option](../server-configuration-parameters/settings.md#server_configuration_parameters-openssl) enables mandatory certificate validation for the incoming connections. In this case, only connections with trusted certificates can be established. Connections with untrusted certificates will be rejected. Thus, certificate validation allows to uniquely authenticate an incoming connection. `Common Name` or `subjectAltName extension` field of the certificate is used to identify the connected user. `subjectAltName extension` supports the usage of one wildcard '*' in the server configuration. This allows to associate multiple certificates with the same user. Additionally, reissuing and revoking of the certificates does not affect the ClickHouse configuration.
 
 To enable SSL certificate authentication, a list of `Common Name`'s or `Subject Alt Name`'s for each ClickHouse user must be specified in the settings file `users.xml `:
 
@@ -30,6 +30,12 @@ To enable SSL certificate authentication, a list of `Common Name`'s or `Subject
             </ssl_certificates>
             <!-- Other settings -->
         </user_name_2>
+        <user_name_3>
+            <ssl_certificates>
+                <!-- Wildcard support -->
+                <subject_alt_name>URI:spiffe://foo.com/*/bar</subject_alt_name>
+            </ssl_certificates>
+        </user_name_3>
     </users>
 </clickhouse>
 ```

docs/en/operations/settings/settings-formats.md

@@ -752,6 +752,17 @@ Possible values:
 
 Default value: 0.
 
+### input_format_json_empty_as_default {#input_format_json_empty_as_default}
+
+When enabled, replace empty input fields in JSON with default values. For complex default expressions `input_format_defaults_for_omitted_fields` must be enabled too.
+
+Possible values:
+
++ 0 — Disable.
++ 1 — Enable.
+
+Default value: 0.
+
 ## TSV format settings {#tsv-format-settings}
 
 ### input_format_tsv_empty_as_default {#input_format_tsv_empty_as_default}

docs/en/operations/utilities/clickhouse-keeper-client.md

@@ -47,13 +47,15 @@ keeper foo bar
 
 -   `ls '[path]'` -- Lists the nodes for the given path (default: cwd)
 -   `cd '[path]'` -- Changes the working path (default `.`)
+-   `cp '<src>' '<dest>'`  -- Copies 'src' node to 'dest' path
+-   `mv '<src>' '<dest>'`  -- Moves 'src' node to the 'dest' path
 -   `exists '<path>'` -- Returns `1` if node exists, `0` otherwise
 -   `set '<path>' <value> [version]` -- Updates the node's value. Only updates if version matches (default: -1)
 -   `create '<path>' <value> [mode]` -- Creates new node with the set value
 -   `touch '<path>'` -- Creates new node with an empty string as value. Doesn't throw an exception if the node already exists
 -   `get '<path>'` -- Returns the node's value
 -   `rm '<path>' [version]` -- Removes the node only if version matches (default: -1)
--   `rmr '<path>'` -- Recursively deletes path. Confirmation required
+-   `rmr '<path>' [limit]` -- Recursively deletes path if the subtree size is smaller than the limit. Confirmation required (default limit = 100)
 -   `flwc <command>` -- Executes four-letter-word command
 -   `help` -- Prints this message
 -   `get_direct_children_number '[path]'` -- Get numbers of direct children nodes under a specific path

docs/en/sql-reference/functions/date-time-functions.md

@@ -1636,37 +1636,329 @@ SELECT toStartOfInterval(toDateTime('2023-01-01 14:45:00'), INTERVAL 1 MINUTE, t
 
 Converts a date with time to a certain fixed date, while preserving the time.
 
+**Syntax**
+
+```sql
+toTime(date[,timezone])
+```
+
+**Arguments**
+
+- `date` — Date to convert to a time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+- `timezone` (optional) — Timezone for the returned value. [String](../data-types/string.md).
+
+**Returned value**
+
+- DateTime with date equated to `1970-01-02` while preserving the time. [DateTime](../data-types/datetime.md).
+
+:::note
+If the `date` input argument contained sub-second components,
+they will be dropped in the returned `DateTime` value with second-accuracy.
+:::
+
+**Example**
+
+Query:
+
+```sql
+SELECT toTime(toDateTime64('1970-12-10 01:20:30.3000',3)) AS result, toTypeName(result);
+```
+
+Result:
+
+```response
+┌──────────────result─┬─toTypeName(result)─┐
+│ 1970-01-02 01:20:30 │ DateTime           │
+└─────────────────────┴────────────────────┘
+```
+
 ## toRelativeYearNum
 
-Converts a date, or date with time, to the number of the year, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of years elapsed since a certain fixed point in the past. 
+
+**Syntax**
+
+```sql
+toRelativeYearNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of years from a fixed reference point in the past. [UInt16](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+    toRelativeYearNum(toDate('2002-12-08')) AS y1,
+    toRelativeYearNum(toDate('2010-10-26')) AS y2
+```
+
+Result:
+
+```response
+┌───y1─┬───y2─┐
+│ 2002 │ 2010 │
+└──────┴──────┘
+```
 
 ## toRelativeQuarterNum
 
-Converts a date, or date with time, to the number of the quarter, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of quarters elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeQuarterNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of quarters from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeQuarterNum(toDate('1993-11-25')) AS q1,
+  toRelativeQuarterNum(toDate('2005-01-05')) AS q2
+```
+
+Result:
+
+```response
+┌───q1─┬───q2─┐
+│ 7975 │ 8020 │
+└──────┴──────┘
+```
 
 ## toRelativeMonthNum
 
-Converts a date, or date with time, to the number of the month, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of months elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeMonthNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of months from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeMonthNum(toDate('2001-04-25')) AS m1,
+  toRelativeMonthNum(toDate('2009-07-08')) AS m2
+```
+
+Result:
+
+```response
+┌────m1─┬────m2─┐
+│ 24016 │ 24115 │
+└───────┴───────┘
+```
 
 ## toRelativeWeekNum
 
-Converts a date, or date with time, to the number of the week, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of weeks elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeWeekNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of weeks from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeWeekNum(toDate('2000-02-29')) AS w1,
+  toRelativeWeekNum(toDate('2001-01-12')) AS w2
+```
+
+Result:
+
+```response
+┌───w1─┬───w2─┐
+│ 1574 │ 1619 │
+└──────┴──────┘
+```
 
 ## toRelativeDayNum
 
-Converts a date, or date with time, to the number of the day, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of days elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeDayNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of days from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeDayNum(toDate('1993-10-05')) AS d1,
+  toRelativeDayNum(toDate('2000-09-20')) AS d2
+```
+
+Result:
+
+```response
+┌───d1─┬────d2─┐
+│ 8678 │ 11220 │
+└──────┴───────┘
+```
 
 ## toRelativeHourNum
 
-Converts a date, or date with time, to the number of the hour, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of hours elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeHourNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of hours from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeHourNum(toDateTime('1993-10-05 05:20:36')) AS h1,
+  toRelativeHourNum(toDateTime('2000-09-20 14:11:29')) AS h2
+```
+
+Result:
+
+```response
+┌─────h1─┬─────h2─┐
+│ 208276 │ 269292 │
+└────────┴────────┘
+```
 
 ## toRelativeMinuteNum
 
-Converts a date, or date with time, to the number of the minute, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of minutes elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeMinuteNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of minutes from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeMinuteNum(toDateTime('1993-10-05 05:20:36')) AS m1,
+  toRelativeMinuteNum(toDateTime('2000-09-20 14:11:29')) AS m2
+```
+
+Result:
+
+```response
+┌───────m1─┬───────m2─┐
+│ 12496580 │ 16157531 │
+└──────────┴──────────┘
+```
 
 ## toRelativeSecondNum
 
-Converts a date, or date with time, to the number of the second, starting from a certain fixed point in the past.
+Converts a date, or date with time, to the number of the seconds elapsed since a certain fixed point in the past.
+
+**Syntax**
+
+```sql
+toRelativeSecondNum(date)
+```
+
+**Arguments**
+
+- `date` — Date or date with time. [Date](../data-types/date.md)/[DateTime](../data-types/datetime.md)/[DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- The number of seconds from a fixed reference point in the past. [UInt32](../data-types/int-uint.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+  toRelativeSecondNum(toDateTime('1993-10-05 05:20:36')) AS s1,
+  toRelativeSecondNum(toDateTime('2000-09-20 14:11:29')) AS s2
+```
+
+Result:
+
+```response
+┌────────s1─┬────────s2─┐
+│ 749794836 │ 969451889 │
+└───────────┴───────────┘
+```
 
 ## toISOYear
 

docs/en/sql-reference/functions/type-conversion-functions.md

@@ -3906,7 +3906,7 @@ Result:
 
 ## toDateTime64
 
-Converts the argument to the [DateTime64](../data-types/datetime64.md) data type.
+Converts an input value to a value of type [DateTime64](../data-types/datetime64.md).
 
 **Syntax**
 
@@ -3918,7 +3918,7 @@ toDateTime64(expr, scale, [timezone])
 
 - `expr` — The value. [String](../data-types/string.md), [UInt32](../data-types/int-uint.md), [Float](../data-types/float.md) or [DateTime](../data-types/datetime.md).
 - `scale` - Tick size (precision): 10<sup>-precision</sup> seconds. Valid range: [ 0 : 9 ].
-- `timezone` - Time zone of the specified datetime64 object.
+- `timezone` (optional) - Time zone of the specified datetime64 object.
 
 **Returned value**
 
@@ -3977,10 +3977,137 @@ SELECT toDateTime64('2019-01-01 00:00:00', 3, 'Asia/Istanbul') AS value, toTypeN
 
 ## toDateTime64OrZero
 
+Like [toDateTime64](#todatetime64), this function converts an input value to a value of type [DateTime64](../data-types/datetime64.md) but returns the min value of [DateTime64](../data-types/datetime64.md) if an invalid argument is received.
+
+**Syntax**
+
+``` sql
+toDateTime64OrZero(expr, scale, [timezone])
+```
+
+**Arguments**
+
+- `expr` — The value. [String](../data-types/string.md), [UInt32](../data-types/int-uint.md), [Float](../data-types/float.md) or [DateTime](../data-types/datetime.md).
+- `scale` - Tick size (precision): 10<sup>-precision</sup> seconds. Valid range: [ 0 : 9 ].
+- `timezone` (optional) - Time zone of the specified DateTime64 object.
+
+**Returned value**
+
+- A calendar date and time of day, with sub-second precision, otherwise the minimum value of `DateTime64`: `1970-01-01 01:00:00.000`. [DateTime64](../data-types/datetime64.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT toDateTime64OrZero('2008-10-12 00:00:00 00:30:30', 3) AS invalid_arg
+```
+
+Result:
+
+```response
+┌─────────────invalid_arg─┐
+│ 1970-01-01 01:00:00.000 │
+└─────────────────────────┘
+```
+
+**See also**
+
+- [toDateTime64](#todatetime64).
+- [toDateTime64OrNull](#todatetime64ornull).
+- [toDateTime64OrDefault](#todatetime64ordefault).
+
 ## toDateTime64OrNull
 
+Like [toDateTime64](#todatetime64), this function converts an input value to a value of type [DateTime64](../data-types/datetime64.md) but returns `NULL` if an invalid argument is received.
+
+**Syntax**
+
+``` sql
+toDateTime64OrNull(expr, scale, [timezone])
+```
+
+**Arguments**
+
+- `expr` — The value. [String](../data-types/string.md), [UInt32](../data-types/int-uint.md), [Float](../data-types/float.md) or [DateTime](../data-types/datetime.md).
+- `scale` - Tick size (precision): 10<sup>-precision</sup> seconds. Valid range: [ 0 : 9 ].
+- `timezone` (optional) - Time zone of the specified DateTime64 object.
+
+**Returned value**
+
+- A calendar date and time of day, with sub-second precision, otherwise `NULL`. [DateTime64](../data-types/datetime64.md)/[NULL](../data-types/nullable.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+    toDateTime64OrNull('1976-10-18 00:00:00.30', 3) AS valid_arg,
+    toDateTime64OrNull('1976-10-18 00:00:00 30', 3) AS invalid_arg
+```
+
+Result:
+
+```response
+┌───────────────valid_arg─┬─invalid_arg─┐
+│ 1976-10-18 00:00:00.300 │        ᴺᵁᴸᴸ │
+└─────────────────────────┴─────────────┘
+```
+
+**See also**
+
+- [toDateTime64](#todatetime64).
+- [toDateTime64OrZero](#todatetime64orzero).
+- [toDateTime64OrDefault](#todatetime64ordefault).
+
 ## toDateTime64OrDefault
 
+Like [toDateTime64](#todatetime64), this function converts an input value to a value of type [DateTime64](../data-types/datetime64.md), 
+but returns either the default value of [DateTime64](../data-types/datetime64.md)
+or the provided default if an invalid argument is received.
+
+**Syntax**
+
+``` sql
+toDateTime64OrNull(expr, scale, [timezone, default])
+```
+
+**Arguments**
+
+- `expr` — The value. [String](../data-types/string.md), [UInt32](../data-types/int-uint.md), [Float](../data-types/float.md) or [DateTime](../data-types/datetime.md).
+- `scale` - Tick size (precision): 10<sup>-precision</sup> seconds. Valid range: [ 0 : 9 ].
+- `timezone` (optional) - Time zone of the specified DateTime64 object.
+- `default` (optional) - Default value to return if an invalid argument is received. [DateTime64](../data-types/datetime64.md).
+
+**Returned value**
+
+- A calendar date and time of day, with sub-second precision, otherwise the minimum value of `DateTime64` or the `default` value if provided. [DateTime64](../data-types/datetime64.md).
+
+**Example**
+
+Query:
+
+```sql
+SELECT
+    toDateTime64OrDefault('1976-10-18 00:00:00 30', 3) AS invalid_arg,
+    toDateTime64OrDefault('1976-10-18 00:00:00 30', 3, 'UTC', toDateTime64('2001-01-01 00:00:00.00',3)) AS invalid_arg_with_default
+```
+
+Result:
+
+```response
+┌─────────────invalid_arg─┬─invalid_arg_with_default─┐
+│ 1970-01-01 01:00:00.000 │  2000-12-31 23:00:00.000 │
+└─────────────────────────┴──────────────────────────┘
+```
+
+**See also**
+
+- [toDateTime64](#todatetime64).
+- [toDateTime64OrZero](#todatetime64orzero).
+- [toDateTime64OrNull](#todatetime64ornull).
+
 ## toDecimal32
 
 Converts an input value to a value of type [`Decimal(9, S)`](../data-types/decimal.md) with scale of `S`. Throws an exception in case of an error.

programs/keeper-client/Commands.cpp

@@ -506,14 +506,23 @@ bool RMRCommand::parse(IParser::Pos & pos, std::shared_ptr<ASTKeeperQuery> & nod
         return false;
     node->args.push_back(std::move(path));
 
+    ASTPtr remove_nodes_limit;
+    if (ParserUnsignedInteger{}.parse(pos, remove_nodes_limit, expected))
+        node->args.push_back(remove_nodes_limit->as<ASTLiteral &>().value);
+    else
+        node->args.push_back(UInt64(100));
+
     return true;
 }
 
 void RMRCommand::execute(const ASTKeeperQuery * query, KeeperClient * client) const
 {
     String path = client->getAbsolutePath(query->args[0].safeGet<String>());
+    UInt64 remove_nodes_limit = query->args[1].safeGet<UInt64>();
+
     client->askConfirmation(
-        "You are going to recursively delete path " + path, [client, path] { client->zookeeper->removeRecursive(path); });
+        "You are going to recursively delete path " + path,
+        [client, path, remove_nodes_limit] { client->zookeeper->removeRecursive(path, static_cast<UInt32>(remove_nodes_limit)); });
 }
 
 bool ReconfigCommand::parse(IParser::Pos & pos, std::shared_ptr<ASTKeeperQuery> & node, DB::Expected & expected) const

programs/keeper-client/Commands.h

@@ -184,7 +184,7 @@ class RMRCommand : public IKeeperClientCommand
 
     void execute(const ASTKeeperQuery * query, KeeperClient * client) const override;
 
-    String getHelpMessage() const override { return "{} <path> -- Recursively deletes path. Confirmation required"; }
+    String getHelpMessage() const override { return "{} <path> [limit] -- Recursively deletes path if the subtree size is smaller than the limit. Confirmation required (default limit = 100)"; }
 };
 
 class ReconfigCommand : public IKeeperClientCommand

programs/keeper/Keeper.cpp

@@ -11,6 +11,7 @@
 #include <Core/ServerUUID.h>
 #include <Common/logger_useful.h>
 #include <Common/CgroupsMemoryUsageObserver.h>
+#include <Common/MemoryWorker.h>
 #include <Common/ErrorHandlers.h>
 #include <Common/assertProcessUserMatchesDataOwner.h>
 #include <Common/makeSocketAddress.h>
@@ -384,6 +385,9 @@ try
         LOG_INFO(log, "Background threads finished in {} ms", watch.elapsedMilliseconds());
     });
 
+    MemoryWorker memory_worker(config().getUInt64("memory_worker_period_ms", 0));
+    memory_worker.start();
+
     static ServerErrorHandler error_handler;
     Poco::ErrorHandler::set(&error_handler);
 
@@ -425,8 +429,9 @@ try
             for (const auto & server : *servers)
                 metrics.emplace_back(ProtocolServerMetrics{server.getPortName(), server.currentThreads(), server.refusedConnections()});
             return metrics;
-        }
-    );
+        },
+        /*update_jemalloc_epoch_=*/memory_worker.getSource() != MemoryWorker::MemoryUsageSource::Jemalloc,
+        /*update_rss_=*/memory_worker.getSource() == MemoryWorker::MemoryUsageSource::None);
 
     std::vector<std::string> listen_hosts = DB::getMultipleValuesFromConfig(config(), "", "listen_host");
 
@@ -655,7 +660,6 @@ try
     GWPAsan::initFinished();
 #endif
 
-
     LOG_INFO(log, "Ready for connections.");
 
     waitForTerminationRequest();

programs/server/Server.cpp

@@ -11,7 +11,6 @@
 #include <Poco/Util/HelpFormatter.h>
 #include <Poco/Environment.h>
 #include <Poco/Config.h>
-#include <Common/Jemalloc.h>
 #include <Common/scope_guard_safe.h>
 #include <Common/logger_useful.h>
 #include <base/phdr_cache.h>
@@ -25,6 +24,7 @@
 #include <base/Numa.h>
 #include <Common/PoolId.h>
 #include <Common/MemoryTracker.h>
+#include <Common/MemoryWorker.h>
 #include <Common/ClickHouseRevision.h>
 #include <Common/DNSResolver.h>
 #include <Common/CgroupsMemoryUsageObserver.h>
@@ -111,6 +111,8 @@
 #include <filesystem>
 #include <unordered_set>
 
+#include <Common/Jemalloc.h>
+
 #include "config.h"
 #include <Common/config_version.h>
 
@@ -449,9 +451,12 @@ void checkForUsersNotInMainConfig(
     }
 }
 
+namespace
+{
+
 /// Unused in other builds
 #if defined(OS_LINUX)
-static String readLine(const String & path)
+String readLine(const String & path)
 {
     ReadBufferFromFile in(path);
     String contents;
@@ -459,7 +464,7 @@ static String readLine(const String & path)
     return contents;
 }
 
-static int readNumber(const String & path)
+int readNumber(const String & path)
 {
     ReadBufferFromFile in(path);
     int result;
@@ -469,7 +474,7 @@ static int readNumber(const String & path)
 
 #endif
 
-static void sanityChecks(Server & server)
+void sanityChecks(Server & server)
 {
     std::string data_path = getCanonicalPath(server.config().getString("path", DBMS_DEFAULT_PATH));
     std::string logs_path = server.config().getString("logger.log", "");
@@ -590,6 +595,8 @@ static void sanityChecks(Server & server)
     }
 }
 
+}
+
 void loadStartupScripts(const Poco::Util::AbstractConfiguration & config, ContextMutablePtr context, Poco::Logger * log)
 {
     try
@@ -906,6 +913,8 @@ try
         LOG_INFO(log, "Background threads finished in {} ms", watch.elapsedMilliseconds());
     });
 
+    MemoryWorker memory_worker(global_context->getServerSettings().memory_worker_period_ms);
+
     /// This object will periodically calculate some metrics.
     ServerAsynchronousMetrics async_metrics(
         global_context,
@@ -924,8 +933,9 @@ try
             for (const auto & server : servers)
                 metrics.emplace_back(ProtocolServerMetrics{server.getPortName(), server.currentThreads(), server.refusedConnections()});
             return metrics;
-        }
-    );
+        },
+        /*update_jemalloc_epoch_=*/memory_worker.getSource() != MemoryWorker::MemoryUsageSource::Jemalloc,
+        /*update_rss_=*/memory_worker.getSource() == MemoryWorker::MemoryUsageSource::None);
 
     /// NOTE: global context should be destroyed *before* GlobalThreadPool::shutdown()
     /// Otherwise GlobalThreadPool::shutdown() will hang, since Context holds some threads.
