Merge branch 'master' of github.com:ClickHouse/ClickHouse into fix-test-replace-current-query

.github/PULL_REQUEST_TEMPLATE.md

@@ -59,6 +59,9 @@ At a minimum, the following information should be added (but add more as needed)
 - [ ] <!---ci_exclude_tsan|msan|ubsan|coverage--> Exclude: All with TSAN, MSAN, UBSAN, Coverage
 - [ ] <!---ci_exclude_aarch64|release|debug--> Exclude: All with aarch64, release, debug
 ---
+- [ ] <!---ci_include_fuzzer--> Run only fuzzers related jobs (libFuzzer fuzzers, AST fuzzers, etc.)
+- [ ] <!---ci_exclude_ast--> Exclude: AST fuzzers
+---
 - [ ] <!---do_not_test--> Do not test
 - [ ] <!---woolen_wolfdog--> Woolen Wolfdog
 - [ ] <!---upload_all--> Upload binaries for special builds

.github/workflows/create_release.yml

@@ -129,9 +129,9 @@ jobs:
         if: ${{ inputs.type == 'patch' && ! inputs.only-repo }}
         shell: bash
         run: |
-          python3 ./tests/ci/create_release.py --set-progress-completed
           git reset --hard HEAD
           git checkout "$GITHUB_REF_NAME"
+          python3 ./tests/ci/create_release.py --set-progress-completed
       - name: Create GH Release
         if: ${{ inputs.type == 'patch' && ! inputs.only-repo }}
         shell: bash

.gitmodules

@@ -230,9 +230,6 @@
 [submodule "contrib/minizip-ng"]
 	path = contrib/minizip-ng
 	url = https://github.com/zlib-ng/minizip-ng
-[submodule "contrib/annoy"]
-	path = contrib/annoy
-	url = https://github.com/ClickHouse/annoy
 [submodule "contrib/qpl"]
 	path = contrib/qpl
 	url = https://github.com/intel/qpl

base/base/CMakeLists.txt

@@ -1,4 +1,4 @@
-add_compile_options($<$<OR:$<COMPILE_LANGUAGE:C>,$<COMPILE_LANGUAGE:CXX>>:${COVERAGE_FLAGS}>)
+add_compile_options("$<$<OR:$<COMPILE_LANGUAGE:C>,$<COMPILE_LANGUAGE:CXX>>:${COVERAGE_FLAGS}>")
 
 if (USE_CLANG_TIDY)
     set (CMAKE_CXX_CLANG_TIDY "${CLANG_TIDY_PATH}")

base/base/cgroupsv2.cpp

@@ -27,27 +27,6 @@ bool cgroupsV2Enabled()
 #endif
 }
 
-bool cgroupsV2MemoryControllerEnabled()
-{
-#if defined(OS_LINUX)
-    chassert(cgroupsV2Enabled());
-    /// According to https://docs.kernel.org/admin-guide/cgroup-v2.html, file "cgroup.controllers" defines which controllers are available
-    /// for the current + child cgroups. The set of available controllers can be restricted from level to level using file
-    /// "cgroups.subtree_control". It is therefore sufficient to check the bottom-most nested "cgroup.controllers" file.
-    fs::path cgroup_dir = cgroupV2PathOfProcess();
-    if (cgroup_dir.empty())
-        return false;
-    std::ifstream controllers_file(cgroup_dir / "cgroup.controllers");
-    if (!controllers_file.is_open())
-        return false;
-    std::string controllers;
-    std::getline(controllers_file, controllers);
-    return controllers.find("memory") != std::string::npos;
-#else
-    return false;
-#endif
-}
-
 fs::path cgroupV2PathOfProcess()
 {
 #if defined(OS_LINUX)
@@ -71,3 +50,28 @@ fs::path cgroupV2PathOfProcess()
     return {};
 #endif
 }
+
+std::optional<std::string> getCgroupsV2PathContainingFile([[maybe_unused]] std::string_view file_name)
+{
+#if defined(OS_LINUX)
+    if (!cgroupsV2Enabled())
+        return {};
+
+    fs::path current_cgroup = cgroupV2PathOfProcess();
+    if (current_cgroup.empty())
+        return {};
+
+    /// Return the bottom-most nested file. If there is no such file at the current
+    /// level, try again at the parent level as settings are inherited.
+    while (current_cgroup != default_cgroups_mount.parent_path())
+    {
+        const auto path = current_cgroup / file_name;
+        if (fs::exists(path))
+            return {current_cgroup};
+        current_cgroup = current_cgroup.parent_path();
+    }
+    return {};
+#else
+    return {};
+#endif
+}