@@ -1204,6 +1214,8 @@ try
 
     FailPointInjection::enableFromGlobalConfig(config());
 
+    memory_worker.start();
+
     int default_oom_score = 0;
 
 #if !defined(NDEBUG)
@@ -1547,15 +1559,6 @@ try
             total_memory_tracker.setDescription("(total)");
             total_memory_tracker.setMetric(CurrentMetrics::MemoryTracking);
 
-            if (cgroups_memory_usage_observer)
-            {
-                double hard_limit_ratio = new_server_settings.cgroup_memory_watcher_hard_limit_ratio;
-                double soft_limit_ratio = new_server_settings.cgroup_memory_watcher_soft_limit_ratio;
-                cgroups_memory_usage_observer->setMemoryUsageLimits(
-                    static_cast<uint64_t>(max_server_memory_usage * hard_limit_ratio),
-                    static_cast<uint64_t>(max_server_memory_usage * soft_limit_ratio));
-            }
-
             size_t merges_mutations_memory_usage_soft_limit = new_server_settings.merges_mutations_memory_usage_soft_limit;
 
             size_t default_merges_mutations_server_memory_usage = static_cast<size_t>(current_physical_server_memory * new_server_settings.merges_mutations_memory_usage_to_ram_ratio);
@@ -1584,8 +1587,6 @@ try
             background_memory_tracker.setDescription("(background)");
             background_memory_tracker.setMetric(CurrentMetrics::MergesMutationsMemoryTracking);
 
-            total_memory_tracker.setAllowUseJemallocMemory(new_server_settings.allow_use_jemalloc_memory);
-
             auto * global_overcommit_tracker = global_context->getGlobalOvercommitTracker();
             total_memory_tracker.setOvercommitTracker(global_overcommit_tracker);
 

src/Access/Authentication.cpp

@@ -239,15 +239,36 @@ bool Authentication::areCredentialsValid(
                 throw Authentication::Require<GSSAcceptorContext>(auth_data.getKerberosRealm());
 
             case AuthenticationType::SSL_CERTIFICATE:
+            {
                 for (SSLCertificateSubjects::Type type : {SSLCertificateSubjects::Type::CN, SSLCertificateSubjects::Type::SAN})
                 {
                     for (const auto & subject : auth_data.getSSLCertificateSubjects().at(type))
                     {
                         if (ssl_certificate_credentials->getSSLCertificateSubjects().at(type).contains(subject))
                             return true;
+
+                        // Wildcard support (1 only)
+                        if (subject.contains('*'))
+                        {
+                            auto prefix = std::string_view(subject).substr(0, subject.find('*'));
+                            auto suffix = std::string_view(subject).substr(subject.find('*') + 1);
+                            auto slashes = std::count(subject.begin(), subject.end(), '/');
+
+                            for (const auto & certificate_subject : ssl_certificate_credentials->getSSLCertificateSubjects().at(type))
+                            {
+                                bool matches_wildcard = certificate_subject.starts_with(prefix) && certificate_subject.ends_with(suffix);
+
+                                // '*' must not represent a '/' in URI, so check if the number of '/' are equal
+                                bool matches_slashes = slashes == count(certificate_subject.begin(), certificate_subject.end(), '/');
+
+                                if (matches_wildcard && matches_slashes)
+                                    return true;
+                            }
+                        }
                     }
                 }
                 return false;
+            }
 
             case AuthenticationType::SSH_KEY:
 #if USE_SSH

src/AggregateFunctions/AggregateFunctionGroupConcat.cpp

@@ -116,15 +116,17 @@ class GroupConcatImpl final
     SerializationPtr serialization;
     UInt64 limit;
     const String delimiter;
+    const DataTypePtr type;
 
 public:
     GroupConcatImpl(const DataTypePtr & data_type_, const Array & parameters_, UInt64 limit_, const String & delimiter_)
         : IAggregateFunctionDataHelper<GroupConcatData<has_limit>, GroupConcatImpl<has_limit>>(
             {data_type_}, parameters_, std::make_shared<DataTypeString>())
-        , serialization(this->argument_types[0]->getDefaultSerialization())
         , limit(limit_)
         , delimiter(delimiter_)
+        , type(data_type_)
     {
+        serialization = isFixedString(type) ? std::make_shared<DataTypeString>()->getDefaultSerialization() : this->argument_types[0]->getDefaultSerialization();
     }
 
     String getName() const override { return name; }
@@ -140,6 +142,13 @@ public:
         if (cur_data.data_size != 0)
             cur_data.insertChar(delimiter.c_str(), delimiter.size(), arena);
 
+        if (isFixedString(type))
+        {
+            ColumnWithTypeAndName col = {columns[0]->getPtr(), type, "column"};
+            const auto & col_str = castColumn(col, std::make_shared<DataTypeString>());
+            cur_data.insert(col_str.get(), serialization, row_num, arena);
+        }
+        else
             cur_data.insert(columns[0], serialization, row_num, arena);
     }
 

src/CMakeLists.txt

@@ -176,7 +176,7 @@ add_library (clickhouse_new_delete STATIC Common/new_delete.cpp)
 target_link_libraries (clickhouse_new_delete PRIVATE clickhouse_common_io)
 if (TARGET ch_contrib::jemalloc)
     target_link_libraries (clickhouse_new_delete PRIVATE ch_contrib::jemalloc)
-    target_link_libraries (clickhouse_common_io PRIVATE ch_contrib::jemalloc)
+    target_link_libraries (clickhouse_common_io PUBLIC ch_contrib::jemalloc)
     target_link_libraries (clickhouse_storages_system PRIVATE ch_contrib::jemalloc)
 endif()
 

src/Client/ConnectionPoolWithFailover.cpp

@@ -168,7 +168,7 @@ std::vector<ConnectionPoolWithFailover::TryResult> ConnectionPoolWithFailover::g
     { return tryGetEntry(pool, timeouts, fail_message, settings, &table_to_check, /*async_callback=*/ {}); };
 
     return getManyImpl(settings, pool_mode, try_get_entry,
-        /*skip_unavailable_endpoints=*/ std::nullopt,
+        /*skip_unavailable_endpoints=*/ false, /// skip_unavailable_endpoints is used to get the min number of entries, and we need at least one
         /*priority_func=*/ {},
         settings.distributed_insert_skip_read_only_replicas);
 }

src/Client/ConnectionPoolWithFailover.h

@@ -42,7 +42,7 @@ public:
             size_t max_error_cap = DBMS_CONNECTION_POOL_WITH_FAILOVER_MAX_ERROR_COUNT);
 
     using Entry = IConnectionPool::Entry;
-    using PoolWithFailoverBase<IConnectionPool>::checkTryResultIsValid;
+    using PoolWithFailoverBase<IConnectionPool>::getValidTryResult;
 
     /** Allocates connection to work. */
     Entry get(const ConnectionTimeouts & timeouts) override;
@@ -98,7 +98,7 @@ public:
 
     std::vector<Base::ShuffledPool> getShuffledPools(const Settings & settings, GetPriorityFunc priority_func = {}, bool use_slowdown_count = false);
 
-    size_t getMaxErrorCup() const { return Base::max_error_cap; }
+    size_t getMaxErrorCap() const { return Base::max_error_cap; }
 
     void updateSharedError(std::vector<ShuffledPool> & shuffled_pools)
     {

src/Client/HedgedConnectionsFactory.cpp

@@ -327,7 +327,7 @@ HedgedConnectionsFactory::State HedgedConnectionsFactory::processFinishedConnect
         ShuffledPool & shuffled_pool = shuffled_pools[index];
         LOG_INFO(log, "Connection failed at try №{}, reason: {}", (shuffled_pool.error_count + 1), fail_message);
 
-        shuffled_pool.error_count = std::min(pool->getMaxErrorCup(), shuffled_pool.error_count + 1);
+        shuffled_pool.error_count = std::min(pool->getMaxErrorCap(), shuffled_pool.error_count + 1);
         shuffled_pool.slowdown_count = 0;
 
         if (shuffled_pool.error_count >= max_tries)

src/Common/AsynchronousMetrics.cpp

@@ -1,5 +1,3 @@
-#include <Common/AsynchronousMetrics.h>
-
 #include <IO/MMappedFileCache.h>
 #include <IO/ReadHelpers.h>
 #include <IO/UncompressedCache.h>
@@ -8,8 +6,10 @@
 #include <base/find_symbols.h>
 #include <base/getPageSize.h>
 #include <sys/resource.h>
+#include <Common/AsynchronousMetrics.h>
 #include <Common/CurrentMetrics.h>
 #include <Common/Exception.h>
+#include <Common/Jemalloc.h>
 #include <Common/filesystemHelpers.h>
 #include <Common/formatReadable.h>
 #include <Common/logger_useful.h>
@@ -69,10 +69,14 @@ static void openCgroupv2MetricFile(const std::string & filename, std::optional<R
 
 AsynchronousMetrics::AsynchronousMetrics(
     unsigned update_period_seconds,
-    const ProtocolServerMetricsFunc & protocol_server_metrics_func_)
+    const ProtocolServerMetricsFunc & protocol_server_metrics_func_,
+    bool update_jemalloc_epoch_,
+    bool update_rss_)
     : update_period(update_period_seconds)
     , log(getLogger("AsynchronousMetrics"))
     , protocol_server_metrics_func(protocol_server_metrics_func_)
+    , update_jemalloc_epoch(update_jemalloc_epoch_)
+    , update_rss(update_rss_)
 {
 #if defined(OS_LINUX)
     openFileIfExists("/proc/cpuinfo", cpuinfo);
@@ -411,9 +415,7 @@ Value saveJemallocMetricImpl(
     const std::string & jemalloc_full_name,
     const std::string & clickhouse_full_name)
 {
-    Value value{};
-    size_t size = sizeof(value);
-    mallctl(jemalloc_full_name.c_str(), &value, &size, nullptr, 0);
+    auto value = getJemallocValue<Value>(jemalloc_full_name.c_str());
     values[clickhouse_full_name] = AsynchronousMetricValue(value, "An internal metric of the low-level memory allocator (jemalloc). See https://jemalloc.net/jemalloc.3.html");
     return value;
 }
@@ -768,8 +770,11 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
     // 'epoch' is a special mallctl -- it updates the statistics. Without it, all
     // the following calls will return stale values. It increments and returns
     // the current epoch number, which might be useful to log as a sanity check.
-    auto epoch = updateJemallocEpoch();
-    new_values["jemalloc.epoch"] = { epoch, "An internal incremental update number of the statistics of jemalloc (Jason Evans' memory allocator), used in all other `jemalloc` metrics." };
+    auto epoch = update_jemalloc_epoch ? updateJemallocEpoch() : getJemallocValue<uint64_t>("epoch");
+    new_values["jemalloc.epoch"]
+        = {epoch,
+           "An internal incremental update number of the statistics of jemalloc (Jason Evans' memory allocator), used in all other "
+           "`jemalloc` metrics."};
 
     // Collect the statistics themselves.
     saveJemallocMetric<size_t>(new_values, "allocated");
@@ -782,10 +787,10 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
     saveJemallocMetric<size_t>(new_values, "background_thread.num_threads");
     saveJemallocMetric<uint64_t>(new_values, "background_thread.num_runs");
     saveJemallocMetric<uint64_t>(new_values, "background_thread.run_intervals");
-    saveJemallocProf<size_t>(new_values, "active");
+    saveJemallocProf<bool>(new_values, "active");
     saveAllArenasMetric<size_t>(new_values, "pactive");
-    [[maybe_unused]] size_t je_malloc_pdirty = saveAllArenasMetric<size_t>(new_values, "pdirty");
-    [[maybe_unused]] size_t je_malloc_pmuzzy = saveAllArenasMetric<size_t>(new_values, "pmuzzy");
+    saveAllArenasMetric<size_t>(new_values, "pdirty");
+    saveAllArenasMetric<size_t>(new_values, "pmuzzy");
     saveAllArenasMetric<size_t>(new_values, "dirty_purged");
     saveAllArenasMetric<size_t>(new_values, "muzzy_purged");
 #endif
@@ -814,41 +819,8 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
             " It is unspecified whether it includes the per-thread stacks and most of the allocated memory, that is allocated with the 'mmap' system call."
             " This metric exists only for completeness reasons. I recommend to use the `MemoryResident` metric for monitoring."};
 
-        /// We must update the value of total_memory_tracker periodically.
-        /// Otherwise it might be calculated incorrectly - it can include a "drift" of memory amount.
-        /// See https://github.com/ClickHouse/ClickHouse/issues/10293
-        {
-            Int64 amount = total_memory_tracker.get();
-            Int64 peak = total_memory_tracker.getPeak();
-            Int64 rss = data.resident;
-            Int64 free_memory_in_allocator_arenas = 0;
-
-#if USE_JEMALLOC
-            /// According to jemalloc man, pdirty is:
-            ///
-            ///     Number of pages within unused extents that are potentially
-            ///     dirty, and for which madvise() or similar has not been called.
-            ///
-            /// So they will be subtracted from RSS to make accounting more
-            /// accurate, since those pages are not really RSS but a memory
-            /// that can be used at anytime via jemalloc.
-            free_memory_in_allocator_arenas = je_malloc_pdirty * getPageSize();
-#endif
-
-            Int64 difference = rss - amount;
-
-            /// Log only if difference is high. This is for convenience. The threshold is arbitrary.
-            if (difference >= 1048576 || difference <= -1048576)
-                LOG_TRACE(log,
-                    "MemoryTracking: was {}, peak {}, free memory in arenas {}, will set to {} (RSS), difference: {}",
-                    ReadableSize(amount),
-                    ReadableSize(peak),
-                    ReadableSize(free_memory_in_allocator_arenas),
-                    ReadableSize(rss),
-                    ReadableSize(difference));
-
-            MemoryTracker::setRSS(rss, free_memory_in_allocator_arenas);
-        }
+        if (update_rss)
+            MemoryTracker::updateRSS(data.resident);
     }
 
     {

src/Common/AsynchronousMetrics.h

@@ -1,15 +1,14 @@
 #pragma once
 
+#include <Common/CgroupsMemoryUsageObserver.h>
 #include <Common/MemoryStatisticsOS.h>
 #include <Common/ThreadPool.h>
 #include <Common/Stopwatch.h>
 #include <IO/ReadBufferFromFile.h>
 
 #include <condition_variable>
-#include <map>
 #include <mutex>
 #include <string>
-#include <thread>
 #include <vector>
 #include <optional>
 #include <unordered_map>
@@ -69,7 +68,9 @@ public:
 
     AsynchronousMetrics(
         unsigned update_period_seconds,
-        const ProtocolServerMetricsFunc & protocol_server_metrics_func_);
+        const ProtocolServerMetricsFunc & protocol_server_metrics_func_,
+        bool update_jemalloc_epoch_,
+        bool update_rss_);
 
     virtual ~AsynchronousMetrics();
 
@@ -112,6 +113,9 @@ private:
     MemoryStatisticsOS memory_stat TSA_GUARDED_BY(data_mutex);
 #endif
 
+    [[maybe_unused]] const bool update_jemalloc_epoch;
+    [[maybe_unused]] const bool update_rss;
+
 #if defined(OS_LINUX)
     std::optional<ReadBufferFromFilePRead> meminfo TSA_GUARDED_BY(data_mutex);
     std::optional<ReadBufferFromFilePRead> loadavg TSA_GUARDED_BY(data_mutex);

src/Common/CgroupsMemoryUsageObserver.cpp

@@ -14,239 +14,21 @@
 #include <fmt/ranges.h>
 
 #include <cstdint>
-#include <filesystem>
-#include <memory>
-#include <optional>
-
-#include "config.h"
-#if USE_JEMALLOC
-#    include <jemalloc/jemalloc.h>
-#define STRINGIFY_HELPER(x) #x
-#define STRINGIFY(x) STRINGIFY_HELPER(x)
-#endif
 
 using namespace DB;
-namespace fs = std::filesystem;
-
-namespace DB
-{
-
-namespace ErrorCodes
-{
-extern const int FILE_DOESNT_EXIST;
-extern const int INCORRECT_DATA;
-}
-
-}
-
-namespace
-{
-
-/// Format is
-///   kernel 5
-///   rss 15
-///   [...]
-using Metrics = std::map<std::string, uint64_t>;
-
-Metrics readAllMetricsFromStatFile(ReadBufferFromFile & buf)
-{
-    Metrics metrics;
-    while (!buf.eof())
-    {
-        std::string current_key;
-        readStringUntilWhitespace(current_key, buf);
-
-        assertChar(' ', buf);
-
-        uint64_t value = 0;
-        readIntText(value, buf);
-        assertChar('\n', buf);
-
-        auto [_, inserted] = metrics.emplace(std::move(current_key), value);
-        chassert(inserted, "Duplicate keys in stat file");
-    }
-    return metrics;
-}
-
-uint64_t readMetricFromStatFile(ReadBufferFromFile & buf, const std::string & key)
-{
-    const auto all_metrics = readAllMetricsFromStatFile(buf);
-    if (const auto it = all_metrics.find(key); it != all_metrics.end())
-        return it->second;
-    throw Exception(ErrorCodes::INCORRECT_DATA, "Cannot find '{}' in '{}'", key, buf.getFileName());
-}
-
-struct CgroupsV1Reader : ICgroupsReader
-{
-    explicit CgroupsV1Reader(const fs::path & stat_file_dir) : buf(stat_file_dir / "memory.stat") { }
-
-    uint64_t readMemoryUsage() override
-    {
-        std::lock_guard lock(mutex);
-        buf.rewind();
-        return readMetricFromStatFile(buf, "rss");
-    }
-
-    std::string dumpAllStats() override
-    {
-        std::lock_guard lock(mutex);
-        buf.rewind();
-        return fmt::format("{}", readAllMetricsFromStatFile(buf));
-    }
-
-private:
-    std::mutex mutex;
-    ReadBufferFromFile buf TSA_GUARDED_BY(mutex);
-};
-
-struct CgroupsV2Reader : ICgroupsReader
-{
-    explicit CgroupsV2Reader(const fs::path & stat_file_dir)
-        : current_buf(stat_file_dir / "memory.current"), stat_buf(stat_file_dir / "memory.stat")
-    {
-    }
-
-    uint64_t readMemoryUsage() override
-    {
-        std::lock_guard lock(mutex);
-        current_buf.rewind();
-        stat_buf.rewind();
-
-        int64_t mem_usage = 0;
-        /// memory.current contains a single number
-        /// the reason why we subtract it described here: https://github.com/ClickHouse/ClickHouse/issues/64652#issuecomment-2149630667
-        readIntText(mem_usage, current_buf);
-        mem_usage -= readMetricFromStatFile(stat_buf, "inactive_file");
-        chassert(mem_usage >= 0, "Negative memory usage");
-        return mem_usage;
-    }
-
-    std::string dumpAllStats() override
-    {
-        std::lock_guard lock(mutex);
-        stat_buf.rewind();
-        return fmt::format("{}", readAllMetricsFromStatFile(stat_buf));
-    }
-
-private:
-    std::mutex mutex;
-    ReadBufferFromFile current_buf TSA_GUARDED_BY(mutex);
-    ReadBufferFromFile stat_buf TSA_GUARDED_BY(mutex);
-};
-
-/// Caveats:
-/// - All of the logic in this file assumes that the current process is the only process in the
-///   containing cgroup (or more precisely: the only process with significant memory consumption).
-///   If this is not the case, then other processe's memory consumption may affect the internal
-///   memory tracker ...
-/// - Cgroups v1 and v2 allow nested cgroup hierarchies. As v1 is deprecated for over half a
-///   decade and will go away at some point, hierarchical detection is only implemented for v2.
-/// - I did not test what happens if a host has v1 and v2 simultaneously enabled. I believe such
-///   systems existed only for a short transition period.
-
-std::optional<std::string> getCgroupsV1Path()
-{
-    auto path = default_cgroups_mount / "memory/memory.stat";
-    if (!fs::exists(path))
-        return {};
-    return {default_cgroups_mount / "memory"};
-}
-
-std::pair<std::string, CgroupsMemoryUsageObserver::CgroupsVersion> getCgroupsPath()
-{
-    auto v2_path = getCgroupsV2PathContainingFile("memory.current");
-    if (v2_path.has_value())
-        return {*v2_path, CgroupsMemoryUsageObserver::CgroupsVersion::V2};
-
-    auto v1_path = getCgroupsV1Path();
-    if (v1_path.has_value())
-        return {*v1_path, CgroupsMemoryUsageObserver::CgroupsVersion::V1};
-
-    throw Exception(ErrorCodes::FILE_DOESNT_EXIST, "Cannot find cgroups v1 or v2 current memory file");
-}
-
-}
 
 namespace DB
 {
 
 CgroupsMemoryUsageObserver::CgroupsMemoryUsageObserver(std::chrono::seconds wait_time_)
     : log(getLogger("CgroupsMemoryUsageObserver")), wait_time(wait_time_)
-{
-    const auto [cgroup_path, version] = getCgroupsPath();
-
-    cgroup_reader = createCgroupsReader(version, cgroup_path);
-
-    LOG_INFO(
-        log,
-        "Will read the current memory usage from '{}' (cgroups version: {}), wait time is {} sec",
-        cgroup_path,
-        (version == CgroupsVersion::V1) ? "v1" : "v2",
-        wait_time.count());
-}
+{}
 
 CgroupsMemoryUsageObserver::~CgroupsMemoryUsageObserver()
 {
     stopThread();
 }
 
-void CgroupsMemoryUsageObserver::setMemoryUsageLimits(uint64_t hard_limit_, uint64_t soft_limit_)
-{
-    std::lock_guard<std::mutex> limit_lock(limit_mutex);
-
-    if (hard_limit_ == hard_limit && soft_limit_ == soft_limit)
-        return;
-
-    hard_limit = hard_limit_;
-    soft_limit = soft_limit_;
-
-    on_hard_limit = [this, hard_limit_](bool up)
-    {
-        if (up)
-        {
-            LOG_WARNING(log, "Exceeded hard memory limit ({})", ReadableSize(hard_limit_));
-
-            /// Update current usage in memory tracker. Also reset free_memory_in_allocator_arenas to zero though we don't know if they are
-            /// really zero. Trying to avoid OOM ...
-            MemoryTracker::setRSS(hard_limit_, 0);
-        }
-        else
-        {
-            LOG_INFO(log, "Dropped below hard memory limit ({})", ReadableSize(hard_limit_));
-        }
-    };
-
-    on_soft_limit = [this, soft_limit_](bool up)
-    {
-        if (up)
-        {
-            LOG_WARNING(log, "Exceeded soft memory limit ({})", ReadableSize(soft_limit_));
-
-#    if USE_JEMALLOC
-            LOG_INFO(log, "Purging jemalloc arenas");
-            mallctl("arena." STRINGIFY(MALLCTL_ARENAS_ALL) ".purge", nullptr, nullptr, nullptr, 0);
-#    endif
-            /// Reset current usage in memory tracker. Expect zero for free_memory_in_allocator_arenas as we just purged them.
-            uint64_t memory_usage = cgroup_reader->readMemoryUsage();
-            LOG_TRACE(
-                log,
-                "Read current memory usage {} bytes ({}) from cgroups, full available stats: {}",
-                memory_usage,
-                ReadableSize(memory_usage),
-                cgroup_reader->dumpAllStats());
-            MemoryTracker::setRSS(memory_usage, 0);
-
-            LOG_INFO(log, "Purged jemalloc arenas. Current memory usage is {}", ReadableSize(memory_usage));
-        }
-        else
-        {
-            LOG_INFO(log, "Dropped below soft memory limit ({})", ReadableSize(soft_limit_));
-        }
-    };
-
-    LOG_INFO(log, "Set new limits, soft limit: {}, hard limit: {}", ReadableSize(soft_limit_), ReadableSize(hard_limit_));
-}
-
 void CgroupsMemoryUsageObserver::setOnMemoryAmountAvailableChangedFn(OnMemoryAmountAvailableChangedFn on_memory_amount_available_changed_)
 {
     std::lock_guard<std::mutex> memory_amount_available_changed_lock(memory_amount_available_changed_mutex);
@@ -300,35 +82,6 @@ void CgroupsMemoryUsageObserver::runThread()
                 std::lock_guard<std::mutex> memory_amount_available_changed_lock(memory_amount_available_changed_mutex);
                 on_memory_amount_available_changed();
             }
-
-            std::lock_guard<std::mutex> limit_lock(limit_mutex);
-            if (soft_limit > 0 && hard_limit > 0)
-            {
-                uint64_t memory_usage = cgroup_reader->readMemoryUsage();
-                LOG_TRACE(log, "Read current memory usage {} bytes ({}) from cgroups", memory_usage, ReadableSize(memory_usage));
-                if (memory_usage > hard_limit)
-                {
-                    if (last_memory_usage <= hard_limit)
-                        on_hard_limit(true);
-                }
-                else
-                {
-                    if (last_memory_usage > hard_limit)
-                        on_hard_limit(false);
-                }
-
-                if (memory_usage > soft_limit)
-                {
-                    if (last_memory_usage <= soft_limit)
-                        on_soft_limit(true);
-                }
-                else
-                {
-                    if (last_memory_usage > soft_limit)
-                        on_soft_limit(false);
-                }
-                last_memory_usage = memory_usage;
-            }
         }
         catch (...)
         {
@@ -337,13 +90,6 @@ void CgroupsMemoryUsageObserver::runThread()
     }
 }
 
-std::unique_ptr<ICgroupsReader> createCgroupsReader(CgroupsMemoryUsageObserver::CgroupsVersion version, const fs::path & cgroup_path)
-{
-    if (version == CgroupsMemoryUsageObserver::CgroupsVersion::V2)
-        return std::make_unique<CgroupsV2Reader>(cgroup_path);
-    else
-        return std::make_unique<CgroupsV1Reader>(cgroup_path);
-}
 }
 
 #endif

src/Common/CgroupsMemoryUsageObserver.h

@@ -3,30 +3,12 @@
 #include <Common/ThreadPool.h>
 
 #include <chrono>
-#include <memory>
 #include <mutex>
 
 namespace DB
 {
 
-struct ICgroupsReader
-{
-    virtual ~ICgroupsReader() = default;
-
-    virtual uint64_t readMemoryUsage() = 0;
-
-    virtual std::string dumpAllStats() = 0;
-};
-
-/// Does two things:
-/// 1. Periodically reads the memory usage of the process from Linux cgroups.
-///    You can specify soft or hard memory limits:
-///    - When the soft memory limit is hit, drop jemalloc cache.
-///    - When the hard memory limit is hit, update MemoryTracking metric to throw memory exceptions faster.
-///    The goal of this is to avoid that the process hits the maximum allowed memory limit at which there is a good
-///    chance that the Limux OOM killer terminates it. All of this is done is because internal memory tracking in
-///    ClickHouse can unfortunately under-estimate the actually used memory.
-/// 2. Periodically reads the the maximum memory available to the process (which can change due to cgroups settings).
+///  Periodically reads the the maximum memory available to the process (which can change due to cgroups settings).
 ///  You can specify a callback to react on changes. The callback typically reloads the configuration, i.e. Server
 ///  or Keeper configuration file. This reloads settings 'max_server_memory_usage' (Server) and 'max_memory_usage_soft_limit'
 ///  (Keeper) from which various other internal limits are calculated, including the soft and hard limits for (1.).
@@ -37,19 +19,11 @@ struct ICgroupsReader
 class CgroupsMemoryUsageObserver
 {
 public:
-    using OnMemoryLimitFn = std::function<void(bool)>;
     using OnMemoryAmountAvailableChangedFn = std::function<void()>;
 
-    enum class CgroupsVersion : uint8_t
-    {
-        V1,
-        V2
-    };
-
     explicit CgroupsMemoryUsageObserver(std::chrono::seconds wait_time_);
     ~CgroupsMemoryUsageObserver();
 
-    void setMemoryUsageLimits(uint64_t hard_limit_, uint64_t soft_limit_);
     void setOnMemoryAmountAvailableChangedFn(OnMemoryAmountAvailableChangedFn on_memory_amount_available_changed_);
 
     void startThread();
@@ -60,32 +34,22 @@ private:
     const std::chrono::seconds wait_time;
 
     std::mutex limit_mutex;
-    size_t hard_limit TSA_GUARDED_BY(limit_mutex) = 0;
-    size_t soft_limit TSA_GUARDED_BY(limit_mutex) = 0;
-    OnMemoryLimitFn on_hard_limit TSA_GUARDED_BY(limit_mutex);
-    OnMemoryLimitFn on_soft_limit TSA_GUARDED_BY(limit_mutex);
 
     std::mutex memory_amount_available_changed_mutex;
     OnMemoryAmountAvailableChangedFn on_memory_amount_available_changed TSA_GUARDED_BY(memory_amount_available_changed_mutex);
 
-    uint64_t last_memory_usage = 0;        /// how much memory does the process use
     uint64_t last_available_memory_amount; /// how much memory can the process use
 
     void stopThread();
 
     void runThread();
 
-    std::unique_ptr<ICgroupsReader> cgroup_reader;
-
     std::mutex thread_mutex;
     std::condition_variable cond;
     ThreadFromGlobalPool thread;
     bool quit = false;
 };
 
-std::unique_ptr<ICgroupsReader>
-createCgroupsReader(CgroupsMemoryUsageObserver::CgroupsVersion version, const std::filesystem::path & cgroup_path);
-
 #else
 class CgroupsMemoryUsageObserver
 {

src/Common/ErrorCodes.cpp

@@ -609,6 +609,7 @@
     M(728, UNEXPECTED_DATA_TYPE) \
     M(729, ILLEGAL_TIME_SERIES_TAGS) \
     M(730, REFRESH_FAILED) \
+    M(731, QUERY_CACHE_USED_WITH_NON_THROW_OVERFLOW_MODE) \
     \
     M(900, DISTRIBUTED_CACHE_ERROR) \
     M(901, CANNOT_USE_DISTRIBUTED_CACHE) \

src/Common/Jemalloc.cpp

@@ -5,7 +5,6 @@
 #include <Common/Exception.h>
 #include <Common/Stopwatch.h>
 #include <Common/logger_useful.h>
-#include <jemalloc/jemalloc.h>
 
 #define STRINGIFY_HELPER(x) #x
 #define STRINGIFY(x) STRINGIFY_HELPER(x)
@@ -26,7 +25,6 @@ namespace ErrorCodes
 
 void purgeJemallocArenas()
 {
-    LOG_TRACE(getLogger("SystemJemalloc"), "Purging unused memory");
     Stopwatch watch;
     mallctl("arena." STRINGIFY(MALLCTL_ARENAS_ALL) ".purge", nullptr, nullptr, nullptr, 0);
     ProfileEvents::increment(ProfileEvents::MemoryAllocatorPurge);
@@ -46,20 +44,6 @@ void checkJemallocProfilingEnabled()
             "set: MALLOC_CONF=background_thread:true,prof:true");
 }
 
-template <typename T>
-void setJemallocValue(const char * name, T value)
-{
-    T old_value;
-    size_t old_value_size = sizeof(T);
-    if (mallctl(name, &old_value, &old_value_size, reinterpret_cast<void*>(&value), sizeof(T)))
-    {
-        LOG_WARNING(getLogger("Jemalloc"), "mallctl for {} failed", name);
-        return;
-    }
-
-    LOG_INFO(getLogger("Jemalloc"), "Value for {} set to {} (from {})", name, value, old_value);
-}
-
 void setJemallocProfileActive(bool value)
 {
     checkJemallocProfilingEnabled();

src/Common/Jemalloc.h

@@ -5,6 +5,8 @@
 #if USE_JEMALLOC
 
 #include <string>
+#include <Common/logger_useful.h>
+#include <jemalloc/jemalloc.h>
 
 namespace DB
 {
@@ -21,6 +23,59 @@ void setJemallocBackgroundThreads(bool enabled);
 
 void setJemallocMaxBackgroundThreads(size_t max_threads);
 
+template <typename T>
+void setJemallocValue(const char * name, T value)
+{
+    T old_value;
+    size_t old_value_size = sizeof(T);
+    mallctl(name, &old_value, &old_value_size, reinterpret_cast<void*>(&value), sizeof(T));
+    LOG_INFO(getLogger("Jemalloc"), "Value for {} set to {} (from {})", name, value, old_value);
+}
+
+template <typename T>
+T getJemallocValue(const char * name)
+{
+    T value;
+    size_t value_size = sizeof(T);
+    mallctl(name, &value, &value_size, nullptr, 0);
+    return value;
+}
+
+/// Each mallctl call consists of string name lookup which can be expensive.
+/// This can be avoided by translating name to "Management Information Base" (MIB)
+/// and using it in mallctlbymib calls
+template <typename T>
+struct JemallocMibCache
+{
+    explicit JemallocMibCache(const char * name)
+    {
+        mallctlnametomib(name, mib, &mib_length);
+    }
+
+    void setValue(T value)
+    {
+        mallctlbymib(mib, mib_length, nullptr, nullptr, reinterpret_cast<void*>(&value), sizeof(T));
+    }
+
+    T getValue()
+    {
+        T value;
+        size_t value_size = sizeof(T);
+        mallctlbymib(mib, mib_length, &value, &value_size, nullptr, 0);
+        return value;
+    }
+
+    void run()
+    {
+        mallctlbymib(mib, mib_length, nullptr, nullptr, nullptr, 0);
+    }
+
+private:
+    static constexpr size_t max_mib_length = 4;
+    size_t mib[max_mib_length];
+    size_t mib_length = max_mib_length;
+};
+
 }
 
 #endif

src/Common/MemoryTracker.cpp

@@ -20,13 +20,9 @@
 #if USE_JEMALLOC
 #    include <jemalloc/jemalloc.h>
 
-#define STRINGIFY_HELPER(x) #x
-#define STRINGIFY(x) STRINGIFY_HELPER(x)
-
 #endif
 
 #include <atomic>
-#include <cmath>
 #include <random>
 #include <cstdlib>
 #include <string>
@@ -115,8 +111,6 @@ void AllocationTrace::onFreeImpl(void * ptr, size_t size) const
 namespace ProfileEvents
 {
     extern const Event QueryMemoryLimitExceeded;
-    extern const Event MemoryAllocatorPurge;
-    extern const Event MemoryAllocatorPurgeTimeMicroseconds;
 }
 
 using namespace std::chrono_literals;
@@ -126,15 +120,13 @@ static constexpr size_t log_peak_memory_usage_every = 1ULL << 30;
 MemoryTracker total_memory_tracker(nullptr, VariableContext::Global);
 MemoryTracker background_memory_tracker(&total_memory_tracker, VariableContext::User, false);
 
-std::atomic<Int64> MemoryTracker::free_memory_in_allocator_arenas;
-
 MemoryTracker::MemoryTracker(VariableContext level_) : parent(&total_memory_tracker), level(level_) {}
 MemoryTracker::MemoryTracker(MemoryTracker * parent_, VariableContext level_) : parent(parent_), level(level_) {}
+
 MemoryTracker::MemoryTracker(MemoryTracker * parent_, VariableContext level_, bool log_peak_memory_usage_in_destructor_)
-    : parent(parent_)
-    , log_peak_memory_usage_in_destructor(log_peak_memory_usage_in_destructor_)
-    , level(level_)
-{}
+    : parent(parent_), log_peak_memory_usage_in_destructor(log_peak_memory_usage_in_destructor_), level(level_)
+{
+}
 
 MemoryTracker::~MemoryTracker()
 {
@@ -204,10 +196,14 @@ void MemoryTracker::debugLogBigAllocationWithoutCheck(Int64 size [[maybe_unused]
         return;
 
     MemoryTrackerBlockerInThread blocker(VariableContext::Global);
-    LOG_TEST(getLogger("MemoryTracker"), "Too big allocation ({} bytes) without checking memory limits, "
-                                                   "it may lead to OOM. Stack trace: {}", size, StackTrace().toString());
+    LOG_TEST(
+        getLogger("MemoryTracker"),
+        "Too big allocation ({} bytes) without checking memory limits, "
+        "it may lead to OOM. Stack trace: {}",
+        size,
+        StackTrace().toString());
 #else
-    return;     /// Avoid trash logging in release builds
+    /// Avoid trash logging in release builds
 #endif
 }
 
@@ -228,6 +224,7 @@ AllocationTrace MemoryTracker::allocImpl(Int64 size, bool throw_if_memory_exceed
         {
             /// For global memory tracker always update memory usage.
             amount.fetch_add(size, std::memory_order_relaxed);
+            rss.fetch_add(size, std::memory_order_relaxed);
 
             auto metric_loaded = metric.load(std::memory_order_relaxed);
             if (metric_loaded != CurrentMetrics::end())
@@ -249,6 +246,7 @@ AllocationTrace MemoryTracker::allocImpl(Int64 size, bool throw_if_memory_exceed
       * So, we allow over-allocations.
       */
     Int64 will_be = size ? size + amount.fetch_add(size, std::memory_order_relaxed) : amount.load(std::memory_order_relaxed);
+    Int64 will_be_rss = size ? size + rss.fetch_add(size, std::memory_order_relaxed) : rss.load(std::memory_order_relaxed);
 
     auto metric_loaded = metric.load(std::memory_order_relaxed);
     if (metric_loaded != CurrentMetrics::end() && size)
@@ -275,6 +273,7 @@ AllocationTrace MemoryTracker::allocImpl(Int64 size, bool throw_if_memory_exceed
         {
             /// Revert
             amount.fetch_sub(size, std::memory_order_relaxed);
+            rss.fetch_sub(size, std::memory_order_relaxed);
 
             /// Prevent recursion. Exception::ctor -> std::string -> new[] -> MemoryTracker::alloc
             MemoryTrackerBlockerInThread untrack_lock(VariableContext::Global);
@@ -297,33 +296,8 @@ AllocationTrace MemoryTracker::allocImpl(Int64 size, bool throw_if_memory_exceed
         }
     }
 
-    Int64 limit_to_check = current_hard_limit;
-
-#if USE_JEMALLOC
-    if (level == VariableContext::Global && allow_use_jemalloc_memory.load(std::memory_order_relaxed))
-    {
-        /// Jemalloc arenas may keep some extra memory.
-        /// This memory was substucted from RSS to decrease memory drift.
-        /// In case memory is close to limit, try to pugre the arenas.
-        /// This is needed to avoid OOM, because some allocations are directly done with mmap.
-        Int64 current_free_memory_in_allocator_arenas = free_memory_in_allocator_arenas.load(std::memory_order_relaxed);
-
-        if (current_free_memory_in_allocator_arenas > 0 && current_hard_limit && current_free_memory_in_allocator_arenas + will_be > current_hard_limit)
-        {
-            if (free_memory_in_allocator_arenas.exchange(-current_free_memory_in_allocator_arenas) > 0)
-            {
-                Stopwatch watch;
-                mallctl("arena." STRINGIFY(MALLCTL_ARENAS_ALL) ".purge", nullptr, nullptr, nullptr, 0);
-                ProfileEvents::increment(ProfileEvents::MemoryAllocatorPurge);
-                ProfileEvents::increment(ProfileEvents::MemoryAllocatorPurgeTimeMicroseconds, watch.elapsedMicroseconds());
-            }
-        }
-
-        limit_to_check += abs(current_free_memory_in_allocator_arenas);
-    }
-#endif
-
-    if (unlikely(current_hard_limit && will_be > limit_to_check))
+    if (unlikely(
+            current_hard_limit && (will_be > current_hard_limit || (level == VariableContext::Global && will_be_rss > current_hard_limit))))
     {
         if (memoryTrackerCanThrow(level, false) && throw_if_memory_exceeded)
         {
@@ -335,6 +309,7 @@ AllocationTrace MemoryTracker::allocImpl(Int64 size, bool throw_if_memory_exceed
             {
                 /// Revert
                 amount.fetch_sub(size, std::memory_order_relaxed);
+                rss.fetch_sub(size, std::memory_order_relaxed);
 
                 /// Prevent recursion. Exception::ctor -> std::string -> new[] -> MemoryTracker::alloc
                 MemoryTrackerBlockerInThread untrack_lock(VariableContext::Global);
@@ -343,12 +318,13 @@ AllocationTrace MemoryTracker::allocImpl(Int64 size, bool throw_if_memory_exceed
                 throw DB::Exception(
                                     DB::ErrorCodes::MEMORY_LIMIT_EXCEEDED,
                                     "Memory limit{}{} exceeded: "
-                                    "would use {} (attempt to allocate chunk of {} bytes), maximum: {}."
+                                    "would use {} (attempt to allocate chunk of {} bytes), current RSS {}, maximum: {}."
                                     "{}{}",
                                     description ? " " : "",
                                     description ? description : "",
                                     formatReadableSizeWithBinarySuffix(will_be),
                                     size,
+                                    formatReadableSizeWithBinarySuffix(rss.load(std::memory_order_relaxed)),
                                     formatReadableSizeWithBinarySuffix(current_hard_limit),
                                     overcommit_result == OvercommitResult::NONE ? "" : " OvercommitTracker decision: ",
                                     toDescription(overcommit_result));
@@ -442,6 +418,7 @@ AllocationTrace MemoryTracker::free(Int64 size, double _sample_probability)
         {
             /// For global memory tracker always update memory usage.
             amount.fetch_sub(size, std::memory_order_relaxed);
+            rss.fetch_sub(size, std::memory_order_relaxed);
             auto metric_loaded = metric.load(std::memory_order_relaxed);
             if (metric_loaded != CurrentMetrics::end())
                 CurrentMetrics::sub(metric_loaded, size);
@@ -455,7 +432,12 @@ AllocationTrace MemoryTracker::free(Int64 size, double _sample_probability)
     }
 
     Int64 accounted_size = size;
-    if (level == VariableContext::Thread || level == VariableContext::Global)
+    if (level == VariableContext::Global)
+    {
+        amount.fetch_sub(accounted_size, std::memory_order_relaxed);
+        rss.fetch_sub(accounted_size, std::memory_order_relaxed);
+    }
+    else if (level == VariableContext::Thread)
     {
         /// Could become negative if memory allocated in this thread is freed in another one
         amount.fetch_sub(accounted_size, std::memory_order_relaxed);
@@ -529,21 +511,29 @@ void MemoryTracker::reset()
 }
 
 
-void MemoryTracker::setRSS(Int64 rss_, Int64 free_memory_in_allocator_arenas_)
+void MemoryTracker::updateRSS(Int64 rss_)
 {
-    Int64 new_amount = rss_;
+    total_memory_tracker.rss.store(rss_, std::memory_order_relaxed);
+}
+
+void MemoryTracker::updateAllocated(Int64 allocated_)
+{
+    Int64 new_amount = allocated_;
+    LOG_INFO(
+        getLogger("MemoryTracker"),
+        "Correcting the value of global memory tracker from {} to {}",
+        ReadableSize(total_memory_tracker.amount.load(std::memory_order_relaxed)),
+        ReadableSize(allocated_));
     total_memory_tracker.amount.store(new_amount, std::memory_order_relaxed);
-    free_memory_in_allocator_arenas.store(free_memory_in_allocator_arenas_, std::memory_order_relaxed);
 
     auto metric_loaded = total_memory_tracker.metric.load(std::memory_order_relaxed);
     if (metric_loaded != CurrentMetrics::end())
         CurrentMetrics::set(metric_loaded, new_amount);
 
     bool log_memory_usage = true;
-    total_memory_tracker.updatePeak(rss_, log_memory_usage);
+    total_memory_tracker.updatePeak(new_amount, log_memory_usage);
 }
 
-
 void MemoryTracker::setSoftLimit(Int64 value)
 {
     soft_limit.store(value, std::memory_order_relaxed);

src/Common/MemoryTracker.h

@@ -2,7 +2,6 @@
 
 #include <atomic>
 #include <chrono>
-#include <optional>
 #include <base/types.h>
 #include <Common/CurrentMetrics.h>
 #include <Common/VariableContext.h>
@@ -57,9 +56,8 @@ private:
     std::atomic<Int64> soft_limit {0};
     std::atomic<Int64> hard_limit {0};
     std::atomic<Int64> profiler_limit {0};
-    std::atomic_bool allow_use_jemalloc_memory {true};
 
-    static std::atomic<Int64> free_memory_in_allocator_arenas;
+    std::atomic<Int64> rss{0};
 
     Int64 profiler_step = 0;
 
@@ -122,6 +120,11 @@ public:
         return amount.load(std::memory_order_relaxed);
     }
 
+    Int64 getRSS() const
+    {
+        return rss.load(std::memory_order_relaxed);
+    }
+
     // Merges and mutations may pass memory ownership to other threads thus in the end of execution
     // MemoryTracker for background task may have a non-zero counter.
     // This method is intended to fix the counter inside of background_memory_tracker.
@@ -154,14 +157,6 @@ public:
     {
         return soft_limit.load(std::memory_order_relaxed);
     }
-    void setAllowUseJemallocMemory(bool value)
-    {
-        allow_use_jemalloc_memory.store(value, std::memory_order_relaxed);
-    }
-    bool getAllowUseJemallocMmemory() const
-    {
-        return allow_use_jemalloc_memory.load(std::memory_order_relaxed);
-    }
 
     /** Set limit if it was not set.
       * Otherwise, set limit to new value, if new value is greater than previous limit.
@@ -249,10 +244,9 @@ public:
     /// Reset the accumulated data.
     void reset();
 
-    /// Reset current counter to an RSS value.
-    /// Jemalloc may have pre-allocated arenas, they are accounted in RSS.
-    /// We can free this arenas in case of exception to avoid OOM.
-    static void setRSS(Int64 rss_, Int64 free_memory_in_allocator_arenas_);
+    /// update values based on external information (e.g. jemalloc's stat)
+    static void updateRSS(Int64 rss_);
+    static void updateAllocated(Int64 allocated_);
 
     /// Prints info about peak memory consumption into log.
     void logPeakMemoryUsage();

src/Common/MemoryWorker.cpp

@@ -0,0 +1,333 @@
+#include <Common/MemoryWorker.h>
+
+#include <IO/ReadBufferFromFile.h>
+#include <IO/ReadBufferFromFileDescriptor.h>
+#include <IO/ReadHelpers.h>
+#include <base/cgroupsv2.h>
+#include <Common/Jemalloc.h>
+#include <Common/MemoryTracker.h>
+#include <Common/ProfileEvents.h>
+#include <Common/formatReadable.h>
+#include <Common/logger_useful.h>
+
+#include <fmt/ranges.h>
+
+#include <filesystem>
+#include <optional>
+
+namespace fs = std::filesystem;
+
+namespace ProfileEvents
+{
+    extern const Event MemoryAllocatorPurge;
+    extern const Event MemoryAllocatorPurgeTimeMicroseconds;
+    extern const Event MemoryWorkerRun;
+    extern const Event MemoryWorkerRunElapsedMicroseconds;
+}
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int FILE_DOESNT_EXIST;
+    extern const int LOGICAL_ERROR;
+}
+
+#if defined(OS_LINUX)
+namespace
+{
+
+using Metrics = std::map<std::string, uint64_t>;
+
+/// Format is
+///   kernel 5
+///   rss 15
+///   [...]
+Metrics readAllMetricsFromStatFile(ReadBufferFromFile & buf)
+{
+    Metrics metrics;
+    while (!buf.eof())
+    {
+        std::string current_key;
+        readStringUntilWhitespace(current_key, buf);
+
+        assertChar(' ', buf);
+
+        uint64_t value = 0;
+        readIntText(value, buf);
+        assertChar('\n', buf);
+
+        auto [_, inserted] = metrics.emplace(std::move(current_key), value);
+        chassert(inserted, "Duplicate keys in stat file");
+    }
+    return metrics;
+}
+
+uint64_t readMetricFromStatFile(ReadBufferFromFile & buf, std::string_view key)
+{
+    while (!buf.eof())
+    {
+        std::string current_key;
+        readStringUntilWhitespace(current_key, buf);
+        if (current_key != key)
+        {
+            std::string dummy;
+            readStringUntilNewlineInto(dummy, buf);
+            buf.ignore();
+            continue;
+        }
+
+        assertChar(' ', buf);
+        uint64_t value = 0;
+        readIntText(value, buf);
+        return value;
+    }
+    LOG_ERROR(getLogger("CgroupsReader"), "Cannot find '{}' in '{}'", key, buf.getFileName());
+    return 0;
+}
+
+struct CgroupsV1Reader : ICgroupsReader
+{
+    explicit CgroupsV1Reader(const fs::path & stat_file_dir) : buf(stat_file_dir / "memory.stat") { }
+
+    uint64_t readMemoryUsage() override
+    {
+        std::lock_guard lock(mutex);
+        buf.rewind();
+        return readMetricFromStatFile(buf, "rss");
+    }
+
+    std::string dumpAllStats() override
+    {
+        std::lock_guard lock(mutex);
+        buf.rewind();
+        return fmt::format("{}", readAllMetricsFromStatFile(buf));
+    }
+
+private:
+    std::mutex mutex;
+    ReadBufferFromFile buf TSA_GUARDED_BY(mutex);
+};
+
+struct CgroupsV2Reader : ICgroupsReader
+{
+    explicit CgroupsV2Reader(const fs::path & stat_file_dir) : stat_buf(stat_file_dir / "memory.stat") { }
+
+    uint64_t readMemoryUsage() override
+    {
+        std::lock_guard lock(mutex);
+        stat_buf.rewind();
+        return readMetricFromStatFile(stat_buf, "anon");
+    }
+
+    std::string dumpAllStats() override
+    {
+        std::lock_guard lock(mutex);
+        stat_buf.rewind();
+        return fmt::format("{}", readAllMetricsFromStatFile(stat_buf));
+    }
+
+private:
+    std::mutex mutex;
+    ReadBufferFromFile stat_buf TSA_GUARDED_BY(mutex);
+};
+
+/// Caveats:
+/// - All of the logic in this file assumes that the current process is the only process in the
+///   containing cgroup (or more precisely: the only process with significant memory consumption).
+///   If this is not the case, then other processe's memory consumption may affect the internal
+///   memory tracker ...
+/// - Cgroups v1 and v2 allow nested cgroup hierarchies. As v1 is deprecated for over half a
+///   decade and will go away at some point, hierarchical detection is only implemented for v2.
+/// - I did not test what happens if a host has v1 and v2 simultaneously enabled. I believe such
+///   systems existed only for a short transition period.
+
+std::optional<std::string> getCgroupsV1Path()
+{
+    auto path = default_cgroups_mount / "memory/memory.stat";
+    if (!fs::exists(path))
+        return {};
+    return {default_cgroups_mount / "memory"};
+}
+
+std::pair<std::string, ICgroupsReader::CgroupsVersion> getCgroupsPath()
+{
+    auto v2_path = getCgroupsV2PathContainingFile("memory.current");
+    if (v2_path.has_value())
+        return {*v2_path, ICgroupsReader::CgroupsVersion::V2};
+
+    auto v1_path = getCgroupsV1Path();
+    if (v1_path.has_value())
+        return {*v1_path, ICgroupsReader::CgroupsVersion::V1};
+
+    throw Exception(ErrorCodes::FILE_DOESNT_EXIST, "Cannot find cgroups v1 or v2 current memory file");
+}
+
+}
+
+std::shared_ptr<ICgroupsReader> ICgroupsReader::createCgroupsReader(ICgroupsReader::CgroupsVersion version, const std::filesystem::path & cgroup_path)
+{
+    if (version == CgroupsVersion::V2)
+        return std::make_shared<CgroupsV2Reader>(cgroup_path);
+    else
+    {
+        chassert(version == CgroupsVersion::V1);
+        return std::make_shared<CgroupsV1Reader>(cgroup_path);
+    }
+}
+#endif
+
+namespace
+{
+
+std::string_view sourceToString(MemoryWorker::MemoryUsageSource source)
+{
+    switch (source)
+    {
+        case MemoryWorker::MemoryUsageSource::Cgroups: return "Cgroups";
+        case MemoryWorker::MemoryUsageSource::Jemalloc: return "Jemalloc";
+        case MemoryWorker::MemoryUsageSource::None: return "None";
+    }
+}
+
+}
+
+/// We try to pick the best possible supported source for reading memory usage.
+/// Supported sources in order of priority
+/// - reading from cgroups' pseudo-files (fastest and most accurate)
+/// - reading jemalloc's resident stat (doesn't take into account allocations that didn't use jemalloc)
+/// Also, different tick rates are used because not all options are equally fast
+MemoryWorker::MemoryWorker(uint64_t period_ms_)
+    : log(getLogger("MemoryWorker"))
+    , period_ms(period_ms_)
+{
+#if defined(OS_LINUX)
+    try
+    {
+        static constexpr uint64_t cgroups_memory_usage_tick_ms{50};
+
+        const auto [cgroup_path, version] = getCgroupsPath();
+        LOG_INFO(
+            getLogger("CgroupsReader"),
+            "Will create cgroup reader from '{}' (cgroups version: {})",
+            cgroup_path,
+            (version == ICgroupsReader::CgroupsVersion::V1) ? "v1" : "v2");
+
+        cgroups_reader = ICgroupsReader::createCgroupsReader(version, cgroup_path);
+        source = MemoryUsageSource::Cgroups;
+        if (period_ms == 0)
+            period_ms = cgroups_memory_usage_tick_ms;
+
+        return;
+    }
+    catch (...)
+    {
+        tryLogCurrentException(log, "Cannot use cgroups reader");
+    }
+#endif
+
+#if USE_JEMALLOC
+    static constexpr uint64_t jemalloc_memory_usage_tick_ms{100};
+
+    source = MemoryUsageSource::Jemalloc;
+    if (period_ms == 0)
+        period_ms = jemalloc_memory_usage_tick_ms;
+#endif
+}
+
+MemoryWorker::MemoryUsageSource MemoryWorker::getSource()
+{
+    return source;
+}
+
+void MemoryWorker::start()
+{
+    if (source == MemoryUsageSource::None)
+        return;
+
+    LOG_INFO(
+        getLogger("MemoryWorker"),
+        "Starting background memory thread with period of {}ms, using {} as source",
+        period_ms,
+        sourceToString(source));
+    background_thread = ThreadFromGlobalPool([this] { backgroundThread(); });
+}
+
+MemoryWorker::~MemoryWorker()
+{
+    {
+        std::unique_lock lock(mutex);
+        shutdown = true;
+    }
+    cv.notify_all();
+
+    if (background_thread.joinable())
+        background_thread.join();
+}
+
+uint64_t MemoryWorker::getMemoryUsage()
+{
+    switch (source)
+    {
+        case MemoryUsageSource::Cgroups:
+            return cgroups_reader != nullptr ? cgroups_reader->readMemoryUsage() : 0;
+        case MemoryUsageSource::Jemalloc:
+#if USE_JEMALLOC
+            return resident_mib.getValue();
+#else
+            return 0;
+#endif
+        case MemoryUsageSource::None:
+            throw DB::Exception(ErrorCodes::LOGICAL_ERROR, "Trying to fetch memory usage while no memory source can be used");
+    }
+}
+
+void MemoryWorker::backgroundThread()
+{
+    std::chrono::milliseconds chrono_period_ms{period_ms};
+    [[maybe_unused]] bool first_run = true;
+    std::unique_lock lock(mutex);
+    while (true)
+    {
+        cv.wait_for(lock, chrono_period_ms, [this] { return shutdown; });
+        if (shutdown)
+            return;
+
+        Stopwatch total_watch;
+
+#if USE_JEMALLOC
+        if (source == MemoryUsageSource::Jemalloc)
+            epoch_mib.setValue(0);
+#endif
+
+        Int64 resident = getMemoryUsage();
+        MemoryTracker::updateRSS(resident);
+
+#if USE_JEMALLOC
+        if (resident > total_memory_tracker.getHardLimit())
+        {
+            Stopwatch purge_watch;
+            purge_mib.run();
+            ProfileEvents::increment(ProfileEvents::MemoryAllocatorPurge);
+            ProfileEvents::increment(ProfileEvents::MemoryAllocatorPurgeTimeMicroseconds, purge_watch.elapsedMicroseconds());
+        }
+#endif
+
+#if USE_JEMALLOC
+        if (unlikely(first_run || total_memory_tracker.get() < 0))
+        {
+            if (source != MemoryUsageSource::Jemalloc)
+                epoch_mib.setValue(0);
+
+            MemoryTracker::updateAllocated(allocated_mib.getValue());
+        }
+#endif
+
+        ProfileEvents::increment(ProfileEvents::MemoryWorkerRun);
+        ProfileEvents::increment(ProfileEvents::MemoryWorkerRunElapsedMicroseconds, total_watch.elapsedMicroseconds());
+        first_run = false;
+    }
+}
+
+}

src/Common/MemoryWorker.h

@@ -0,0 +1,84 @@
+#pragma once
+
+#include <Common/CgroupsMemoryUsageObserver.h>
+#include <Common/ThreadPool.h>
+#include <Common/Jemalloc.h>
+
+namespace DB
+{
+
+struct ICgroupsReader
+{
+    enum class CgroupsVersion : uint8_t
+    {
+        V1,
+        V2
+    };
+
+#if defined(OS_LINUX)
+    static std::shared_ptr<ICgroupsReader>
+    createCgroupsReader(ICgroupsReader::CgroupsVersion version, const std::filesystem::path & cgroup_path);
+#endif
+
+    virtual ~ICgroupsReader() = default;
+
+    virtual uint64_t readMemoryUsage() = 0;
+
+    virtual std::string dumpAllStats() = 0;
+};
+
+
+/// Correct MemoryTracker based on external information (e.g. Cgroups or stats.resident from jemalloc)
+/// The worker spawns a background thread which periodically reads current resident memory from the source,
+/// whose value is sent to global MemoryTracker.
+/// It can do additional things like purging jemalloc dirty pages if the current memory usage is higher than global hard limit.
+class MemoryWorker
+{
+public:
+    explicit MemoryWorker(uint64_t period_ms_);
+
+    enum class MemoryUsageSource : uint8_t
+    {
+        None,
+        Cgroups,
+        Jemalloc
+    };
+
+    MemoryUsageSource getSource();
+
+    void start();
+
+    ~MemoryWorker();
+private:
+    uint64_t getMemoryUsage();
+
+    void backgroundThread();
+
+    ThreadFromGlobalPool background_thread;
+
+    std::mutex mutex;
+    std::condition_variable cv;
+    bool shutdown = false;
+
+    LoggerPtr log;
+
+    uint64_t period_ms;
+
+    MemoryUsageSource source{MemoryUsageSource::None};
+
+    std::shared_ptr<ICgroupsReader> cgroups_reader;
+
+#if USE_JEMALLOC
+    JemallocMibCache<uint64_t> epoch_mib{"epoch"};
+    JemallocMibCache<size_t> resident_mib{"stats.resident"};
+    JemallocMibCache<size_t> allocated_mib{"stats.allocated"};
+
+#define STRINGIFY_HELPER(x) #x
+#define STRINGIFY(x) STRINGIFY_HELPER(x)
+    JemallocMibCache<size_t> purge_mib{"arena." STRINGIFY(MALLCTL_ARENAS_ALL) ".purge"};
+#undef STRINGIFY
+#undef STRINGIFY_HELPER
+#endif
+};
+
+}

src/Common/PoolWithFailoverBase.h

@@ -122,12 +122,18 @@ public:
         return result.entry.isNull() || !result.is_usable || (skip_read_only_replicas && result.is_readonly);
     }
 
-    void checkTryResultIsValid(const TryResult & result, bool skip_read_only_replicas) const
+    TryResult getValidTryResult(const std::vector<TryResult> & results, bool skip_read_only_replicas) const
     {
+        if (results.empty())
+            throw DB::Exception(DB::ErrorCodes::ALL_CONNECTION_TRIES_FAILED, "Cannot get any valid connection because all connection tries failed");
+
+        auto result = results.front();
         if (isTryResultInvalid(result, skip_read_only_replicas))
             throw DB::Exception(DB::ErrorCodes::LOGICAL_ERROR,
                 "Got an invalid connection result: entry.isNull {}, is_usable {}, is_up_to_date {}, delay {}, is_readonly {}, skip_read_only_replicas {}",
                 result.entry.isNull(), result.is_usable, result.is_up_to_date, result.delay, result.is_readonly, skip_read_only_replicas);
+
+        return result;
     }
 
     size_t getPoolSize() const { return nested_pools.size(); }

src/Common/ProfileEvents.cpp

@@ -827,6 +827,9 @@ The server successfully detected this situation and will download merged part fr
     M(GWPAsanAllocateSuccess, "Number of successful allocations done by GWPAsan") \
     M(GWPAsanAllocateFailed, "Number of failed allocations done by GWPAsan (i.e. filled pool)") \
     M(GWPAsanFree, "Number of free operations done by GWPAsan") \
+    \
+    M(MemoryWorkerRun, "Number of runs done by MemoryWorker in background") \
+    M(MemoryWorkerRunElapsedMicroseconds, "Total time spent by MemoryWorker for background work") \
 
 
 #ifdef APPLY_FOR_EXTERNAL_EVENTS

src/Common/ZooKeeper/IKeeper.cpp

@@ -171,6 +171,7 @@ bool isUserError(Error zk_return_code)
 
 void CreateRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
 void RemoveRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
+void RemoveRecursiveRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
 void ExistsRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
 void GetRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }
 void SetRequest::addRootPath(const String & root_path) { Coordination::addRootPath(path, root_path); }

src/Common/ZooKeeper/IKeeper.h

@@ -248,6 +248,23 @@ struct RemoveResponse : virtual Response
 {
 };
 
+struct RemoveRecursiveRequest : virtual Request
+{
+    String path;
+
+    /// strict limit for number of deleted nodes
+    uint32_t remove_nodes_limit = 1;
+
+    void addRootPath(const String & root_path) override;
+    String getPath() const override { return path; }
+
+    size_t bytesSize() const override { return path.size() + sizeof(remove_nodes_limit); }
+};
+
+struct RemoveRecursiveResponse : virtual Response
+{
+};
+
 struct ExistsRequest : virtual Request
 {
     String path;
@@ -430,6 +447,7 @@ struct ErrorResponse : virtual Response
 
 using CreateCallback = std::function<void(const CreateResponse &)>;
 using RemoveCallback = std::function<void(const RemoveResponse &)>;
+using RemoveRecursiveCallback = std::function<void(const RemoveRecursiveResponse &)>;
 using ExistsCallback = std::function<void(const ExistsResponse &)>;
 using GetCallback = std::function<void(const GetResponse &)>;
 using SetCallback = std::function<void(const SetResponse &)>;
@@ -587,6 +605,11 @@ public:
         int32_t version,
         RemoveCallback callback) = 0;
 
+    virtual void removeRecursive(
+        const String & path,
+        uint32_t remove_nodes_limit,
+        RemoveRecursiveCallback callback) = 0;
+
     virtual void exists(
         const String & path,
         ExistsCallback callback,

src/Common/ZooKeeper/TestKeeper.cpp

@@ -90,6 +90,36 @@ struct TestKeeperRemoveRequest final : RemoveRequest, TestKeeperRequest
     }
 };
 
+struct TestKeeperRemoveRecursiveRequest final : RemoveRecursiveRequest, TestKeeperRequest
+{
+    TestKeeperRemoveRecursiveRequest() = default;
+    explicit TestKeeperRemoveRecursiveRequest(const RemoveRecursiveRequest & base) : RemoveRecursiveRequest(base) {}
+    ResponsePtr createResponse() const override;
+    std::pair<ResponsePtr, Undo> process(TestKeeper::Container & container, int64_t zxid) const override;
+
+    void processWatches(TestKeeper::Watches & node_watches, TestKeeper::Watches & list_watches) const override
+    {
+        std::vector<std::pair<String, size_t>> deleted;
+
+        auto add_deleted_watches = [&](TestKeeper::Watches & w)
+        {
+            for (const auto & [watch_path, _] : w)
+                if (watch_path.starts_with(path))
+                    deleted.emplace_back(watch_path, std::count(watch_path.begin(), watch_path.end(), '/'));
+        };
+
+        add_deleted_watches(node_watches);
+        add_deleted_watches(list_watches);
+        std::sort(deleted.begin(), deleted.end(), [](const auto & lhs, const auto & rhs)
+        {
+            return lhs.second < rhs.second;
+        });
+
+        for (const auto & [watch_path, _] : deleted)
+            processWatchesImpl(watch_path, node_watches, list_watches);
+    }
+};
+
 struct TestKeeperExistsRequest final : ExistsRequest, TestKeeperRequest
 {
     ResponsePtr createResponse() const override;
@@ -175,6 +205,10 @@ struct TestKeeperMultiRequest final : MultiRequest, TestKeeperRequest
             {
                 requests.push_back(std::make_shared<TestKeeperRemoveRequest>(*concrete_request_remove));
             }
+            else if (const auto * concrete_request_remove_recursive = dynamic_cast<const RemoveRecursiveRequest *>(generic_request.get()))
+            {
+                requests.push_back(std::make_shared<TestKeeperRemoveRecursiveRequest>(*concrete_request_remove_recursive));
+            }
             else if (const auto * concrete_request_set = dynamic_cast<const SetRequest *>(generic_request.get()))
             {
                 requests.push_back(std::make_shared<TestKeeperSetRequest>(*concrete_request_set));
@@ -313,6 +347,62 @@ std::pair<ResponsePtr, Undo> TestKeeperRemoveRequest::process(TestKeeper::Contai
     return { std::make_shared<RemoveResponse>(response), undo };
 }
 
+std::pair<ResponsePtr, Undo> TestKeeperRemoveRecursiveRequest::process(TestKeeper::Container & container, int64_t zxid) const
+{
+    RemoveRecursiveResponse response;
+    response.zxid = zxid;
+    Undo undo;
+
+    auto root_it = container.find(path);
+    if (root_it == container.end())
+    {
+        response.error = Error::ZNONODE;
+        return { std::make_shared<RemoveRecursiveResponse>(response), undo };
+    }
+
+    std::vector<std::pair<std::string, Coordination::TestKeeper::Node>> children;
+
+    for (auto it = std::next(root_it); it != container.end(); ++it)
+    {
+        const auto & [child_path, child_node] = *it;
+
+        if (child_path.starts_with(path))
+            children.emplace_back(child_path, child_node);
+        else
+            break;
+    }
+
+    if (children.size() > remove_nodes_limit)
+    {
+        response.error = Error::ZNOTEMPTY;
+        return { std::make_shared<RemoveRecursiveResponse>(response), undo };
+    }
+
+    auto & parent = container.at(parentPath(path));
+    --parent.stat.numChildren;
+    ++parent.stat.cversion;
+
+    for (const auto & [child_path, child_node] : children)
+    {
+        auto child_it = container.find(child_path);
+        chassert(child_it != container.end());
+        container.erase(child_it);
+    }
+
+    response.error = Error::ZOK;
+    undo = [&container, dead = std::move(children), root_path = path]()
+    {
+        for (auto && [child_path, child_node] : dead)
+            container.emplace(child_path, child_node);
+
+        auto & undo_parent = container.at(parentPath(root_path));
+        ++undo_parent.stat.numChildren;
+        --undo_parent.stat.cversion;
+    };
+
+    return { std::make_shared<RemoveRecursiveResponse>(response), undo };
+}
+
 std::pair<ResponsePtr, Undo> TestKeeperExistsRequest::process(TestKeeper::Container & container, int64_t zxid) const
 {
     ExistsResponse response;
@@ -530,6 +620,7 @@ std::pair<ResponsePtr, Undo> TestKeeperMultiRequest::process(TestKeeper::Contain
 
 ResponsePtr TestKeeperCreateRequest::createResponse() const { return std::make_shared<CreateResponse>(); }
 ResponsePtr TestKeeperRemoveRequest::createResponse() const { return std::make_shared<RemoveResponse>(); }
+ResponsePtr TestKeeperRemoveRecursiveRequest::createResponse() const { return std::make_shared<RemoveRecursiveResponse>(); }
 ResponsePtr TestKeeperExistsRequest::createResponse() const { return std::make_shared<ExistsResponse>(); }
 ResponsePtr TestKeeperGetRequest::createResponse() const { return std::make_shared<GetResponse>(); }
 ResponsePtr TestKeeperSetRequest::createResponse() const { return std::make_shared<SetResponse>(); }
@@ -771,6 +862,21 @@ void TestKeeper::remove(
     pushRequest(std::move(request_info));
 }
 
+void TestKeeper::removeRecursive(
+    const String & path,
+    uint32_t remove_nodes_limit,
+    RemoveRecursiveCallback callback)
+{
+    TestKeeperRemoveRecursiveRequest request;
+    request.path = path;
+    request.remove_nodes_limit = remove_nodes_limit;
+
+    RequestInfo request_info;
+    request_info.request = std::make_shared<TestKeeperRemoveRecursiveRequest>(std::move(request));
+    request_info.callback = [callback](const Response & response) { callback(dynamic_cast<const RemoveRecursiveResponse &>(response)); };
+    pushRequest(std::move(request_info));
+}
+
 void TestKeeper::exists(
         const String & path,
         ExistsCallback callback,

src/Common/ZooKeeper/TestKeeper.h

@@ -58,6 +58,11 @@ public:
             int32_t version,
             RemoveCallback callback) override;
 
+    void removeRecursive(
+        const String & path,
+        uint32_t remove_nodes_limit,
+        RemoveRecursiveCallback callback) override;
+
     void exists(
             const String & path,
             ExistsCallback callback,

src/Common/ZooKeeper/Types.h

@@ -31,6 +31,7 @@ using AsyncResponses = std::vector<std::pair<std::string, std::future<R>>>;
 
 Coordination::RequestPtr makeCreateRequest(const std::string & path, const std::string & data, int create_mode, bool ignore_if_exists = false);
 Coordination::RequestPtr makeRemoveRequest(const std::string & path, int version);
+Coordination::RequestPtr makeRemoveRecursiveRequest(const std::string & path, uint32_t remove_nodes_limit);
 Coordination::RequestPtr makeSetRequest(const std::string & path, const std::string & data, int version);
 Coordination::RequestPtr makeCheckRequest(const std::string & path, int version);
 

src/Common/ZooKeeper/ZooKeeper.cpp

@@ -979,18 +979,47 @@ bool ZooKeeper::tryRemoveChildrenRecursive(const std::string & path, bool probab
     return removed_as_expected;
 }
 
-void ZooKeeper::removeRecursive(const std::string & path)
+void ZooKeeper::removeRecursive(const std::string & path, uint32_t remove_nodes_limit)
+{
+    if (!isFeatureEnabled(DB::KeeperFeatureFlag::REMOVE_RECURSIVE))
     {
         removeChildrenRecursive(path);
         remove(path);
+        return;
     }
 
-void ZooKeeper::tryRemoveRecursive(const std::string & path)
+    check(tryRemoveRecursive(path, remove_nodes_limit), path);
+}
+
+Coordination::Error ZooKeeper::tryRemoveRecursive(const std::string & path, uint32_t remove_nodes_limit)
+{
+    if (!isFeatureEnabled(DB::KeeperFeatureFlag::REMOVE_RECURSIVE))
     {
         tryRemoveChildrenRecursive(path);
-    tryRemove(path);
+        return tryRemove(path);
     }
 
+    auto promise = std::make_shared<std::promise<Coordination::RemoveRecursiveResponse>>();
+    auto future = promise->get_future();
+
+    auto callback = [promise](const Coordination::RemoveRecursiveResponse & response) mutable
+    {
+        promise->set_value(response);
+    };
+
+    impl->removeRecursive(path, remove_nodes_limit, std::move(callback));
+
+    if (future.wait_for(std::chrono::milliseconds(args.operation_timeout_ms)) != std::future_status::ready)
+    {
+        impl->finalize(fmt::format("Operation timeout on {} {}", Coordination::OpNum::RemoveRecursive, path));
+        return Coordination::Error::ZOPERATIONTIMEOUT;
+    }
+    else
+    {
+        auto response = future.get();
+        return response.error;
+    }
+}
 
 namespace
 {
@@ -1619,6 +1648,14 @@ Coordination::RequestPtr makeRemoveRequest(const std::string & path, int version
     return request;
 }
 
+Coordination::RequestPtr makeRemoveRecursiveRequest(const std::string & path, uint32_t remove_nodes_limit)
+{
+    auto request = std::make_shared<Coordination::RemoveRecursiveRequest>();
+    request->path = path;
+    request->remove_nodes_limit = remove_nodes_limit;
+    return request;
+}
+
 Coordination::RequestPtr makeSetRequest(const std::string & path, const std::string & data, int version)
 {
     auto request = std::make_shared<Coordination::SetRequest>();

src/Common/ZooKeeper/ZooKeeper.h

@@ -479,15 +479,16 @@ public:
 
     Int64 getClientID();
 
-    /// Remove the node with the subtree. If someone concurrently adds or removes a node
-    /// in the subtree, the result is undefined.
-    void removeRecursive(const std::string & path);
+    /// Remove the node with the subtree.
+    /// If Keeper supports RemoveRecursive operation then it will be performed atomically.
+    /// Otherwise if someone concurrently adds or removes a node in the subtree, the result is undefined.
+    void removeRecursive(const std::string & path, uint32_t remove_nodes_limit = 100);
 
-    /// Remove the node with the subtree. If someone concurrently removes a node in the subtree,
-    /// this will not cause errors.
+    /// Same as removeRecursive but in case if Keeper does not supports RemoveRecursive and
+    /// if someone concurrently removes a node in the subtree, this will not cause errors.
     /// For instance, you can call this method twice concurrently for the same node and the end
     /// result would be the same as for the single call.
-    void tryRemoveRecursive(const std::string & path);
+    Coordination::Error tryRemoveRecursive(const std::string & path, uint32_t remove_nodes_limit = 100);
 
     /// Similar to removeRecursive(...) and tryRemoveRecursive(...), but does not remove path itself.
     /// Node defined as RemoveException will not be deleted.

src/Common/ZooKeeper/ZooKeeperCommon.cpp

@@ -1,5 +1,5 @@
-#include "Common/ZooKeeper/IKeeper.h"
-#include "Common/ZooKeeper/ZooKeeperConstants.h"
+#include <Common/ZooKeeper/IKeeper.h>
+#include <Common/ZooKeeper/ZooKeeperConstants.h>
 #include <Common/ZooKeeper/ZooKeeperCommon.h>
 #include <Common/ZooKeeper/ZooKeeperIO.h>
 #include <Common/Stopwatch.h>
@@ -232,6 +232,27 @@ void ZooKeeperRemoveRequest::readImpl(ReadBuffer & in)
     Coordination::read(version, in);
 }
 
+void ZooKeeperRemoveRecursiveRequest::writeImpl(WriteBuffer & out) const
+{
+    Coordination::write(path, out);
+    Coordination::write(remove_nodes_limit, out);
+}
+
+void ZooKeeperRemoveRecursiveRequest::readImpl(ReadBuffer & in)
+{
+    Coordination::read(path, in);
+    Coordination::read(remove_nodes_limit, in);
+}
+
+std::string ZooKeeperRemoveRecursiveRequest::toStringImpl(bool /*short_format*/) const
+{
+    return fmt::format(
+        "path = {}\n"
+        "remove_nodes_limit = {}",
+        path,
+        remove_nodes_limit);
+}
+
 void ZooKeeperExistsRequest::writeImpl(WriteBuffer & out) const
 {
     Coordination::write(path, out);
@@ -510,6 +531,11 @@ ZooKeeperMultiRequest::ZooKeeperMultiRequest(std::span<const Coordination::Reque
             checkOperationType(Write);
             requests.push_back(std::make_shared<ZooKeeperRemoveRequest>(*concrete_request_remove));
         }
+        else if (const auto * concrete_request_remove_recursive = dynamic_cast<const RemoveRecursiveRequest *>(generic_request.get()))
+        {
+            checkOperationType(Write);
+            requests.push_back(std::make_shared<ZooKeeperRemoveRecursiveRequest>(*concrete_request_remove_recursive));
+        }
         else if (const auto * concrete_request_set = dynamic_cast<const SetRequest *>(generic_request.get()))
         {
             checkOperationType(Write);
@@ -707,6 +733,7 @@ ZooKeeperResponsePtr ZooKeeperHeartbeatRequest::makeResponse() const { return se
 ZooKeeperResponsePtr ZooKeeperSyncRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperSyncResponse>()); }
 ZooKeeperResponsePtr ZooKeeperAuthRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperAuthResponse>()); }
 ZooKeeperResponsePtr ZooKeeperRemoveRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperRemoveResponse>()); }
+ZooKeeperResponsePtr ZooKeeperRemoveRecursiveRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperRemoveRecursiveResponse>()); }
 ZooKeeperResponsePtr ZooKeeperExistsRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperExistsResponse>()); }
 ZooKeeperResponsePtr ZooKeeperGetRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperGetResponse>()); }
 ZooKeeperResponsePtr ZooKeeperSetRequest::makeResponse() const { return setTime(std::make_shared<ZooKeeperSetResponse>()); }
@@ -1024,6 +1051,7 @@ ZooKeeperRequestFactory::ZooKeeperRequestFactory()
     registerZooKeeperRequest<OpNum::SetACL, ZooKeeperSetACLRequest>(*this);
     registerZooKeeperRequest<OpNum::FilteredList, ZooKeeperFilteredListRequest>(*this);
     registerZooKeeperRequest<OpNum::CheckNotExists, ZooKeeperCheckRequest>(*this);
+    registerZooKeeperRequest<OpNum::RemoveRecursive, ZooKeeperRemoveRecursiveRequest>(*this);
 }
 
 PathMatchResult matchPath(std::string_view path, std::string_view match_to)

src/Common/ZooKeeper/ZooKeeperCommon.h

@@ -285,6 +285,31 @@ struct ZooKeeperRemoveResponse final : RemoveResponse, ZooKeeperResponse
     size_t bytesSize() const override { return RemoveResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
 };
 
+struct ZooKeeperRemoveRecursiveRequest final : RemoveRecursiveRequest, ZooKeeperRequest
+{
+    ZooKeeperRemoveRecursiveRequest() = default;
+    explicit ZooKeeperRemoveRecursiveRequest(const RemoveRecursiveRequest & base) : RemoveRecursiveRequest(base) {}
+
+    OpNum getOpNum() const override { return OpNum::RemoveRecursive; }
+    void writeImpl(WriteBuffer & out) const override;
+    void readImpl(ReadBuffer & in) override;
+    std::string toStringImpl(bool short_format) const override;
+
+    ZooKeeperResponsePtr makeResponse() const override;
+    bool isReadRequest() const override { return false; }
+
+    size_t bytesSize() const override { return RemoveRecursiveRequest::bytesSize() + sizeof(xid); }
+};
+
+struct ZooKeeperRemoveRecursiveResponse : RemoveRecursiveResponse, ZooKeeperResponse
+{
+    void readImpl(ReadBuffer &) override {}
+    void writeImpl(WriteBuffer &) const override {}
+    OpNum getOpNum() const override { return OpNum::RemoveRecursive; }
+
+    size_t bytesSize() const override { return RemoveRecursiveResponse::bytesSize() + sizeof(xid) + sizeof(zxid); }
+};
+
 struct ZooKeeperExistsRequest final : ExistsRequest, ZooKeeperRequest
 {
     ZooKeeperExistsRequest() = default;

src/Common/ZooKeeper/ZooKeeperConstants.cpp

@@ -29,6 +29,7 @@ static const std::unordered_set<int32_t> VALID_OPERATIONS =
     static_cast<int32_t>(OpNum::GetACL),
     static_cast<int32_t>(OpNum::FilteredList),
     static_cast<int32_t>(OpNum::CheckNotExists),
+    static_cast<int32_t>(OpNum::RemoveRecursive),
 };
 
 OpNum getOpNum(int32_t raw_op_num)

src/Common/ZooKeeper/ZooKeeperConstants.h

@@ -40,6 +40,7 @@ enum class OpNum : int32_t
     FilteredList = 500,
     CheckNotExists = 501,
     CreateIfNotExists = 502,
+    RemoveRecursive = 503,
 
     SessionID = 997, /// Special internal request
 };

src/Common/ZooKeeper/ZooKeeperImpl.cpp

@@ -1347,6 +1347,25 @@ void ZooKeeper::remove(
     ProfileEvents::increment(ProfileEvents::ZooKeeperRemove);
 }
 
+void ZooKeeper::removeRecursive(
+    const String &path,
+    uint32_t remove_nodes_limit,
+    RemoveRecursiveCallback callback)
+{
+    if (!isFeatureEnabled(KeeperFeatureFlag::REMOVE_RECURSIVE))
+        throw Exception::fromMessage(Error::ZBADARGUMENTS, "RemoveRecursive request type cannot be used because it's not supported by the server");
+
+    ZooKeeperRemoveRecursiveRequest request;
+    request.path = path;
+    request.remove_nodes_limit = remove_nodes_limit;
+
+    RequestInfo request_info;
+    request_info.request = std::make_shared<ZooKeeperRemoveRecursiveRequest>(std::move(request));
+    request_info.callback = [callback](const Response & response) { callback(dynamic_cast<const RemoveRecursiveResponse &>(response)); };
+
+    pushRequest(std::move(request_info));
+    ProfileEvents::increment(ProfileEvents::ZooKeeperRemove);
+}
 
 void ZooKeeper::exists(
     const String & path,

src/Common/ZooKeeper/ZooKeeperImpl.h

@@ -146,6 +146,11 @@ public:
         int32_t version,
         RemoveCallback callback) override;
 
+    void removeRecursive(
+        const String &path,
+        uint32_t remove_nodes_limit,
+        RemoveRecursiveCallback callback) override;
+
     void exists(
         const String & path,
         ExistsCallback callback,

src/Common/atomicRename.cpp

@@ -57,11 +57,13 @@ namespace ErrorCodes
 namespace DB
 {
 
-static bool supportsAtomicRenameImpl()
+static std::optional<std::string> supportsAtomicRenameImpl()
 {
     VersionNumber renameat2_minimal_version(3, 15, 0);
     VersionNumber linux_version(Poco::Environment::osVersion());
-    return linux_version >= renameat2_minimal_version;
+    if (linux_version >= renameat2_minimal_version)
+        return std::nullopt;
+    return fmt::format("Linux kernel 3.15+ is required, got {}", linux_version.toString());
 }
 
 static bool renameat2(const std::string & old_path, const std::string & new_path, int flags)
@@ -97,10 +99,14 @@ static bool renameat2(const std::string & old_path, const std::string & new_path
     ErrnoException::throwFromPath(ErrorCodes::SYSTEM_ERROR, new_path, "Cannot rename {} to {}", old_path, new_path);
 }
 
-bool supportsAtomicRename()
+bool supportsAtomicRename(std::string * out_message)
 {
-    static bool supports = supportsAtomicRenameImpl();
-    return supports;
+    static auto error = supportsAtomicRenameImpl();
+    if (!error.has_value())
+        return true;
+    if (out_message)
+        *out_message = error.value();
+    return false;
 }
 
 }
@@ -152,16 +158,22 @@ static bool renameat2(const std::string & old_path, const std::string & new_path
 }
 
 
-static bool supportsAtomicRenameImpl()
+static std::optional<std::string> supportsAtomicRenameImpl()
 {
     auto fun = dlsym(RTLD_DEFAULT, "renamex_np");
-    return fun != nullptr;
+    if (fun != nullptr)
+        return std::nullopt;
+    return "macOS 10.12 or later is required";
 }
 
-bool supportsAtomicRename()
+bool supportsAtomicRename(std::string * out_message)
 {
-    static bool supports = supportsAtomicRenameImpl();
-    return supports;
+    static auto error = supportsAtomicRenameImpl();
+    if (!error.has_value())
+        return true;
+    if (out_message)
+        *out_message = error.value();
+    return false;
 }
 
 }
@@ -179,8 +191,10 @@ static bool renameat2(const std::string &, const std::string &, int)
     return false;
 }
 
-bool supportsAtomicRename()
+bool supportsAtomicRename(std::string * out_message)
 {
+    if (out_message)
+        *out_message = "only Linux and macOS are supported";
     return false;
 }
 

src/Common/atomicRename.h

@@ -6,7 +6,7 @@ namespace DB
 {
 
 /// Returns true, if the following functions supported by the system
-bool supportsAtomicRename();
+bool supportsAtomicRename(std::string * out_message = nullptr);
 
 /// Atomically rename old_path to new_path. If new_path exists, do not overwrite it and throw exception
 void renameNoReplace(const std::string & old_path, const std::string & new_path);

src/Common/tests/gtest_cgroups_reader.cpp

@@ -6,7 +6,7 @@
 #include <filesystem>
 
 #include <IO/WriteBufferFromFile.h>
-#include <Common/CgroupsMemoryUsageObserver.h>
+#include <Common/MemoryWorker.h>
 #include <Common/filesystemHelpers.h>
 
 using namespace DB;
@@ -126,7 +126,7 @@ const std::string EXPECTED[2]
        "\"workingset_restore_anon\": 0, \"workingset_restore_file\": 0, \"zswap\": 0, \"zswapped\": 0, \"zswpin\": 0, \"zswpout\": 0}"};
 
 
-class CgroupsMemoryUsageObserverFixture : public ::testing::TestWithParam<CgroupsMemoryUsageObserver::CgroupsVersion>
+class CgroupsMemoryUsageObserverFixture : public ::testing::TestWithParam<ICgroupsReader::CgroupsVersion>
 {
     void SetUp() override
     {
@@ -138,7 +138,7 @@ class CgroupsMemoryUsageObserverFixture : public ::testing::TestWithParam<Cgroup
         stat_file.write(SAMPLE_FILE[version].data(), SAMPLE_FILE[version].size());
         stat_file.sync();
 
-        if (GetParam() == CgroupsMemoryUsageObserver::CgroupsVersion::V2)
+        if (GetParam() == ICgroupsReader::CgroupsVersion::V2)
         {
             auto current_file = WriteBufferFromFile(tmp_dir + "/memory.current");
             current_file.write("29645422592", 11);
@@ -154,18 +154,18 @@ protected:
 TEST_P(CgroupsMemoryUsageObserverFixture, ReadMemoryUsageTest)
 {
     const auto version = GetParam();
-    auto reader = createCgroupsReader(version, tmp_dir);
+    auto reader = ICgroupsReader::createCgroupsReader(version, tmp_dir);
     ASSERT_EQ(
         reader->readMemoryUsage(),
-        version == CgroupsMemoryUsageObserver::CgroupsVersion::V1 ? /* rss from memory.stat */ 2232029184
-                                                                  : /* value from memory.current - inactive_file */ 20952338432);
+        version == ICgroupsReader::CgroupsVersion::V1 ? /* rss from memory.stat */ 2232029184
+                                                                  : /* anon from memory.stat */ 10429399040);
 }
 
 
 TEST_P(CgroupsMemoryUsageObserverFixture, DumpAllStatsTest)
 {
     const auto version = GetParam();
-    auto reader = createCgroupsReader(version, tmp_dir);
+    auto reader = ICgroupsReader::createCgroupsReader(version, tmp_dir);
     ASSERT_EQ(reader->dumpAllStats(), EXPECTED[static_cast<uint8_t>(version)]);
 }
 
@@ -173,6 +173,6 @@ TEST_P(CgroupsMemoryUsageObserverFixture, DumpAllStatsTest)
 INSTANTIATE_TEST_SUITE_P(
     CgroupsMemoryUsageObserverTests,
     CgroupsMemoryUsageObserverFixture,
-    ::testing::Values(CgroupsMemoryUsageObserver::CgroupsVersion::V1, CgroupsMemoryUsageObserver::CgroupsVersion::V2));
+    ::testing::Values(ICgroupsReader::CgroupsVersion::V1, ICgroupsReader::CgroupsVersion::V2));
 
 #endif

src/Compression/CompressedReadBufferBase.cpp

@@ -39,7 +39,7 @@ using Checksum = CityHash_v1_0_2::uint128;
 
 
 /// Validate checksum of data, and if it mismatches, find out possible reason and throw exception.
-static void validateChecksum(char * data, size_t size, const Checksum expected_checksum)
+static void validateChecksum(char * data, size_t size, const Checksum expected_checksum, bool external_data)
 {
     auto calculated_checksum = CityHash_v1_0_2::CityHash128(data, size);
     if (expected_checksum == calculated_checksum)
@@ -64,6 +64,8 @@ static void validateChecksum(char * data, size_t size, const Checksum expected_c
                                             "this can be caused by disk bit rot. This exception protects ClickHouse "
                                             "from data corruption due to hardware failures.";
 
+    int error_code = external_data ? ErrorCodes::CANNOT_DECOMPRESS : ErrorCodes::CHECKSUM_DOESNT_MATCH;
+
     auto flip_bit = [](char * buf, size_t pos)
     {
         buf[pos / 8] ^= 1 << pos % 8;
@@ -87,7 +89,7 @@ static void validateChecksum(char * data, size_t size, const Checksum expected_c
             {
                 message << ". The mismatch is caused by single bit flip in data block at byte " << (bit_pos / 8) << ", bit " << (bit_pos % 8) << ". "
                     << message_hardware_failure;
-                throw Exception::createDeprecated(message.str(), ErrorCodes::CHECKSUM_DOESNT_MATCH);
+                throw Exception::createDeprecated(message.str(), error_code);
             }
 
             flip_bit(tmp_data, bit_pos);    /// Restore
@@ -102,10 +104,10 @@ static void validateChecksum(char * data, size_t size, const Checksum expected_c
     {
         message << ". The mismatch is caused by single bit flip in checksum. "
             << message_hardware_failure;
-        throw Exception::createDeprecated(message.str(), ErrorCodes::CHECKSUM_DOESNT_MATCH);
+        throw Exception::createDeprecated(message.str(), error_code);
     }
 
-    throw Exception::createDeprecated(message.str(), ErrorCodes::CHECKSUM_DOESNT_MATCH);
+    throw Exception::createDeprecated(message.str(), error_code);
 }
 
 static void readHeaderAndGetCodecAndSize(
@@ -151,7 +153,7 @@ static void readHeaderAndGetCodecAndSize(
                         "Most likely corrupted data.", size_compressed_without_checksum);
 
     if (size_compressed_without_checksum < header_size)
-        throw Exception(ErrorCodes::CORRUPTED_DATA, "Can't decompress data: "
+        throw Exception(external_data ? ErrorCodes::CANNOT_DECOMPRESS : ErrorCodes::CORRUPTED_DATA, "Can't decompress data: "
             "the compressed data size ({}, this should include header size) is less than the header size ({})",
             size_compressed_without_checksum, static_cast<size_t>(header_size));
 }
@@ -202,7 +204,7 @@ size_t CompressedReadBufferBase::readCompressedData(size_t & size_decompressed,
         readBinaryLittleEndian(checksum.low64, checksum_in);
         readBinaryLittleEndian(checksum.high64, checksum_in);
 
-        validateChecksum(compressed_buffer, size_compressed_without_checksum, checksum);
+        validateChecksum(compressed_buffer, size_compressed_without_checksum, checksum, external_data);
     }
 
     ProfileEvents::increment(ProfileEvents::ReadCompressedBytes, size_compressed_without_checksum + sizeof(Checksum));
@@ -247,7 +249,7 @@ size_t CompressedReadBufferBase::readCompressedDataBlockForAsynchronous(size_t &
             readBinaryLittleEndian(checksum.low64, checksum_in);
             readBinaryLittleEndian(checksum.high64, checksum_in);
 
-            validateChecksum(compressed_buffer, size_compressed_without_checksum, checksum);
+            validateChecksum(compressed_buffer, size_compressed_without_checksum, checksum, external_data);
         }
 
         ProfileEvents::increment(ProfileEvents::ReadCompressedBytes, size_compressed_without_checksum + sizeof(Checksum));
@@ -307,7 +309,7 @@ void CompressedReadBufferBase::decompress(BufferBase::Buffer & to, size_t size_d
 
         UInt8 header_size = ICompressionCodec::getHeaderSize();
         if (size_compressed_without_checksum < header_size)
-            throw Exception(ErrorCodes::CORRUPTED_DATA,
+            throw Exception(external_data ? ErrorCodes::CANNOT_DECOMPRESS : ErrorCodes::CORRUPTED_DATA,
                 "Can't decompress data: the compressed data size ({}, this should include header size) is less than the header size ({})",
                     size_compressed_without_checksum, static_cast<size_t>(header_size));
 

src/Compression/CompressedWriteBuffer.cpp

@@ -55,10 +55,29 @@ void CompressedWriteBuffer::nextImpl()
 
         out.write(compressed_buffer.data(), compressed_size);
     }
+
+    /// Increase buffer size for next data if adaptive buffer size is used and nextImpl was called because of end of buffer.
+    if (!available() && use_adaptive_buffer_size && memory.size() < adaptive_buffer_max_size)
+    {
+        memory.resize(std::min(memory.size() * 2, adaptive_buffer_max_size));
+        BufferBase::set(memory.data(), memory.size(), 0);
+    }
 }
 
-CompressedWriteBuffer::CompressedWriteBuffer(WriteBuffer & out_, CompressionCodecPtr codec_, size_t buf_size)
-    : BufferWithOwnMemory<WriteBuffer>(buf_size), out(out_), codec(std::move(codec_))
+void CompressedWriteBuffer::finalizeImpl()
+{
+    /// Don't try to resize buffer in nextImpl.
+    use_adaptive_buffer_size = false;
+    next();
+}
+
+CompressedWriteBuffer::CompressedWriteBuffer(
+    WriteBuffer & out_, CompressionCodecPtr codec_, size_t buf_size, bool use_adaptive_buffer_size_, size_t adaptive_buffer_initial_size)
+    : BufferWithOwnMemory<WriteBuffer>(use_adaptive_buffer_size_ ? adaptive_buffer_initial_size : buf_size)
+    , out(out_)
+    , codec(std::move(codec_))
+    , use_adaptive_buffer_size(use_adaptive_buffer_size_)
+    , adaptive_buffer_max_size(buf_size)
 {
 }
 

src/Compression/CompressedWriteBuffer.h

@@ -19,7 +19,9 @@ public:
     explicit CompressedWriteBuffer(
         WriteBuffer & out_,
         CompressionCodecPtr codec_ = CompressionCodecFactory::instance().getDefaultCodec(),
-        size_t buf_size = DBMS_DEFAULT_BUFFER_SIZE);
+        size_t buf_size = DBMS_DEFAULT_BUFFER_SIZE,
+        bool use_adaptive_buffer_size_ = false,
+        size_t adaptive_buffer_initial_size = DBMS_DEFAULT_INITIAL_ADAPTIVE_BUFFER_SIZE);
 
     ~CompressedWriteBuffer() override;
 
@@ -45,10 +47,17 @@ public:
 
 private:
     void nextImpl() override;
+    void finalizeImpl() override;
 
     WriteBuffer & out;
     CompressionCodecPtr codec;
 
+    /// If true, the size of internal buffer will be exponentially increased up to
+    /// adaptive_buffer_max_size after each nextImpl call. It can be used to avoid
+    /// large buffer allocation when actual size of written data is small.
+    bool use_adaptive_buffer_size;
+    size_t adaptive_buffer_max_size;
+
     PODArray<char> compressed_buffer;
 };
 

src/Coordination/KeeperAsynchronousMetrics.cpp

@@ -114,8 +114,13 @@ void updateKeeperInformation(KeeperDispatcher & keeper_dispatcher, AsynchronousM
 }
 
 KeeperAsynchronousMetrics::KeeperAsynchronousMetrics(
-    ContextPtr context_, unsigned update_period_seconds, const ProtocolServerMetricsFunc & protocol_server_metrics_func_)
-    : AsynchronousMetrics(update_period_seconds, protocol_server_metrics_func_), context(std::move(context_))
+    ContextPtr context_,
+    unsigned update_period_seconds,
+    const ProtocolServerMetricsFunc & protocol_server_metrics_func_,
+    bool update_jemalloc_epoch_,
+    bool update_rss_)
+    : AsynchronousMetrics(update_period_seconds, protocol_server_metrics_func_, update_jemalloc_epoch_, update_rss_)
+    , context(std::move(context_))
 {
 }
 

src/Coordination/KeeperAsynchronousMetrics.h

@@ -13,9 +13,13 @@ class KeeperAsynchronousMetrics : public AsynchronousMetrics
 {
 public:
     KeeperAsynchronousMetrics(
-        ContextPtr context_, unsigned update_period_seconds, const ProtocolServerMetricsFunc & protocol_server_metrics_func_);
-    ~KeeperAsynchronousMetrics() override;
+        ContextPtr context_,
+        unsigned update_period_seconds,
+        const ProtocolServerMetricsFunc & protocol_server_metrics_func_,
+        bool update_jemalloc_epoch_,
+        bool update_rss_);
 
+    ~KeeperAsynchronousMetrics() override;
 private:
     ContextPtr context;
 

src/Coordination/KeeperConstants.h

@@ -11,6 +11,7 @@ enum class KeeperApiVersion : uint8_t
     WITH_FILTERED_LIST,
     WITH_MULTI_READ,
     WITH_CHECK_NOT_EXISTS,
+    WITH_REMOVE_RECURSIVE,
 };
 
 const String keeper_system_path = "/keeper";

src/Coordination/KeeperDispatcher.cpp

@@ -91,6 +91,12 @@ bool checkIfRequestIncreaseMem(const Coordination::ZooKeeperRequestPtr & request
                     memory_delta -= remove_req.bytesSize();
                     break;
                 }
+                case Coordination::OpNum::RemoveRecursive:
+                {
+                    Coordination::ZooKeeperRemoveRecursiveRequest & remove_req = dynamic_cast<Coordination::ZooKeeperRemoveRecursiveRequest &>(*sub_zk_request);
+                    memory_delta -= remove_req.bytesSize();
+                    break;
+                }
                 default:
                     break;
             }
@@ -148,7 +154,14 @@ void KeeperDispatcher::requestThread()
                 Int64 mem_soft_limit = keeper_context->getKeeperMemorySoftLimit();
                 if (configuration_and_settings->standalone_keeper && isExceedingMemorySoftLimit() && checkIfRequestIncreaseMem(request.request))
                 {
-                    LOG_WARNING(log, "Processing requests refused because of max_memory_usage_soft_limit {}, the total used memory is {}, request type is {}", ReadableSize(mem_soft_limit), ReadableSize(total_memory_tracker.get()), request.request->getOpNum());
+                    LOG_WARNING(
+                        log,
+                        "Processing requests refused because of max_memory_usage_soft_limit {}, the total allocated memory is {}, RSS is {}, request type "
+                        "is {}",
+                        ReadableSize(mem_soft_limit),
+                        ReadableSize(total_memory_tracker.get()),
+                        ReadableSize(total_memory_tracker.getRSS()),
+                        request.request->getOpNum());
                     addErrorResponses({request}, Coordination::Error::ZCONNECTIONLOSS);
                     continue;
                 }

src/Coordination/KeeperFeatureFlags.h

@@ -12,6 +12,7 @@ enum class KeeperFeatureFlag : size_t
     MULTI_READ,
     CHECK_NOT_EXISTS,
     CREATE_IF_NOT_EXISTS,
+    REMOVE_RECURSIVE,
 };
 
 class KeeperFeatureFlags

src/Coordination/KeeperServer.cpp

@@ -35,6 +35,7 @@
 #    include <Poco/Net/Context.h>
 #    include <Poco/Net/SSLManager.h>
 #    include <Poco/Net/Utility.h>
+#    include <Poco/StringTokenizer.h>
 #endif
 
 #include <chrono>
@@ -107,15 +108,28 @@ void setSSLParams(nuraft::asio_service::options & asio_opts)
     params.loadDefaultCAs = config.getBool("openSSL.server.loadDefaultCAFile", false);
     params.verificationMode = Poco::Net::Utility::convertVerificationMode(config.getString("openSSL.server.verificationMode", "none"));
 
-    asio_opts.ssl_context_provider_server_ = [ctx_params = params, certificate_data]
+    std::string disabled_protocols_list = config.getString("openSSL.server.disableProtocols", "");
+    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)
     {
-        Poco::Net::Context context(Poco::Net::Context::Usage::TLSV1_2_SERVER_USE, ctx_params);
+        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();
-        uint64_t options = 0;
-        options |= SSL_OP_ALL;
-        options |= SSL_OP_NO_SSLv2;
-        options |= SSL_OP_SINGLE_DH_USE;
-        SSL_CTX_set_options(ssl_ctx, options);
         SSL_CTX_set_cert_cb(ssl_ctx, callSetCertificate, reinterpret_cast<void *>(certificate_data.get()));
         return ssl_ctx;
     };
@@ -646,7 +660,7 @@ bool KeeperServer::isLeaderAlive() const
 bool KeeperServer::isExceedingMemorySoftLimit() const
 {
     Int64 mem_soft_limit = keeper_context->getKeeperMemorySoftLimit();
-    return mem_soft_limit > 0 && total_memory_tracker.get() >= mem_soft_limit;
+    return mem_soft_limit > 0 && std::max(total_memory_tracker.get(), total_memory_tracker.getRSS()) >= mem_soft_limit;
 }
 
 /// TODO test whether taking failed peer in count

src/Coordination/KeeperStorage.cpp

@@ -832,6 +832,15 @@ std::shared_ptr<typename Container::Node> KeeperStorage<Container>::UncommittedS
     return tryGetNodeFromStorage(path);
 }
 
+template<typename Container>
+const typename Container::Node * KeeperStorage<Container>::UncommittedState::getActualNodeView(StringRef path, const Node & storage_node) const
+{
+    if (auto node_it = nodes.find(path.toView()); node_it != nodes.end())
+        return node_it->second.node.get();
+
+    return &storage_node;
+}
+
 template<typename Container>
 Coordination::ACLs KeeperStorage<Container>::UncommittedState::getACLs(StringRef path) const
 {
@@ -1124,7 +1133,7 @@ struct KeeperStorageRequestProcessor
     }
 
     virtual KeeperStorageBase::ResponsesForSessions
-    processWatches(KeeperStorageBase::Watches & /*watches*/, KeeperStorageBase::Watches & /*list_watches*/) const
+    processWatches(const Storage & /*storage*/, int64_t /*zxid*/, KeeperStorageBase::Watches & /*watches*/, KeeperStorageBase::Watches & /*list_watches*/) const
     {
         return {};
     }
@@ -1241,7 +1250,7 @@ struct KeeperStorageCreateRequestProcessor final : public KeeperStorageRequestPr
     using KeeperStorageRequestProcessor<Storage>::KeeperStorageRequestProcessor;
 
     KeeperStorageBase::ResponsesForSessions
-    processWatches(KeeperStorageBase::Watches & watches, KeeperStorageBase::Watches & list_watches) const override
+    processWatches(const Storage & /*storage*/, int64_t /*zxid*/, KeeperStorageBase::Watches & watches, KeeperStorageBase::Watches & list_watches) const override
     {
         return processWatchesImpl(this->zk_request->getPath(), watches, list_watches, Coordination::Event::CREATED);
     }
@@ -1462,16 +1471,41 @@ struct KeeperStorageGetRequestProcessor final : public KeeperStorageRequestProce
     }
 };
 
+namespace
+{
+
+template <typename Storage>
+void addUpdateParentPzxidDelta(Storage & storage, std::vector<typename Storage::Delta> & deltas, int64_t zxid, StringRef path)
+{
+    auto parent_path = parentNodePath(path);
+    if (!storage.uncommitted_state.getNode(parent_path))
+        return;
+
+    deltas.emplace_back(
+        std::string{parent_path},
+        zxid,
+        typename Storage::UpdateNodeDelta
+        {
+            [zxid](Storage::Node & parent)
+            {
+                parent.pzxid = std::max(parent.pzxid, zxid);
+            }
+        }
+    );
+}
+
+}
+
 template<typename Storage>
 struct KeeperStorageRemoveRequestProcessor final : public KeeperStorageRequestProcessor<Storage>
 {
+    using KeeperStorageRequestProcessor<Storage>::KeeperStorageRequestProcessor;
+
     bool checkAuth(Storage & storage, int64_t session_id, bool is_local) const override
     {
         return storage.checkACL(parentNodePath(this->zk_request->getPath()), Coordination::ACL::Delete, session_id, is_local);
     }
 
-    using KeeperStorageRequestProcessor<Storage>::KeeperStorageRequestProcessor;
-
     std::vector<typename Storage::Delta>
     preprocess(Storage & storage, int64_t zxid, int64_t /*session_id*/, int64_t /*time*/, uint64_t & digest, const KeeperContext & keeper_context) const override
     {
@@ -1488,31 +1522,12 @@ struct KeeperStorageRemoveRequestProcessor final : public KeeperStorageRequestPr
             return {typename Storage::Delta{zxid, Coordination::Error::ZBADARGUMENTS}};
         }
 
-        const auto update_parent_pzxid = [&]()
-        {
-            auto parent_path = parentNodePath(request.path);
-            if (!storage.uncommitted_state.getNode(parent_path))
-                return;
-
-            new_deltas.emplace_back(
-                std::string{parent_path},
-                zxid,
-                typename Storage::UpdateNodeDelta
-                {
-                    [zxid](Storage::Node & parent)
-                    {
-                        parent.pzxid = std::max(parent.pzxid, zxid);
-                    }
-                }
-            );
-        };
-
         auto node = storage.uncommitted_state.getNode(request.path);
 
         if (!node)
         {
             if (request.restored_from_zookeeper_log)
-                update_parent_pzxid();
+                addUpdateParentPzxidDelta(storage, new_deltas, zxid, request.path);
             return {typename Storage::Delta{zxid, Coordination::Error::ZNONODE}};
         }
         else if (request.version != -1 && request.version != node->version)
@@ -1521,7 +1536,7 @@ struct KeeperStorageRemoveRequestProcessor final : public KeeperStorageRequestPr
             return {typename Storage::Delta{zxid, Coordination::Error::ZNOTEMPTY}};
 
         if (request.restored_from_zookeeper_log)
-            update_parent_pzxid();
+            addUpdateParentPzxidDelta(storage, new_deltas, zxid, request.path);
 
         new_deltas.emplace_back(
             std::string{parentNodePath(request.path)},
@@ -1552,12 +1567,318 @@ struct KeeperStorageRemoveRequestProcessor final : public KeeperStorageRequestPr
     }
 
     KeeperStorageBase::ResponsesForSessions
-    processWatches(KeeperStorageBase::Watches & watches, KeeperStorageBase::Watches & list_watches) const override
+    processWatches(const Storage & /*storage*/, int64_t /*zxid*/, KeeperStorageBase::Watches & watches, KeeperStorageBase::Watches & list_watches) const override
     {
         return processWatchesImpl(this->zk_request->getPath(), watches, list_watches, Coordination::Event::DELETED);
     }
 };
 
+template<typename Storage>
+struct KeeperStorageRemoveRecursiveRequestProcessor final : public KeeperStorageRequestProcessor<Storage>
+{
+    using KeeperStorageRequestProcessor<Storage>::KeeperStorageRequestProcessor;
+
+    bool checkAuth(Storage & storage, int64_t session_id, bool is_local) const override
+    {
+        return storage.checkACL(parentNodePath(this->zk_request->getPath()), Coordination::ACL::Delete, session_id, is_local);
+    }
+
+    std::vector<typename Storage::Delta>
+    preprocess(Storage & storage, int64_t zxid, int64_t session_id, int64_t /*time*/, uint64_t & digest, const KeeperContext & keeper_context) const override
+    {
+        ProfileEvents::increment(ProfileEvents::KeeperRemoveRequest);
+        Coordination::ZooKeeperRemoveRecursiveRequest & request = dynamic_cast<Coordination::ZooKeeperRemoveRecursiveRequest &>(*this->zk_request);
+
+        std::vector<typename Storage::Delta> new_deltas;
+
+        if (Coordination::matchPath(request.path, keeper_system_path) != Coordination::PathMatchResult::NOT_MATCH)
+        {
+            auto error_msg = fmt::format("Trying to delete an internal Keeper path ({}) which is not allowed", request.path);
+
+            handleSystemNodeModification(keeper_context, error_msg);
+            return {typename Storage::Delta{zxid, Coordination::Error::ZBADARGUMENTS}};
+        }
+
+        auto node = storage.uncommitted_state.getNode(request.path);
+
+        if (!node)
+        {
+            if (request.restored_from_zookeeper_log)
+                addUpdateParentPzxidDelta(storage, new_deltas, zxid, request.path);
+
+            return {typename Storage::Delta{zxid, Coordination::Error::ZNONODE}};
+        }
+
+        ToDeleteTreeCollector collector(storage, zxid, session_id, request.remove_nodes_limit);
+        auto collect_status = collector.collect(request.path, *node);
+
+        if (collect_status == ToDeleteTreeCollector::CollectStatus::NoAuth)
+            return {typename Storage::Delta{zxid, Coordination::Error::ZNOAUTH}};
+
+        if (collect_status == ToDeleteTreeCollector::CollectStatus::LimitExceeded)
+            return {typename Storage::Delta{zxid, Coordination::Error::ZNOTEMPTY}};
+
+        if (request.restored_from_zookeeper_log)
+            addUpdateParentPzxidDelta(storage, new_deltas, zxid, request.path);
+
+        auto delete_deltas = collector.extractDeltas();
+
+        for (const auto & delta : delete_deltas)
+        {
+            const auto * remove_delta = std::get_if<typename Storage::RemoveNodeDelta>(&delta.operation);
+            if (remove_delta && remove_delta->ephemeral_owner)
+                storage.unregisterEphemeralPath(remove_delta->ephemeral_owner, delta.path);
+        }
+
+        new_deltas.insert(new_deltas.end(), std::make_move_iterator(delete_deltas.begin()), std::make_move_iterator(delete_deltas.end()));
+
+        digest = storage.calculateNodesDigest(digest, new_deltas);
+
+        return new_deltas;
+    }
+
+    Coordination::ZooKeeperResponsePtr process(Storage & storage, int64_t zxid) const override
+    {
+        Coordination::ZooKeeperResponsePtr response_ptr = this->zk_request->makeResponse();
+        Coordination::ZooKeeperRemoveRecursiveResponse & response = dynamic_cast<Coordination::ZooKeeperRemoveRecursiveResponse &>(*response_ptr);
+
+        response.error = storage.commit(zxid);
+        return response_ptr;
+    }
+
+    KeeperStorageBase::ResponsesForSessions
+    processWatches(const Storage & storage, int64_t zxid, KeeperStorageBase::Watches & watches, KeeperStorageBase::Watches & list_watches) const override
+    {
+        /// need to iterate over zxid deltas and update watches for deleted tree.
+        const auto & deltas = storage.uncommitted_state.deltas;
+
+        KeeperStorageBase::ResponsesForSessions responses;
+        for (auto it = deltas.rbegin(); it != deltas.rend() && it->zxid == zxid; ++it)
+        {
+            const auto * remove_delta = std::get_if<typename Storage::RemoveNodeDelta>(&it->operation);
+            if (remove_delta)
+            {
+                auto new_responses = processWatchesImpl(it->path, watches, list_watches, Coordination::Event::DELETED);
+                responses.insert(responses.end(), std::make_move_iterator(new_responses.begin()), std::make_move_iterator(new_responses.end()));
+            }
+        }
+
+        return responses;
+    }
+
+private:
+    using SNode = typename Storage::Node;
+
+    class ToDeleteTreeCollector
+    {
+        Storage & storage;
+        int64_t zxid;
+        int64_t session_id;
+        uint32_t limit;
+
+        uint32_t max_level = 0;
+        uint32_t nodes_observed = 1;  /// root node
+        std::unordered_map<uint32_t, std::vector<typename Storage::Delta>> by_level_deltas;
+
+        struct Step
+        {
+            String path;
+            std::variant<SNode, const SNode *> node;
+            uint32_t level;
+        };
+
+        enum class CollectStatus
+        {
+            Ok,
+            NoAuth,
+            LimitExceeded,
+        };
+
+        friend struct KeeperStorageRemoveRecursiveRequestProcessor;
+
+    public:
+        ToDeleteTreeCollector(Storage & storage_, int64_t zxid_, int64_t session_id_, uint32_t limit_)
+            : storage(storage_)
+            , zxid(zxid_)
+            , session_id(session_id_)
+            , limit(limit_)
+        {
+        }
+
+        CollectStatus collect(StringRef root_path, const SNode & root_node)
+        {
+            std::deque<Step> steps;
+
+            if (checkLimits(&root_node))
+                return CollectStatus::LimitExceeded;
+
+            steps.push_back(Step{root_path.toString(), &root_node, 0});
+
+            while (!steps.empty())
+            {
+                Step step = std::move(steps.front());
+                steps.pop_front();
+
+                StringRef path = step.path;
+                uint32_t level = step.level;
+                const SNode * node_ptr = nullptr;
+
+                if (auto * rdb = std::get_if<SNode>(&step.node))
+                    node_ptr = rdb;
+                else
+                    node_ptr = std::get<const SNode *>(step.node);
+
+                chassert(!path.empty());
+                chassert(node_ptr != nullptr);
+
+                const auto & node = *node_ptr;
+                auto actual_node_ptr = storage.uncommitted_state.getActualNodeView(path, node);
+                chassert(actual_node_ptr != nullptr); /// explicitly check that node is not deleted
+
+                if (actual_node_ptr->numChildren() > 0 && !storage.checkACL(path, Coordination::ACL::Delete, session_id, /*is_local=*/false))
+                    return CollectStatus::NoAuth;
+
+                if (auto status = visitRocksDBNode(steps, path, level); status != CollectStatus::Ok)
+                    return status;
+
+                if (auto status = visitMemNode(steps, path, level); status != CollectStatus::Ok)
+                    return status;
+
+                if (auto status = visitRootAndUncommitted(steps, path, node, level); status != CollectStatus::Ok)
+                    return status;
+            }
+
+            return CollectStatus::Ok;
+        }
+
+        std::vector<typename Storage::Delta> extractDeltas()
+        {
+            std::vector<typename Storage::Delta> deltas;
+
+            for (ssize_t level = max_level; level >= 0; --level)
+            {
+                auto & level_deltas = by_level_deltas[static_cast<uint32_t>(level)];
+                deltas.insert(deltas.end(), std::make_move_iterator(level_deltas.begin()), std::make_move_iterator(level_deltas.end()));
+            }
+
+            return std::move(deltas);
+        }
+
+    private:
+        CollectStatus visitRocksDBNode(std::deque<Step> & steps, StringRef root_path, uint32_t level)
+        {
+            if constexpr (Storage::use_rocksdb)
+            {
+                std::filesystem::path root_fs_path(root_path.toString());
+                auto children = storage.container.getChildren(root_path.toString());
+
+                for (auto && [child_name, child_node] : children)
+                {
+                    auto child_path = (root_fs_path / child_name).generic_string();
+                    const auto actual_child_node_ptr = storage.uncommitted_state.getActualNodeView(child_path, child_node);
+
+                    if (actual_child_node_ptr == nullptr) /// node was deleted in previous step of multi transaction
+                        continue;
+
+                    if (checkLimits(actual_child_node_ptr))
+                        return CollectStatus::LimitExceeded;
+
+                    steps.push_back(Step{std::move(child_path), std::move(child_node), level + 1});
+                }
+            }
+
+            return CollectStatus::Ok;
+        }
+
+        CollectStatus visitMemNode(std::deque<Step> & steps, StringRef root_path, uint32_t level)
+        {
+            if constexpr (!Storage::use_rocksdb)
+            {
+                auto node_it = storage.container.find(root_path);
+                if (node_it == storage.container.end())
+                    return CollectStatus::Ok;
+
+                std::filesystem::path root_fs_path(root_path.toString());
+                const auto & children = node_it->value.getChildren();
+
+                for (const auto & child_name : children)
+                {
+                    auto child_path = (root_fs_path / child_name.toView()).generic_string();
+
+                    auto child_it = storage.container.find(child_path);
+                    chassert(child_it != storage.container.end());
+                    const auto & child_node = child_it->value;
+
+                    const auto actual_child_node_ptr = storage.uncommitted_state.getActualNodeView(child_path, child_node);
+
+                    if (actual_child_node_ptr == nullptr) /// node was deleted in previous step of multi transaction
+                        continue;
+
+                    if (checkLimits(actual_child_node_ptr))
+                        return CollectStatus::LimitExceeded;
+
+                    steps.push_back(Step{std::move(child_path), &child_node, level + 1});
+                }
+            }
+
+            return CollectStatus::Ok;
+        }
+
+        CollectStatus visitRootAndUncommitted(std::deque<Step> & steps, StringRef root_path, const SNode & root_node, uint32_t level)
+        {
+            const auto & nodes = storage.uncommitted_state.nodes;
+
+            /// nodes are sorted by paths with level locality
+            auto it = nodes.upper_bound(root_path.toString() + "/");
+
+            for (; it != nodes.end() && parentNodePath(it->first) == root_path; ++it)
+            {
+                const auto actual_child_node_ptr = it->second.node.get();
+
+                if (actual_child_node_ptr == nullptr) /// node was deleted in previous step of multi transaction
+                    continue;
+
+                if (checkLimits(actual_child_node_ptr))
+                    return CollectStatus::LimitExceeded;
+
+                const String & child_path = it->first;
+                const SNode & child_node = *it->second.node;
+
+                steps.push_back(Step{child_path, &child_node, level + 1});
+            }
+
+            addDelta(root_path, root_node, level);
+
+            return CollectStatus::Ok;
+        }
+
+        void addDelta(StringRef root_path, const SNode & root_node, uint32_t level)
+        {
+            max_level = std::max(max_level, level);
+
+            by_level_deltas[level].emplace_back(
+                parentNodePath(root_path).toString(),
+                zxid,
+                typename Storage::UpdateNodeDelta{
+                    [](SNode & parent)
+                    {
+                        ++parent.cversion;
+                        parent.decreaseNumChildren();
+                    }
+                });
+
+            by_level_deltas[level].emplace_back(root_path.toString(), zxid, typename Storage::RemoveNodeDelta{root_node.version, root_node.ephemeralOwner()});
+        }
+
+        bool checkLimits(const SNode * node)
+        {
+            chassert(node != nullptr);
+            nodes_observed += node->numChildren();
+            return nodes_observed > limit;
+        }
+    };
+};
+
 template<typename Storage>
 struct KeeperStorageExistsRequestProcessor final : public KeeperStorageRequestProcessor<Storage>
 {
@@ -1709,7 +2030,7 @@ struct KeeperStorageSetRequestProcessor final : public KeeperStorageRequestProce
     }
 
     KeeperStorageBase::ResponsesForSessions
-    processWatches(typename Storage::Watches & watches, typename Storage::Watches & list_watches) const override
+    processWatches(const Storage & /*storage*/, int64_t /*zxid*/, typename Storage::Watches & watches, typename Storage::Watches & list_watches) const override
     {
         return processWatchesImpl(this->zk_request->getPath(), watches, list_watches, Coordination::Event::CHANGED);
     }
@@ -2131,6 +2452,10 @@ struct KeeperStorageMultiRequestProcessor final : public KeeperStorageRequestPro
                     check_operation_type(OperationType::Write);
                     concrete_requests.push_back(std::make_shared<KeeperStorageRemoveRequestProcessor<Storage>>(sub_zk_request));
                     break;
+                case Coordination::OpNum::RemoveRecursive:
+                    check_operation_type(OperationType::Write);
+                    concrete_requests.push_back(std::make_shared<KeeperStorageRemoveRecursiveRequestProcessor<Storage>>(sub_zk_request));
+                    break;
                 case Coordination::OpNum::Set:
                     check_operation_type(OperationType::Write);
                     concrete_requests.push_back(std::make_shared<KeeperStorageSetRequestProcessor<Storage>>(sub_zk_request));
@@ -2250,12 +2575,12 @@ struct KeeperStorageMultiRequestProcessor final : public KeeperStorageRequestPro
     }
 
     KeeperStorageBase::ResponsesForSessions
-    processWatches(typename Storage::Watches & watches, typename Storage::Watches & list_watches) const override
+    processWatches(const Storage & storage, int64_t zxid, typename Storage::Watches & watches, typename Storage::Watches & list_watches) const override
     {
         typename Storage::ResponsesForSessions result;
         for (const auto & generic_request : concrete_requests)
         {
-            auto responses = generic_request->processWatches(watches, list_watches);
+            auto responses = generic_request->processWatches(storage, zxid, watches, list_watches);
             result.insert(result.end(), responses.begin(), responses.end());
         }
         return result;
@@ -2400,6 +2725,7 @@ KeeperStorageRequestProcessorsFactory<Storage>::KeeperStorageRequestProcessorsFa
     registerKeeperRequestProcessor<Coordination::OpNum::SetACL, KeeperStorageSetACLRequestProcessor<Storage>>(*this);
     registerKeeperRequestProcessor<Coordination::OpNum::GetACL, KeeperStorageGetACLRequestProcessor<Storage>>(*this);
     registerKeeperRequestProcessor<Coordination::OpNum::CheckNotExists, KeeperStorageCheckRequestProcessor<Storage>>(*this);
+    registerKeeperRequestProcessor<Coordination::OpNum::RemoveRecursive, KeeperStorageRemoveRecursiveRequestProcessor<Storage>>(*this);
 }
 
 
@@ -2718,7 +3044,7 @@ KeeperStorage<Container>::ResponsesForSessions KeeperStorage<Container>::process
         /// If this requests processed successfully we need to check watches
         if (response->error == Coordination::Error::ZOK)
         {
-            auto watch_responses = request_processor->processWatches(watches, list_watches);
+            auto watch_responses = request_processor->processWatches(*this, zxid, watches, list_watches);
             results.insert(results.end(), watch_responses.begin(), watch_responses.end());
         }
 

src/Coordination/KeeperStorage.h

@@ -566,6 +566,7 @@ public:
         void rollback(int64_t rollback_zxid);
 
         std::shared_ptr<Node> getNode(StringRef path) const;
+        const Node * getActualNodeView(StringRef path, const Node & storage_node) const;
         Coordination::ACLs getACLs(StringRef path) const;
 
         void applyDelta(const Delta & delta);
@@ -609,7 +610,18 @@ public:
             using is_transparent = void; // required to make find() work with different type than key_type
         };
 
-        mutable std::unordered_map<std::string, UncommittedNode, Hash, Equal> nodes;
+        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);
+            }
+        };
+
+        mutable std::map<std::string, UncommittedNode, PathCmp> nodes;
         std::unordered_map<std::string, std::list<const Delta *>, Hash, Equal> deltas_for_path;
 
         std::list<Delta> deltas;

src/Coordination/tests/gtest_coordination.cpp

@@ -3113,6 +3113,8 @@ TYPED_TEST(CoordinationTest, TestFeatureFlags)
     ASSERT_TRUE(feature_flags.isEnabled(KeeperFeatureFlag::FILTERED_LIST));
     ASSERT_TRUE(feature_flags.isEnabled(KeeperFeatureFlag::MULTI_READ));
     ASSERT_FALSE(feature_flags.isEnabled(KeeperFeatureFlag::CHECK_NOT_EXISTS));
+    ASSERT_FALSE(feature_flags.isEnabled(KeeperFeatureFlag::CREATE_IF_NOT_EXISTS));
+    ASSERT_FALSE(feature_flags.isEnabled(KeeperFeatureFlag::REMOVE_RECURSIVE));
 }
 
 TYPED_TEST(CoordinationTest, TestSystemNodeModify)
@@ -3374,6 +3376,474 @@ TYPED_TEST(CoordinationTest, TestReapplyingDeltas)
     ASSERT_TRUE(children1_set == children2_set);
 }
 
+TYPED_TEST(CoordinationTest, TestRemoveRecursiveRequest)
+{
+    using namespace DB;
+    using namespace Coordination;
+
+    using Storage = typename TestFixture::Storage;
+
+    ChangelogDirTest rocks("./rocksdb");
+    this->setRocksDBDirectory("./rocksdb");
+
+    Storage storage{500, "", this->keeper_context};
+
+    int32_t zxid = 0;
+
+    const auto create = [&](const String & path, int create_mode)
+    {
+        int new_zxid = ++zxid;
+
+        const auto create_request = std::make_shared<ZooKeeperCreateRequest>();
+        create_request->path = path;
+        create_request->is_ephemeral = create_mode == zkutil::CreateMode::Ephemeral || create_mode == zkutil::CreateMode::EphemeralSequential;
+        create_request->is_sequential = create_mode == zkutil::CreateMode::PersistentSequential || create_mode == zkutil::CreateMode::EphemeralSequential;
+
+        storage.preprocessRequest(create_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(create_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZOK) << "Failed to create " << path;
+    };
+
+    const auto remove = [&](const String & path, int32_t version = -1)
+    {
+        int new_zxid = ++zxid;
+
+        auto remove_request = std::make_shared<ZooKeeperRemoveRequest>();
+        remove_request->path = path;
+        remove_request->version = version;
+
+        storage.preprocessRequest(remove_request, 1, 0, new_zxid);
+        return storage.processRequest(remove_request, 1, new_zxid);
+    };
+
+    const auto remove_recursive = [&](const String & path, uint32_t remove_nodes_limit = 1)
+    {
+        int new_zxid = ++zxid;
+
+        auto remove_request = std::make_shared<ZooKeeperRemoveRecursiveRequest>();
+        remove_request->path = path;
+        remove_request->remove_nodes_limit = remove_nodes_limit;
+
+        storage.preprocessRequest(remove_request, 1, 0, new_zxid);
+        return storage.processRequest(remove_request, 1, new_zxid);
+    };
+
+    const auto exists = [&](const String & path)
+    {
+        int new_zxid = ++zxid;
+
+        const auto exists_request = std::make_shared<ZooKeeperExistsRequest>();
+        exists_request->path = path;
+
+        storage.preprocessRequest(exists_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(exists_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        return responses[0].response->error == Coordination::Error::ZOK;
+    };
+
+    {
+        SCOPED_TRACE("Single Remove Single Node");
+        create("/T1", zkutil::CreateMode::Persistent);
+
+        auto responses = remove("/T1");
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_EQ(responses[0].response->error, Coordination::Error::ZOK);
+        ASSERT_FALSE(exists("/T1"));
+    }
+
+    {
+        SCOPED_TRACE("Single Remove Tree");
+        create("/T2", zkutil::CreateMode::Persistent);
+        create("/T2/A", zkutil::CreateMode::Persistent);
+
+        auto responses = remove("/T2");
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_EQ(responses[0].response->error, Coordination::Error::ZNOTEMPTY);
+        ASSERT_TRUE(exists("/T2"));
+    }
+
+    {
+        SCOPED_TRACE("Recursive Remove Single Node");
+        create("/T3", zkutil::CreateMode::Persistent);
+
+        auto responses = remove_recursive("/T3", 100);
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_EQ(responses[0].response->error, Coordination::Error::ZOK);
+        ASSERT_FALSE(exists("/T3"));
+    }
+
+    {
+        SCOPED_TRACE("Recursive Remove Tree Small Limit");
+        create("/T5", zkutil::CreateMode::Persistent);
+        create("/T5/A", zkutil::CreateMode::Persistent);
+        create("/T5/B", zkutil::CreateMode::Persistent);
+        create("/T5/A/C", zkutil::CreateMode::Persistent);
+
+        auto responses = remove_recursive("/T5", 2);
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_EQ(responses[0].response->error, Coordination::Error::ZNOTEMPTY);
+        ASSERT_TRUE(exists("/T5"));
+        ASSERT_TRUE(exists("/T5/A"));
+        ASSERT_TRUE(exists("/T5/B"));
+        ASSERT_TRUE(exists("/T5/A/C"));
+    }
+
+    {
+        SCOPED_TRACE("Recursive Remove Tree Big Limit");
+        create("/T6", zkutil::CreateMode::Persistent);
+        create("/T6/A", zkutil::CreateMode::Persistent);
+        create("/T6/B", zkutil::CreateMode::Persistent);
+        create("/T6/A/C", zkutil::CreateMode::Persistent);
+
+        auto responses = remove_recursive("/T6", 4);
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_EQ(responses[0].response->error, Coordination::Error::ZOK);
+        ASSERT_FALSE(exists("/T6"));
+        ASSERT_FALSE(exists("/T6/A"));
+        ASSERT_FALSE(exists("/T6/B"));
+        ASSERT_FALSE(exists("/T6/A/C"));
+    }
+
+    {
+        SCOPED_TRACE("Recursive Remove Ephemeral");
+        create("/T7", zkutil::CreateMode::Ephemeral);
+        ASSERT_EQ(storage.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_FALSE(exists("/T7"));
+    }
+
+    {
+        SCOPED_TRACE("Recursive Remove Tree With Ephemeral");
+        create("/T8", zkutil::CreateMode::Persistent);
+        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);
+
+        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_FALSE(exists("/T8"));
+        ASSERT_FALSE(exists("/T8/A"));
+        ASSERT_FALSE(exists("/T8/B"));
+        ASSERT_FALSE(exists("/T8/A/C"));
+    }
+}
+
+TYPED_TEST(CoordinationTest, TestRemoveRecursiveInMultiRequest)
+{
+    using namespace DB;
+    using namespace Coordination;
+
+    using Storage = typename TestFixture::Storage;
+
+    ChangelogDirTest rocks("./rocksdb");
+    this->setRocksDBDirectory("./rocksdb");
+
+    Storage storage{500, "", this->keeper_context};
+    int zxid = 0;
+
+    auto prepare_create_tree = []()
+    {
+        return Coordination::Requests{
+            zkutil::makeCreateRequest("/A", "A", zkutil::CreateMode::Persistent),
+            zkutil::makeCreateRequest("/A/B", "B", zkutil::CreateMode::Persistent),
+            zkutil::makeCreateRequest("/A/C", "C", zkutil::CreateMode::Ephemeral),
+            zkutil::makeCreateRequest("/A/B/D", "D", zkutil::CreateMode::Ephemeral),
+        };
+    };
+
+    const auto exists = [&](const String & path)
+    {
+        int new_zxid = ++zxid;
+
+        const auto exists_request = std::make_shared<ZooKeeperExistsRequest>();
+        exists_request->path = path;
+
+        storage.preprocessRequest(exists_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(exists_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        return responses[0].response->error == Coordination::Error::ZOK;
+    };
+
+    const auto is_multi_ok = [&](Coordination::ZooKeeperResponsePtr response)
+    {
+        const auto & multi_response = dynamic_cast<Coordination::ZooKeeperMultiResponse &>(*response);
+
+        for (const auto & op_response : multi_response.responses)
+            if (op_response->error != Coordination::Error::ZOK)
+                return false;
+
+        return true;
+    };
+
+    {
+        SCOPED_TRACE("Remove In Multi Tx");
+        int new_zxid = ++zxid;
+        auto ops = prepare_create_tree();
+
+        ops.push_back(zkutil::makeRemoveRequest("/A", -1));
+        const auto request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+
+        storage.preprocessRequest(request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(request, 1, new_zxid);
+        ops.pop_back();
+
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_FALSE(is_multi_ok(responses[0].response));
+    }
+
+    {
+        SCOPED_TRACE("Recursive Remove In Multi Tx");
+        int new_zxid = ++zxid;
+        auto ops = prepare_create_tree();
+
+        ops.push_back(zkutil::makeRemoveRecursiveRequest("/A", 4));
+        const auto request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+
+        storage.preprocessRequest(request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(request, 1, new_zxid);
+        ops.pop_back();
+
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_TRUE(is_multi_ok(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("Recursive Remove With Regular In Multi Tx");
+        int new_zxid = ++zxid;
+        auto ops = prepare_create_tree();
+
+        ops.push_back(zkutil::makeRemoveRequest("/A/C", -1));
+        ops.push_back(zkutil::makeRemoveRecursiveRequest("/A", 3));
+        const auto request = std::make_shared<ZooKeeperMultiRequest>(ops, ACLs{});
+
+        storage.preprocessRequest(request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(request, 1, new_zxid);
+        ops.pop_back();
+        ops.pop_back();
+
+        ASSERT_EQ(responses.size(), 1);
+        ASSERT_TRUE(is_multi_ok(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("Recursive Remove From Committed and Uncommitted states");
+        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));
+        ASSERT_TRUE(exists("/A"));
+        ASSERT_TRUE(exists("/A/C"));
+        ASSERT_TRUE(exists("/A/B"));
+        ASSERT_TRUE(exists("/A/B/D"));
+
+        /// Remove node A/C as a single remove request.
+        /// Remove all other as remove recursive request.
+        /// In this case we should list storage to understand the tree topology
+        /// but ignore already deleted nodes in uncommitted state.
+
+        int remove_zxid = ++zxid;
+        ops = {
+            zkutil::makeRemoveRequest("/A/C", -1),
+            zkutil::makeRemoveRecursiveRequest("/A", 3),
+        };
+        const 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_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"));
+    }
+}
+
+TYPED_TEST(CoordinationTest, TestRemoveRecursiveWatches)
+{
+    using namespace DB;
+    using namespace Coordination;
+
+    using Storage = typename TestFixture::Storage;
+
+    ChangelogDirTest rocks("./rocksdb");
+    this->setRocksDBDirectory("./rocksdb");
+
+    Storage storage{500, "", this->keeper_context};
+    int zxid = 0;
+
+    const auto create = [&](const String & path, int create_mode)
+    {
+        int new_zxid = ++zxid;
+
+        const auto create_request = std::make_shared<ZooKeeperCreateRequest>();
+        create_request->path = path;
+        create_request->is_ephemeral = create_mode == zkutil::CreateMode::Ephemeral || create_mode == zkutil::CreateMode::EphemeralSequential;
+        create_request->is_sequential = create_mode == zkutil::CreateMode::PersistentSequential || create_mode == zkutil::CreateMode::EphemeralSequential;
+
+        storage.preprocessRequest(create_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(create_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZOK) << "Failed to create " << path;
+    };
+
+    const auto add_watch = [&](const String & path)
+    {
+        int new_zxid = ++zxid;
+
+        const auto exists_request = std::make_shared<ZooKeeperExistsRequest>();
+        exists_request->path = path;
+        exists_request->has_watch = true;
+
+        storage.preprocessRequest(exists_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(exists_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZOK);
+    };
+
+    const auto add_list_watch = [&](const String & path)
+    {
+        int new_zxid = ++zxid;
+
+        const auto list_request = std::make_shared<ZooKeeperListRequest>();
+        list_request->path = path;
+        list_request->has_watch = true;
+
+        storage.preprocessRequest(list_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(list_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZOK);
+    };
+
+    create("/A", zkutil::CreateMode::Persistent);
+    create("/A/B", zkutil::CreateMode::Persistent);
+    create("/A/C", zkutil::CreateMode::Ephemeral);
+    create("/A/B/D", zkutil::CreateMode::Ephemeral);
+
+    add_watch("/A");
+    add_watch("/A/B");
+    add_watch("/A/C");
+    add_watch("/A/B/D");
+    add_list_watch("/A");
+    add_list_watch("/A/B");
+    ASSERT_EQ(storage.watches.size(), 4);
+    ASSERT_EQ(storage.list_watches.size(), 2);
+
+    int new_zxid = ++zxid;
+
+    auto remove_request = std::make_shared<ZooKeeperRemoveRecursiveRequest>();
+    remove_request->path = "/A";
+    remove_request->remove_nodes_limit = 4;
+
+    storage.preprocessRequest(remove_request, 1, 0, new_zxid);
+    auto responses = storage.processRequest(remove_request, 1, new_zxid);
+
+    ASSERT_EQ(responses.size(), 7);
+
+    for (size_t i = 0; i < 7; ++i)
+    {
+        ASSERT_EQ(responses[i].response->error, Coordination::Error::ZOK);
+
+        if (const auto * watch_response = dynamic_cast<Coordination::ZooKeeperWatchResponse *>(responses[i].response.get()))
+            ASSERT_EQ(watch_response->type, Coordination::Event::DELETED);
+    }
+
+    ASSERT_EQ(storage.watches.size(), 0);
+    ASSERT_EQ(storage.list_watches.size(), 0);
+}
+
+TYPED_TEST(CoordinationTest, TestRemoveRecursiveAcls)
+{
+    using namespace DB;
+    using namespace Coordination;
+
+    using Storage = typename TestFixture::Storage;
+
+    ChangelogDirTest rocks("./rocksdb");
+    this->setRocksDBDirectory("./rocksdb");
+
+    Storage storage{500, "", this->keeper_context};
+    int zxid = 0;
+
+    {
+        int new_zxid = ++zxid;
+        String user_auth_data = "test_user:test_password";
+
+        const auto auth_request = std::make_shared<ZooKeeperAuthRequest>();
+        auth_request->scheme = "digest";
+        auth_request->data = user_auth_data;
+
+        storage.preprocessRequest(auth_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(auth_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZOK) << "Failed to add auth to session";
+    }
+
+    const auto create = [&](const String & path)
+    {
+        int new_zxid = ++zxid;
+
+        const auto create_request = std::make_shared<ZooKeeperCreateRequest>();
+        create_request->path = path;
+        create_request->acls = {{.permissions = ACL::Create, .scheme = "auth", .id = ""}};
+
+        storage.preprocessRequest(create_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(create_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZOK) << "Failed to create " << path;
+    };
+
+    /// Add nodes with only Create ACL
+    create("/A");
+    create("/A/B");
+    create("/A/C");
+    create("/A/B/D");
+
+    {
+        int new_zxid = ++zxid;
+
+        auto remove_request = std::make_shared<ZooKeeperRemoveRecursiveRequest>();
+        remove_request->path = "/A";
+        remove_request->remove_nodes_limit = 4;
+
+        storage.preprocessRequest(remove_request, 1, 0, new_zxid);
+        auto responses = storage.processRequest(remove_request, 1, new_zxid);
+
+        EXPECT_EQ(responses.size(), 1);
+        EXPECT_EQ(responses[0].response->error, Coordination::Error::ZNOAUTH);
+    }
+}
+
 /// INSTANTIATE_TEST_SUITE_P(CoordinationTestSuite,
 ///     CoordinationTest,
 ///     ::testing::ValuesIn(std::initializer_list<CompressionParam>{CompressionParam{true, ".zstd"}, CompressionParam{false, ""}}));

src/Core/Defines.h

@@ -20,6 +20,9 @@ static constexpr auto DBMS_DEFAULT_POLL_INTERVAL = 10;
 /// The size of the I/O buffer by default.
 static constexpr auto DBMS_DEFAULT_BUFFER_SIZE = 1048576ULL;
 
+/// The initial size of adaptive I/O buffer by default.
+static constexpr auto DBMS_DEFAULT_INITIAL_ADAPTIVE_BUFFER_SIZE = 16384ULL;
+
 static constexpr auto PADDING_FOR_SIMD = 64;
 
 /** Which blocks by default read the data (by number of rows).
@@ -40,7 +43,7 @@ static constexpr auto SHOW_CHARS_ON_SYNTAX_ERROR = ptrdiff_t(160);
 /// each period reduces the error counter by 2 times
 /// too short a period can cause errors to disappear immediately after creation.
 static constexpr auto DBMS_CONNECTION_POOL_WITH_FAILOVER_DEFAULT_DECREASE_ERROR_PERIOD = 60;
-/// replica error max cap, this is to prevent replica from accumulating too many errors and taking to long to recover.
+/// replica error max cap, this is to prevent replica from accumulating too many errors and taking too long to recover.
 static constexpr auto DBMS_CONNECTION_POOL_WITH_FAILOVER_MAX_ERROR_COUNT = 1000;
 
 /// The boundary on which the blocks for asynchronous file operations should be aligned.

src/Core/ServerSettings.h

@@ -148,6 +148,7 @@ namespace DB
     M(Bool, storage_metadata_write_full_object_key, false, "Write disk metadata files with VERSION_FULL_OBJECT_KEY format", 0) \
     M(UInt64, max_materialized_views_count_for_table, 0, "A limit on the number of materialized views attached to a table.", 0) \
     M(UInt32, max_database_replicated_create_table_thread_pool_size, 1, "The number of threads to create tables during replica recovery in DatabaseReplicated. Zero means number of threads equal number of cores.", 0) \
+    M(Bool, database_replicated_allow_detach_permanently, true, "Allow detaching tables permanently in Replicated databases", 0) \
     M(Bool, format_alter_operations_with_parentheses, false, "If enabled, each operation in alter queries will be surrounded with parentheses in formatted queries to make them less ambiguous.", 0) \
     M(String, default_replica_path, "/clickhouse/tables/{uuid}/{shard}", "The path to the table in ZooKeeper", 0) \
     M(String, default_replica_name, "{replica}", "The replica name in ZooKeeper", 0) \
@@ -169,6 +170,7 @@ namespace DB
     M(Bool, prepare_system_log_tables_on_startup, false, "If true, ClickHouse creates all configured `system.*_log` tables before the startup. It can be helpful if some startup scripts depend on these tables.", 0) \
     M(Double, gwp_asan_force_sample_probability, 0.0003, "Probability that an allocation from specific places will be sampled by GWP Asan (i.e. PODArray allocations)", 0) \
     M(UInt64, config_reload_interval_ms, 2000, "How often clickhouse will reload config and check for new changes", 0) \
+    M(UInt64, memory_worker_period_ms, 0, "Tick period of background memory worker which corrects memory tracker memory usages and cleans up unused pages during higher memory usage. If set to 0, default value will be used depending on the memory usage source", 0) \
     M(Bool, disable_insertion_and_mutation, false, "Disable all insert/alter/delete queries. This setting will be enabled if someone needs read-only nodes to prevent insertion and mutation affect reading performance.", 0)
 
 /// If you add a setting which can be updated at runtime, please update 'changeable_settings' map in StorageSystemServerSettings.cpp

src/Core/Settings.h

@@ -923,6 +923,9 @@ class IColumn;
     M(Bool, implicit_transaction, false, "If enabled and not already inside a transaction, wraps the query inside a full transaction (begin + commit or rollback)", 0) \
     M(UInt64, grace_hash_join_initial_buckets, 1, "Initial number of grace hash join buckets", 0) \
     M(UInt64, grace_hash_join_max_buckets, 1024, "Limit on the number of grace hash join buckets", 0) \
+    M(Int32, join_to_sort_minimum_perkey_rows, 40, "The lower limit of per-key average rows in the right table to determine whether to rerange the right table by key in left or inner join. This setting ensures that the optimization is not applied for sparse table keys", 0) \
+    M(Int32, join_to_sort_maximum_table_rows, 10000, "The maximum number of rows in the right table to determine whether to rerange the right table by key in left or inner join.", 0) \
+    M(Bool, allow_experimental_join_right_table_sorting, false, "If it is set to true, and the conditions of `join_to_sort_minimum_perkey_rows` and `join_to_sort_maximum_table_rows` are met, rerange the right table by key to improve the performance in left or inner hash join.", 0) \
     M(Timezone, session_timezone, "", "This setting can be removed in the future due to potential caveats. It is experimental and is not suitable for production usage. The default timezone for current session or query. The server default timezone if empty.", 0) \
     M(Bool, use_hive_partitioning, false, "Allows to use hive partitioning for File, URL, S3, AzureBlobStorage and HDFS engines.", 0)\
     \
@@ -945,6 +948,7 @@ class IColumn;
     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(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) \
     \
     M(Bool, allow_experimental_inverted_index, false, "If it is set to true, allow to use experimental inverted index.", 0) \
     M(Bool, allow_experimental_full_text_index, false, "If it is set to true, allow to use experimental full-text index.", 0) \
@@ -1140,6 +1144,7 @@ class IColumn;
     M(Bool, input_format_try_infer_variants, false, "Try to infer the Variant type in text formats when there is more than one possible type for column/array elements", 0) \
     M(Bool, type_json_skip_duplicated_paths, false, "When enabled, during parsing JSON object into JSON type duplicated paths will be ignored and only the first one will be inserted instead of an exception", 0) \
     M(UInt64, input_format_json_max_depth, 1000, "Maximum depth of a field in JSON. This is not a strict limit, it does not have to be applied precisely.", 0) \
+    M(Bool, input_format_json_empty_as_default, false, "Treat empty fields in JSON input as default values.", 0) \
     M(Bool, input_format_try_infer_integers, true, "Try to infer integers instead of floats while schema inference in text formats", 0) \
     M(Bool, input_format_try_infer_dates, true, "Try to infer dates from string fields while schema inference in text formats", 0) \
     M(Bool, input_format_try_infer_datetimes, true, "Try to infer datetimes from string fields while schema inference in text formats", 0) \

src/Core/SettingsChangesHistory.cpp

@@ -71,6 +71,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
     },
     {"24.9",
         {
+            {"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."},
             {"create_if_not_exists", false, false, "New setting."},
@@ -79,11 +80,15 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
             {"output_format_identifier_quoting_style", "Backticks", "Backticks", "New setting."},
             {"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"},
+            {"join_to_sort_minimum_perkey_rows", 0, 40, "The lower limit of per-key average rows in the right table to determine whether to rerange the right table by key in left or inner join. This setting ensures that the optimization is not applied for sparse table keys"},
+            {"join_to_sort_maximum_table_rows", 0, 10000, "The maximum number of rows in the right table to determine whether to rerange the right table by key in left or inner join"},
+            {"allow_experimental_join_right_table_sorting", false, false, "If it is set to true, and the conditions of `join_to_sort_minimum_perkey_rows` and `join_to_sort_maximum_table_rows` are met, rerange the right table by key to improve the performance in left or inner hash join"}
         }
     },
     {"24.8",
         {
-            {"rows_before_aggregation", true, true, "Provide exact value for rows_before_aggregation statistic, represents the number of rows read before aggregation"},
+            {"rows_before_aggregation", false, false, "Provide exact value for rows_before_aggregation statistic, represents the number of rows read before aggregation"},
             {"restore_replace_external_table_functions_to_null", false, false, "New setting."},
             {"restore_replace_external_engines_to_null", false, false, "New setting."},
             {"input_format_json_max_depth", 1000000, 1000, "It was unlimited in previous versions, but that was unsafe."},
@@ -99,7 +104,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
             {"use_json_alias_for_old_object_type", true, false, "Use JSON type alias to create new JSON type"},
             {"type_json_skip_duplicated_paths", false, false, "Allow to skip duplicated paths during JSON parsing"},
             {"allow_experimental_vector_similarity_index", false, false, "Added new setting to allow experimental vector similarity indexes"},
-            {"input_format_try_infer_datetimes_only_datetime64", true, false, "Allow to infer DateTime instead of DateTime64 in data formats"}
+            {"input_format_try_infer_datetimes_only_datetime64", true, false, "Allow to infer DateTime instead of DateTime64 in data formats"},
         }
     },
     {"24.7",

src/DataTypes/Serializations/ISerialization.cpp

@@ -420,6 +420,21 @@ bool ISerialization::isEphemeralSubcolumn(const DB::ISerialization::SubstreamPat
     return path[last_elem].type == Substream::VariantElementNullMap;
 }
 
+bool ISerialization::isDynamicSubcolumn(const DB::ISerialization::SubstreamPath & path, size_t prefix_len)
+{
+    if (prefix_len == 0 || prefix_len > path.size())
+        return false;
+
+    for (size_t i = 0; i != prefix_len; ++i)
+    {
+        if (path[i].type == SubstreamType::DynamicData || path[i].type == SubstreamType::DynamicStructure
+            || path[i].type == SubstreamType::ObjectData || path[i].type == SubstreamType::ObjectStructure)
+            return true;
+    }
+
+    return false;
+}
+
 ISerialization::SubstreamData ISerialization::createFromPath(const SubstreamPath & path, size_t prefix_len)
 {
     assert(prefix_len <= path.size());

src/DataTypes/Serializations/ISerialization.h

@@ -457,6 +457,9 @@ public:
     /// for writing/reading data. For example, it's a null-map subcolumn of Variant type (it's always constructed from discriminators);.
     static bool isEphemeralSubcolumn(const SubstreamPath & path, size_t prefix_len);
 
+    /// Returns true if stream with specified path corresponds to dynamic subcolumn.
+    static bool isDynamicSubcolumn(const SubstreamPath & path, size_t prefix_len);
+
 protected:
     template <typename State, typename StatePtr>
     State * checkAndGetState(const StatePtr & state) const;

src/DataTypes/Serializations/SerializationArray.cpp

@@ -11,6 +11,7 @@
 #include <IO/WriteBufferFromString.h>
 
 #include <Formats/FormatSettings.h>
+#include <Formats/JSONUtils.h>
 
 namespace DB
 {
@@ -615,28 +616,49 @@ void SerializationArray::serializeTextJSONPretty(const IColumn & column, size_t
 }
 
 
-void SerializationArray::deserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+template <typename ReturnType>
+ReturnType SerializationArray::deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
 {
-    deserializeTextImpl(column, istr,
-        [&](IColumn & nested_column)
+    auto deserialize_nested = [&settings, this](IColumn & nested_column, ReadBuffer & buf) -> ReturnType
+    {
+        if constexpr (std::is_same_v<ReturnType, void>)
         {
             if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(nested_column))
-                SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(nested_column, istr, settings, nested);
+                SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(nested_column, buf, settings, nested);
             else
-                nested->deserializeTextJSON(nested_column, istr, settings);
+                nested->deserializeTextJSON(nested_column, buf, settings);
+        }
+        else
+        {
+            if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(nested_column))
+                return SerializationNullable::tryDeserializeNullAsDefaultOrNestedTextJSON(nested_column, buf, settings, nested);
+            return nested->tryDeserializeTextJSON(nested_column, buf, settings);
+        }
+    };
+
+    if (settings.json.empty_as_default)
+        return deserializeTextImpl<ReturnType>(column, istr,
+            [&deserialize_nested, &istr](IColumn & nested_column) -> ReturnType
+            {
+                return JSONUtils::deserializeEmpyStringAsDefaultOrNested<ReturnType>(nested_column, istr, deserialize_nested);
             }, false);
+    else
+        return deserializeTextImpl<ReturnType>(column, istr,
+            [&deserialize_nested, &istr](IColumn & nested_column) -> ReturnType
+            {
+                return deserialize_nested(nested_column, istr);
+            }, false);
+}
+
+
+void SerializationArray::deserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+{
+    deserializeTextJSONImpl<void>(column, istr, settings);
 }
 
 bool SerializationArray::tryDeserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
 {
-    auto read_nested = [&](IColumn & nested_column)
-    {
-        if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(nested_column))
-            return SerializationNullable::tryDeserializeNullAsDefaultOrNestedTextJSON(nested_column, istr, settings, nested);
-        return nested->tryDeserializeTextJSON(nested_column, istr, settings);
-    };
-
-    return deserializeTextImpl<bool>(column, istr, std::move(read_nested), false);
+    return deserializeTextJSONImpl<bool>(column, istr, settings);
 }
 
 

src/DataTypes/Serializations/SerializationArray.h

@@ -82,6 +82,10 @@ public:
         SerializationPtr create(const SerializationPtr & prev) const override;
         ColumnPtr create(const ColumnPtr & prev) const override;
     };
+
+private:
+    template <typename ReturnType>
+    ReturnType deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const;
 };
 
 }

src/DataTypes/Serializations/SerializationMap.cpp

@@ -6,6 +6,7 @@
 #include <Columns/ColumnMap.h>
 #include <Core/Field.h>
 #include <Formats/FormatSettings.h>
+#include <Formats/JSONUtils.h>
 #include <Common/assert_cast.h>
 #include <Common/quoteString.h>
 #include <IO/WriteHelpers.h>
@@ -316,28 +317,52 @@ void SerializationMap::serializeTextJSONPretty(const IColumn & column, size_t ro
 }
 
 
-void SerializationMap::deserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+template <typename ReturnType>
+ReturnType SerializationMap::deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
 {
-    deserializeTextImpl(column, istr,
-        [&settings](ReadBuffer & buf, const SerializationPtr & subcolumn_serialization, IColumn & subcolumn)
+    auto deserialize_nested = [&settings](IColumn & subcolumn, ReadBuffer & buf, const SerializationPtr & subcolumn_serialization) -> ReturnType
+    {
+        if constexpr (std::is_same_v<ReturnType, void>)
         {
             if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(subcolumn))
                 SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(subcolumn, buf, settings, subcolumn_serialization);
             else
                 subcolumn_serialization->deserializeTextJSON(subcolumn, buf, settings);
-        });
         }
-
-bool SerializationMap::tryDeserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
-{
-    auto reader = [&settings](ReadBuffer & buf, const SerializationPtr & subcolumn_serialization, IColumn & subcolumn)
+        else
         {
             if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(subcolumn))
                 return SerializationNullable::tryDeserializeNullAsDefaultOrNestedTextJSON(subcolumn, buf, settings, subcolumn_serialization);
             return subcolumn_serialization->tryDeserializeTextJSON(subcolumn, buf, settings);
+        }
     };
 
-    return deserializeTextImpl<bool>(column, istr, reader);
+    if (settings.json.empty_as_default)
+        return deserializeTextImpl<ReturnType>(column, istr,
+            [&deserialize_nested](ReadBuffer & buf, const SerializationPtr & subcolumn_serialization, IColumn & subcolumn) -> ReturnType
+            {
+                return JSONUtils::deserializeEmpyStringAsDefaultOrNested<ReturnType>(subcolumn, buf,
+                    [&deserialize_nested, &subcolumn_serialization](IColumn & subcolumn_, ReadBuffer & buf_) -> ReturnType
+                    {
+                        return deserialize_nested(subcolumn_, buf_, subcolumn_serialization);
+                    });
+            });
+    else
+        return deserializeTextImpl<ReturnType>(column, istr,
+            [&deserialize_nested](ReadBuffer & buf, const SerializationPtr & subcolumn_serialization, IColumn & subcolumn) -> ReturnType
+            {
+                return deserialize_nested(subcolumn, buf, subcolumn_serialization);
+            });
+}
+
+void SerializationMap::deserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+{
+    deserializeTextJSONImpl<void>(column, istr, settings);
+}
+
+bool SerializationMap::tryDeserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+{
+    return deserializeTextJSONImpl<bool>(column, istr, settings);
 }
 
 void SerializationMap::serializeTextXML(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const

src/DataTypes/Serializations/SerializationMap.h

@@ -74,6 +74,9 @@ private:
 
     template <typename ReturnType = void, typename Reader>
     ReturnType deserializeTextImpl(IColumn & column, ReadBuffer & istr, Reader && reader) const;
+
+    template <typename ReturnType>
+    ReturnType deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const;
 };
 
 }

src/DataTypes/Serializations/SerializationTuple.cpp

@@ -5,6 +5,7 @@
 #include <Core/Field.h>
 #include <Columns/ColumnTuple.h>
 #include <Common/assert_cast.h>
+#include <Formats/JSONUtils.h>
 #include <IO/WriteHelpers.h>
 #include <IO/ReadHelpers.h>
 #include <IO/ReadBufferFromString.h>
@@ -313,27 +314,9 @@ void SerializationTuple::serializeTextJSONPretty(const IColumn & column, size_t
 }
 
 template <typename ReturnType>
-ReturnType SerializationTuple::deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+ReturnType SerializationTuple::deserializeTupleJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings, auto && deserialize_element) const
 {
-    static constexpr bool throw_exception = std::is_same_v<ReturnType, void>;
-
-    auto deserialize_element = [&](IColumn & element_column, size_t element_pos)
-    {
-        if constexpr (throw_exception)
-        {
-            if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(element_column))
-                SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(element_column, istr, settings, elems[element_pos]);
-            else
-                elems[element_pos]->deserializeTextJSON(element_column, istr, settings);
-            return true;
-        }
-        else
-        {
-            if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(element_column))
-                return SerializationNullable::tryDeserializeNullAsDefaultOrNestedTextJSON(element_column, istr, settings, elems[element_pos]);
-            return elems[element_pos]->tryDeserializeTextJSON(element_column, istr, settings);
-        }
-    };
+    static constexpr auto throw_exception = std::is_same_v<ReturnType, void>;
 
     if (settings.json.read_named_tuples_as_objects
         && have_explicit_names)
@@ -506,12 +489,51 @@ ReturnType SerializationTuple::deserializeTextJSONImpl(IColumn & column, ReadBuf
     }
 }
 
-void SerializationTuple::deserializeTextJSON(DB::IColumn & column, DB::ReadBuffer & istr, const DB::FormatSettings & settings) const
+template <typename ReturnType>
+ReturnType SerializationTuple::deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
 {
-    deserializeTextJSONImpl(column, istr, settings);
+    auto deserialize_nested = [&settings](IColumn & nested_column, ReadBuffer & buf, const SerializationPtr & nested_column_serialization) -> ReturnType
+    {
+        if constexpr (std::is_same_v<ReturnType, void>)
+        {
+            if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(nested_column))
+                SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(nested_column, buf, settings, nested_column_serialization);
+            else
+                nested_column_serialization->deserializeTextJSON(nested_column, buf, settings);
+        }
+        else
+        {
+            if (settings.null_as_default && !isColumnNullableOrLowCardinalityNullable(nested_column))
+                return SerializationNullable::tryDeserializeNullAsDefaultOrNestedTextJSON(nested_column, buf, settings, nested_column_serialization);
+            else
+                return nested_column_serialization->tryDeserializeTextJSON(nested_column, buf, settings);
+        }
+    };
+
+    if (settings.json.empty_as_default)
+        return deserializeTupleJSONImpl<ReturnType>(column, istr, settings,
+            [&deserialize_nested, &istr, this](IColumn & nested_column, size_t element_pos) -> ReturnType
+            {
+                return JSONUtils::deserializeEmpyStringAsDefaultOrNested<ReturnType>(nested_column, istr,
+                    [&deserialize_nested, element_pos, this](IColumn & nested_column_, ReadBuffer & buf) -> ReturnType
+                    {
+                        return deserialize_nested(nested_column_, buf, elems[element_pos]);
+                    });
+            });
+    else
+        return deserializeTupleJSONImpl<ReturnType>(column, istr, settings,
+            [&deserialize_nested, &istr, this](IColumn & nested_column, size_t element_pos) -> ReturnType
+            {
+                return deserialize_nested(nested_column, istr, elems[element_pos]);
+            });
 }
 
-bool SerializationTuple::tryDeserializeTextJSON(DB::IColumn & column, DB::ReadBuffer & istr, const DB::FormatSettings & settings) const
+void SerializationTuple::deserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
+{
+    deserializeTextJSONImpl<void>(column, istr, settings);
+}
+
+bool SerializationTuple::tryDeserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
 {
     return deserializeTextJSONImpl<bool>(column, istr, settings);
 }

src/DataTypes/Serializations/SerializationTuple.h

@@ -81,7 +81,10 @@ private:
     template <typename ReturnType = void>
     ReturnType deserializeTextImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings, bool whole) const;
 
-    template <typename ReturnType = void>
+    template <typename ReturnType>
+    ReturnType deserializeTupleJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings, auto && deserialize_element) const;
+
+    template <typename ReturnType>
     ReturnType deserializeTextJSONImpl(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const;
 
     template <typename ReturnType = void>

src/Databases/DatabaseAtomic.cpp

@@ -197,8 +197,9 @@ void DatabaseAtomic::renameTable(ContextPtr local_context, const String & table_
             throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Moving tables between databases of different engines is not supported");
     }
 
-    if (exchange && !supportsAtomicRename())
-        throw Exception(ErrorCodes::NOT_IMPLEMENTED, "RENAME EXCHANGE is not supported");
+    std::string message;
+    if (exchange && !supportsAtomicRename(&message))
+        throw Exception(ErrorCodes::NOT_IMPLEMENTED, "RENAME EXCHANGE is not supported because exchanging files is not supported by the OS ({})", message);
 
     waitDatabaseStarted();
 

src/Databases/DatabaseReplicated.cpp

@@ -63,6 +63,7 @@ namespace ErrorCodes
     extern const int NO_ACTIVE_REPLICAS;
     extern const int CANNOT_GET_REPLICATED_DATABASE_SNAPSHOT;
     extern const int CANNOT_RESTORE_TABLE;
+    extern const int SUPPORT_IS_DISABLED;
 }
 
 static constexpr const char * REPLICATED_DATABASE_MARK = "DatabaseReplicated";
@@ -1741,6 +1742,9 @@ void DatabaseReplicated::detachTablePermanently(ContextPtr local_context, const
 {
     waitDatabaseStarted();
 
+    if (!local_context->getServerSettings().database_replicated_allow_detach_permanently)
+        throw Exception(ErrorCodes::SUPPORT_IS_DISABLED, "Support for DETACH TABLE PERMANENTLY is disabled");
+
     auto txn = local_context->getZooKeeperMetadataTransaction();
     assert(!ddl_worker->isCurrentlyActive() || txn);
     if (txn && txn->isInitialQuery())

src/Disks/DiskLocal.cpp

@@ -339,7 +339,15 @@ DiskLocal::writeFile(const String & path, size_t buf_size, WriteMode mode, const
 {
     int flags = (mode == WriteMode::Append) ? (O_APPEND | O_CREAT | O_WRONLY) : -1;
     return std::make_unique<WriteBufferFromFile>(
-        fs::path(disk_path) / path, buf_size, flags, settings.local_throttler);
+        fs::path(disk_path) / path,
+        buf_size,
+        flags,
+        settings.local_throttler,
+        0666,
+        nullptr,
+        0,
+        settings.use_adaptive_write_buffer,
+        settings.adaptive_write_buffer_initial_size);
 }
 
 std::vector<String> DiskLocal::getBlobPath(const String & path) const

src/Disks/IO/WriteBufferFromAzureBlobStorage.cpp

@@ -59,7 +59,7 @@ WriteBufferFromAzureBlobStorage::WriteBufferFromAzureBlobStorage(
     const WriteSettings & write_settings_,
     std::shared_ptr<const AzureBlobStorage::RequestSettings> settings_,
     ThreadPoolCallbackRunnerUnsafe<void> schedule_)
-    : WriteBufferFromFileBase(buf_size_, nullptr, 0)
+    : WriteBufferFromFileBase(std::min(buf_size_, static_cast<size_t>(DBMS_DEFAULT_BUFFER_SIZE)), nullptr, 0)
     , log(getLogger("WriteBufferFromAzureBlobStorage"))
     , buffer_allocation_policy(createBufferAllocationPolicy(*settings_))
     , max_single_part_upload_size(settings_->max_single_part_upload_size)
@@ -244,11 +244,21 @@ void WriteBufferFromAzureBlobStorage::allocateBuffer()
     buffer_allocation_policy->nextBuffer();
     chassert(0 == hidden_size);
 
-    auto size = buffer_allocation_policy->getBufferSize();
-
+    /// First buffer was already allocated in BufferWithOwnMemory constructor with buffer size provided in constructor.
+    /// It will be reallocated in subsequent nextImpl calls up to the desired buffer size from buffer_allocation_policy.
     if (buffer_allocation_policy->getBufferNumber() == 1)
-        size = std::min(size_t(DBMS_DEFAULT_BUFFER_SIZE), size);
+    {
+        /// Reduce memory size if initial size was larger then desired size from buffer_allocation_policy.
+        /// Usually it doesn't happen but we have it in unit tests.
+        if (memory.size() > buffer_allocation_policy->getBufferSize())
+        {
+            memory.resize(buffer_allocation_policy->getBufferSize());
+            WriteBuffer::set(memory.data(), memory.size());
+        }
+        return;
+    }
 
+    auto size = buffer_allocation_policy->getBufferSize();
     memory = Memory(size);
     WriteBuffer::set(memory.data(), memory.size());
 }

src/Disks/ObjectStorages/AzureBlobStorage/AzureObjectStorage.cpp

@@ -289,7 +289,7 @@ std::unique_ptr<WriteBufferFromFileBase> AzureObjectStorage::writeObject( /// NO
     return std::make_unique<WriteBufferFromAzureBlobStorage>(
         client.get(),
         object.remote_path,
-        buf_size,
+        write_settings.use_adaptive_write_buffer ? write_settings.adaptive_write_buffer_initial_size : buf_size,
         patchSettings(write_settings),
         settings.get(),
         std::move(scheduler));

src/Disks/ObjectStorages/S3/S3ObjectStorage.cpp

@@ -282,7 +282,7 @@ std::unique_ptr<WriteBufferFromFileBase> S3ObjectStorage::writeObject( /// NOLIN
         client.get(),
         uri.bucket,
         object.remote_path,
-        buf_size,
+        write_settings.use_adaptive_write_buffer ? write_settings.adaptive_write_buffer_initial_size : buf_size,
         request_settings,
         std::move(blob_storage_log),
         attributes,

src/Formats/FormatFactory.cpp

@@ -152,6 +152,7 @@ FormatSettings getFormatSettings(const ContextPtr & context, const Settings & se
     format_settings.json.try_infer_objects_as_tuples = settings.input_format_json_try_infer_named_tuples_from_objects;
     format_settings.json.throw_on_bad_escape_sequence = settings.input_format_json_throw_on_bad_escape_sequence;
     format_settings.json.ignore_unnecessary_fields = settings.input_format_json_ignore_unnecessary_fields;
+    format_settings.json.empty_as_default = settings.input_format_json_empty_as_default;
     format_settings.json.type_json_skip_duplicated_paths = settings.type_json_skip_duplicated_paths;
     format_settings.null_as_default = settings.input_format_null_as_default;
     format_settings.force_null_for_omitted_fields = settings.input_format_force_null_for_omitted_fields;

src/Formats/FormatSettings.h

@@ -237,6 +237,7 @@ struct FormatSettings
         bool infer_incomplete_types_as_strings = true;
         bool throw_on_bad_escape_sequence = true;
         bool ignore_unnecessary_fields = true;
+        bool empty_as_default = false;
         bool type_json_skip_duplicated_paths = false;
     } json{};
 

src/Formats/JSONUtils.cpp

@@ -2,12 +2,14 @@
 #include <Formats/JSONUtils.h>
 #include <Formats/ReadSchemaUtils.h>
 #include <Formats/EscapingRuleUtils.h>
+#include <IO/PeekableReadBuffer.h>
 #include <IO/ReadBufferFromString.h>
 #include <IO/WriteBufferValidUTF8.h>
 #include <DataTypes/Serializations/SerializationNullable.h>
 #include <DataTypes/DataTypeNullable.h>
 #include <DataTypes/DataTypeObjectDeprecated.h>
 #include <DataTypes/DataTypeFactory.h>
+#include <Common/assert_cast.h>
 
 #include <base/find_symbols.h>
 
@@ -286,11 +288,19 @@ namespace JSONUtils
                 return true;
             }
 
+            auto deserialize = [as_nullable, &format_settings, &serialization](IColumn & column_, ReadBuffer & buf) -> bool
+            {
                 if (as_nullable)
-                return SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(column, in, format_settings, serialization);
+                    return SerializationNullable::deserializeNullAsDefaultOrNestedTextJSON(column_, buf, format_settings, serialization);
 
-            serialization->deserializeTextJSON(column, in, format_settings);
+                serialization->deserializeTextJSON(column_, buf, format_settings);
                 return true;
+            };
+
+            if (format_settings.json.empty_as_default)
+                return JSONUtils::deserializeEmpyStringAsDefaultOrNested<bool, false>(column, in, deserialize);
+            else
+                return deserialize(column, in);
         }
         catch (Exception & e)
         {
@@ -920,6 +930,78 @@ namespace JSONUtils
         }
     }
 
+    template <typename ReturnType, bool default_column_return_value>
+    ReturnType deserializeEmpyStringAsDefaultOrNested(IColumn & column, ReadBuffer & istr, const NestedDeserialize<ReturnType> & deserialize_nested)
+    {
+        static constexpr auto throw_exception = std::is_same_v<ReturnType, void>;
+
+        static constexpr auto EMPTY_STRING = "\"\"";
+        static constexpr auto EMPTY_STRING_LENGTH = std::string_view(EMPTY_STRING).length();
+
+        if (istr.eof() || *istr.position() != EMPTY_STRING[0])
+            return deserialize_nested(column, istr);
+
+        auto do_deserialize = [](IColumn & column_, ReadBuffer & buf, auto && check_for_empty_string, auto && deserialize) -> ReturnType
+        {
+            if (check_for_empty_string(buf))
+            {
+                column_.insertDefault();
+                return ReturnType(default_column_return_value);
+            }
+            return deserialize(column_, buf);
+        };
+
+        if (istr.available() >= EMPTY_STRING_LENGTH)
+        {
+            /// We have enough data in buffer to check if we have an empty string.
+            auto check_for_empty_string = [](ReadBuffer & buf) -> bool
+            {
+                auto * pos = buf.position();
+                if (checkString(EMPTY_STRING, buf))
+                    return true;
+                buf.position() = pos;
+                return false;
+            };
+
+            return do_deserialize(column, istr, check_for_empty_string, deserialize_nested);
+        }
+
+        /// We don't have enough data in buffer to check if we have an empty string.
+        /// Use PeekableReadBuffer to make a checkpoint before checking for an
+        /// empty string and rollback if check was failed.
+
+        auto check_for_empty_string = [](ReadBuffer & buf) -> bool
+        {
+            auto & peekable_buf = assert_cast<PeekableReadBuffer &>(buf);
+            peekable_buf.setCheckpoint();
+            SCOPE_EXIT(peekable_buf.dropCheckpoint());
+            if (checkString(EMPTY_STRING, peekable_buf))
+                return true;
+            peekable_buf.rollbackToCheckpoint();
+            return false;
+        };
+
+        auto deserialize_nested_with_check = [&deserialize_nested](IColumn & column_, ReadBuffer & buf) -> ReturnType
+        {
+            auto & peekable_buf = assert_cast<PeekableReadBuffer &>(buf);
+            if constexpr (throw_exception)
+                deserialize_nested(column_, peekable_buf);
+            else if (!deserialize_nested(column_, peekable_buf))
+                return ReturnType(false);
+
+            if (unlikely(peekable_buf.hasUnreadData()))
+                throw Exception(ErrorCodes::LOGICAL_ERROR, "Incorrect state while parsing JSON: PeekableReadBuffer has unread data in own memory: {}", String(peekable_buf.position(), peekable_buf.available()));
+
+            return ReturnType(true);
+        };
+
+        PeekableReadBuffer peekable_buf(istr, true);
+        return do_deserialize(column, peekable_buf, check_for_empty_string, deserialize_nested_with_check);
+    }
+
+    template void deserializeEmpyStringAsDefaultOrNested<void, true>(IColumn & column, ReadBuffer & istr, const NestedDeserialize<void> & deserialize_nested);
+    template bool deserializeEmpyStringAsDefaultOrNested<bool, true>(IColumn & column, ReadBuffer & istr, const NestedDeserialize<bool> & deserialize_nested);
+    template bool deserializeEmpyStringAsDefaultOrNested<bool, false>(IColumn & column, ReadBuffer & istr, const NestedDeserialize<bool> & deserialize_nested);
 }
 
 }

src/Formats/JSONUtils.h

@@ -8,6 +8,7 @@
 #include <IO/Progress.h>
 #include <Core/NamesAndTypes.h>
 #include <Common/Stopwatch.h>
+#include <functional>
 #include <utility>
 
 namespace DB
@@ -146,6 +147,16 @@ namespace JSONUtils
 
     bool skipUntilFieldInObject(ReadBuffer & in, const String & desired_field_name, const FormatSettings::JSON & settings);
     void skipTheRestOfObject(ReadBuffer & in, const FormatSettings::JSON & settings);
+
+    template <typename ReturnType>
+    using NestedDeserialize = std::function<ReturnType(IColumn &, ReadBuffer &)>;
+
+    template <typename ReturnType, bool default_column_return_value = true>
+    ReturnType deserializeEmpyStringAsDefaultOrNested(IColumn & column, ReadBuffer & istr, const NestedDeserialize<ReturnType> & deserialize_nested);
+
+    extern template void deserializeEmpyStringAsDefaultOrNested<void, true>(IColumn & column, ReadBuffer & istr, const NestedDeserialize<void> & deserialize_nested);
+    extern template bool deserializeEmpyStringAsDefaultOrNested<bool, true>(IColumn & column, ReadBuffer & istr, const NestedDeserialize<bool> & deserialize_nested);
+    extern template bool deserializeEmpyStringAsDefaultOrNested<bool, false>(IColumn & column, ReadBuffer & istr, const NestedDeserialize<bool> & deserialize_nested);
 }
 
 }

src/Functions/materialize.h

@@ -2,7 +2,7 @@
 #include <Functions/IFunction.h>
 #include <Functions/FunctionFactory.h>
 #include <Columns/ColumnLowCardinality.h>
-#include <DataTypes/DataTypeLowCardinality.h>
+#include <Columns/ColumnSparse.h>
 
 namespace DB
 {
@@ -18,11 +18,6 @@ public:
         return std::make_shared<FunctionMaterialize>();
     }
 
-    bool useDefaultImplementationForNulls() const override
-    {
-        return false;
-    }
-
     /// Get the function name.
     String getName() const override
     {
@@ -34,8 +29,16 @@ public:
         return true;
     }
 
+    bool useDefaultImplementationForNulls() const override { return false; }
+
+    bool useDefaultImplementationForNothing() const override { return false; }
+
+    bool useDefaultImplementationForConstants() const override { return false; }
+
     bool useDefaultImplementationForLowCardinalityColumns() const override { return false; }
 
+    bool useDefaultImplementationForSparseColumns() const override { return false; }
+
     bool isSuitableForConstantFolding() const override { return false; }
 
     bool isSuitableForShortCircuitArgumentsExecution(const DataTypesWithConstInfo & /*arguments*/) const override { return false; }
@@ -52,7 +55,7 @@ public:
 
     ColumnPtr executeImpl(const ColumnsWithTypeAndName & arguments, const DataTypePtr &, size_t /*input_rows_count*/) const override
     {
-        return arguments[0].column->convertToFullColumnIfConst();
+        return recursiveRemoveSparse(arguments[0].column->convertToFullColumnIfConst());
     }
 
     bool hasInformationAboutMonotonicity() const override { return true; }

src/Functions/sqid.cpp

@@ -124,7 +124,7 @@ public:
                 std::string_view sqid = col_non_const->getDataAt(i).toView();
                 std::vector<UInt64> integers = sqids.decode(String(sqid));
                 res_nested_data.insert(integers.begin(), integers.end());
-                res_offsets_data.push_back(integers.size());
+                res_offsets_data.push_back(res_offsets_data.back() + integers.size());
             }
         }
         else

src/IO/S3/URI.cpp

@@ -36,7 +36,7 @@ URI::URI(const std::string & uri_, bool allow_archive_path_syntax)
     /// Case when bucket name represented in domain name of S3 URL.
     /// E.g. (https://bucket-name.s3.region.amazonaws.com/key)
     /// https://docs.aws.amazon.com/AmazonS3/latest/dev/VirtualHosting.html#virtual-hosted-style-access
-    static const RE2 virtual_hosted_style_pattern(R"((.+)\.(s3express[\-a-z0-9]+|s3|cos|obs|oss|eos)([.\-][a-z0-9\-.:]+))");
+    static const RE2 virtual_hosted_style_pattern(R"(([^.]+)\.(s3express[\-a-z0-9]+|s3|cos|obs|.*oss[^\/]*|eos)([.\-][a-z0-9\-.:]+))");
 
     /// Case when AWS Private Link Interface is being used
     /// E.g. (bucket.vpce-07a1cd78f1bd55c5f-j3a3vg6w.s3.us-east-1.vpce.amazonaws.com/bucket-name/key)

src/IO/WriteBufferFromFile.cpp

@@ -32,8 +32,10 @@ WriteBufferFromFile::WriteBufferFromFile(
     ThrottlerPtr throttler_,
     mode_t mode,
     char * existing_memory,
-    size_t alignment)
-    : WriteBufferFromFileDescriptor(-1, buf_size, existing_memory, throttler_, alignment, file_name_)
+    size_t alignment,
+    bool use_adaptive_buffer_size_,
+    size_t adaptive_buffer_initial_size)
+    : WriteBufferFromFileDescriptor(-1, buf_size, existing_memory, throttler_, alignment, file_name_, use_adaptive_buffer_size_, adaptive_buffer_initial_size)
 {
     ProfileEvents::increment(ProfileEvents::FileOpen);
 
@@ -66,8 +68,10 @@ WriteBufferFromFile::WriteBufferFromFile(
     size_t buf_size,
     ThrottlerPtr throttler_,
     char * existing_memory,
-    size_t alignment)
-    : WriteBufferFromFileDescriptor(fd_, buf_size, existing_memory, throttler_, alignment, original_file_name)
+    size_t alignment,
+    bool use_adaptive_buffer_size_,
+    size_t adaptive_buffer_initial_size)
+    : WriteBufferFromFileDescriptor(fd_, buf_size, existing_memory, throttler_, alignment, original_file_name, use_adaptive_buffer_size_, adaptive_buffer_initial_size)
 {
     fd_ = -1;
 }

src/IO/WriteBufferFromFile.h

@@ -36,7 +36,9 @@ public:
         ThrottlerPtr throttler_ = {},
         mode_t mode = 0666,
         char * existing_memory = nullptr,
-        size_t alignment = 0);
+        size_t alignment = 0,
+        bool use_adaptive_buffer_size_ = false,
+        size_t adaptive_buffer_initial_size = DBMS_DEFAULT_INITIAL_ADAPTIVE_BUFFER_SIZE);
 
     /// Use pre-opened file descriptor.
     explicit WriteBufferFromFile(
@@ -45,7 +47,9 @@ public:
         size_t buf_size = DBMS_DEFAULT_BUFFER_SIZE,
         ThrottlerPtr throttler_ = {},
         char * existing_memory = nullptr,
-        size_t alignment = 0);
+        size_t alignment = 0,
+        bool use_adaptive_buffer_size_ = false,
+        size_t adaptive_buffer_initial_size = DBMS_DEFAULT_INITIAL_ADAPTIVE_BUFFER_SIZE);
 
     ~WriteBufferFromFile() override;
 

src/IO/WriteBufferFromFileDescriptor.cpp

@@ -83,6 +83,13 @@ void WriteBufferFromFileDescriptor::nextImpl()
 
     ProfileEvents::increment(ProfileEvents::DiskWriteElapsedMicroseconds, watch.elapsedMicroseconds());
     ProfileEvents::increment(ProfileEvents::WriteBufferFromFileDescriptorWriteBytes, bytes_written);
+
+    /// Increase buffer size for next data if adaptive buffer size is used and nextImpl was called because of end of buffer.
+    if (!available() && use_adaptive_buffer_size && memory.size() < adaptive_max_buffer_size)
+    {
+        memory.resize(std::min(memory.size() * 2, adaptive_max_buffer_size));
+        BufferBase::set(memory.data(), memory.size(), 0);
+    }
 }
 
 /// NOTE: This class can be used as a very low-level building block, for example
@@ -94,11 +101,15 @@ WriteBufferFromFileDescriptor::WriteBufferFromFileDescriptor(
     char * existing_memory,
     ThrottlerPtr throttler_,
     size_t alignment,
-    std::string file_name_)
-    : WriteBufferFromFileBase(buf_size, existing_memory, alignment)
+    std::string file_name_,
+    bool use_adaptive_buffer_size_,
+    size_t adaptive_buffer_initial_size)
+    : WriteBufferFromFileBase(use_adaptive_buffer_size_ ? adaptive_buffer_initial_size : buf_size, existing_memory, alignment)
     , fd(fd_)
     , throttler(throttler_)
     , file_name(std::move(file_name_))
+    , use_adaptive_buffer_size(use_adaptive_buffer_size_)
+    , adaptive_max_buffer_size(buf_size)
 {
 }
 
@@ -124,6 +135,7 @@ void WriteBufferFromFileDescriptor::finalizeImpl()
         return;
     }
 
+    use_adaptive_buffer_size = false;
     next();
 }
 

src/IO/WriteBufferFromFileDescriptor.h

@@ -18,7 +18,9 @@ public:
         char * existing_memory = nullptr,
         ThrottlerPtr throttler_ = {},
         size_t alignment = 0,
-        std::string file_name_ = "");
+        std::string file_name_ = "",
+        bool use_adaptive_buffer_size_ = false,
+        size_t adaptive_buffer_initial_size = DBMS_DEFAULT_INITIAL_ADAPTIVE_BUFFER_SIZE);
 
     /** Could be used before initialization if needed 'fd' was not passed to constructor.
       * It's not possible to change 'fd' during work.
@@ -56,6 +58,12 @@ protected:
     /// If file has name contains filename, otherwise contains string "(fd=...)"
     std::string file_name;
 
+    /// If true, the size of internal buffer will be exponentially increased up to
+    /// adaptive_buffer_max_size after each nextImpl call. It can be used to avoid
+    /// large buffer allocation when actual size of written data is small.
+    bool use_adaptive_buffer_size;
+    size_t adaptive_max_buffer_size;
+
     void finalizeImpl() override;
 };
 

src/IO/WriteBufferFromS3.cpp

@@ -95,7 +95,7 @@ WriteBufferFromS3::WriteBufferFromS3(
     std::optional<std::map<String, String>> object_metadata_,
     ThreadPoolCallbackRunnerUnsafe<void> schedule_,
     const WriteSettings & write_settings_)
-    : WriteBufferFromFileBase(buf_size_, nullptr, 0)
+    : WriteBufferFromFileBase(std::min(buf_size_, static_cast<size_t>(DBMS_DEFAULT_BUFFER_SIZE)), nullptr, 0)
     , bucket(bucket_)
     , key(key_)
     , request_settings(request_settings_)
@@ -351,9 +351,17 @@ void WriteBufferFromS3::allocateBuffer()
     buffer_allocation_policy->nextBuffer();
     chassert(0 == hidden_size);
 
+    /// First buffer was already allocated in BufferWithOwnMemory constructor with provided in constructor buffer size.
+    /// It will be reallocated in subsequent nextImpl calls up to the desired buffer size from buffer_allocation_policy.
     if (buffer_allocation_policy->getBufferNumber() == 1)
     {
-        allocateFirstBuffer();
+        /// Reduce memory size if initial size was larger then desired size from buffer_allocation_policy.
+        /// Usually it doesn't happen but we have it in unit tests.
+        if (memory.size() > buffer_allocation_policy->getBufferSize())
+        {
+            memory.resize(buffer_allocation_policy->getBufferSize());
+            WriteBuffer::set(memory.data(), memory.size());
+        }
         return;
     }
 
@@ -361,14 +369,6 @@ void WriteBufferFromS3::allocateBuffer()
     WriteBuffer::set(memory.data(), memory.size());
 }
 
-void WriteBufferFromS3::allocateFirstBuffer()
-{
-    const auto max_first_buffer = buffer_allocation_policy->getBufferSize();
-    const auto size = std::min(size_t(DBMS_DEFAULT_BUFFER_SIZE), max_first_buffer);
-    memory = Memory(size);
-    WriteBuffer::set(memory.data(), memory.size());
-}
-
 void WriteBufferFromS3::setFakeBufferWhenPreFinalized()
 {
     WriteBuffer::set(fake_buffer_when_prefinalized, sizeof(fake_buffer_when_prefinalized));

src/IO/WriteBufferFromS3.h

@@ -64,7 +64,6 @@ private:
     void reallocateFirstBuffer();
     void detachBuffer();
     void allocateBuffer();
-    void allocateFirstBuffer();
     void setFakeBufferWhenPreFinalized();
 
     S3::UploadPartRequest getUploadRequest(size_t part_number, PartData & data);

src/IO/WriteBufferValidUTF8.cpp

@@ -54,7 +54,7 @@ inline void WriteBufferValidUTF8::putReplacement()
 }
 
 
-inline void WriteBufferValidUTF8::putValid(char *data, size_t len)
+inline void WriteBufferValidUTF8::putValid(const char *data, size_t len)
 {
     if (len == 0)
         return;
@@ -149,9 +149,34 @@ void WriteBufferValidUTF8::finalizeImpl()
     /// Write all complete sequences from buffer.
     nextImpl();
 
-    /// If unfinished sequence at end, then write replacement.
+    /// Handle remaining bytes if we have an incomplete sequence
     if (working_buffer.begin() != memory.data())
+    {
+        const char * p = memory.data();
+
+        while (p < pos)
+        {
+            UInt8 len = length_of_utf8_sequence[static_cast<const unsigned char>(*p)];
+            if (p + len > pos)
+            {
+                /// Incomplete sequence. Skip one byte.
                 putReplacement();
+                ++p;
+            }
+            else if (Poco::UTF8Encoding::isLegal(reinterpret_cast<const unsigned char *>(p), len))
+            {
+                /// Valid sequence
+                putValid(p, len);
+                p += len;
+            }
+            else
+            {
+                /// Invalid sequence, skip first byte.
+                putReplacement();
+                ++p;
+            }
+        }
+    }
 }
 
 }