base/base/cgroupsv2.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <filesystem>
+#include <string_view>
 
 #if defined(OS_LINUX)
 /// I think it is possible to mount the cgroups hierarchy somewhere else (e.g. when in containers).
@@ -11,11 +12,11 @@ static inline const std::filesystem::path default_cgroups_mount = "/sys/fs/cgrou
 /// Is cgroups v2 enabled on the system?
 bool cgroupsV2Enabled();
 
-/// Is the memory controller of cgroups v2 enabled on the system?
-/// Assumes that cgroupsV2Enabled() is enabled.
-bool cgroupsV2MemoryControllerEnabled();
-
 /// Detects which cgroup v2 the process belongs to and returns the filesystem path to the cgroup.
 /// Returns an empty path the cgroup cannot be determined.
 /// Assumes that cgroupsV2Enabled() is enabled.
 std::filesystem::path cgroupV2PathOfProcess();
+
+/// Returns the most nested cgroup dir containing the specified file.
+/// If cgroups v2 is not enabled - returns an empty optional.
+std::optional<std::string> getCgroupsV2PathContainingFile([[maybe_unused]] std::string_view file_name);

base/base/getMemoryAmount.cpp

@@ -19,9 +19,6 @@ std::optional<uint64_t> getCgroupsV2MemoryLimit()
     if (!cgroupsV2Enabled())
         return {};
 
-    if (!cgroupsV2MemoryControllerEnabled())
-        return {};
-
     std::filesystem::path current_cgroup = cgroupV2PathOfProcess();
     if (current_cgroup.empty())
         return {};

base/poco/Net/include/Poco/Net/HTTPServerSession.h

@@ -58,6 +58,10 @@ namespace Net
 
         void setKeepAliveTimeout(Poco::Timespan keepAliveTimeout);
 
+        size_t getKeepAliveTimeout() const { return _keepAliveTimeout.totalSeconds(); }
+
+        size_t getMaxKeepAliveRequests() const { return _maxKeepAliveRequests; }
+
     private:
         bool _firstRequest;
         Poco::Timespan _keepAliveTimeout;

base/poco/Net/src/HTTPServerSession.cpp

@@ -19,11 +19,11 @@ namespace Poco {
 namespace Net {
 
 
-HTTPServerSession::HTTPServerSession(const StreamSocket& socket, HTTPServerParams::Ptr pParams):
-	HTTPSession(socket, pParams->getKeepAlive()),
-	_firstRequest(true),
-	_keepAliveTimeout(pParams->getKeepAliveTimeout()),
-	_maxKeepAliveRequests(pParams->getMaxKeepAliveRequests())
+HTTPServerSession::HTTPServerSession(const StreamSocket & socket, HTTPServerParams::Ptr pParams)
+    : HTTPSession(socket, pParams->getKeepAlive())
+    , _firstRequest(true)
+    , _keepAliveTimeout(pParams->getKeepAliveTimeout())
+    , _maxKeepAliveRequests(pParams->getMaxKeepAliveRequests())
 {
 	setTimeout(pParams->getTimeout());
 }
@@ -52,11 +52,12 @@ bool HTTPServerSession::hasMoreRequests()
 	}
 	else if (_maxKeepAliveRequests != 0 && getKeepAlive())
 	{
-		if (_maxKeepAliveRequests > 0) 
-			--_maxKeepAliveRequests;
-		return buffered() > 0 || socket().poll(_keepAliveTimeout, Socket::SELECT_READ);
-	}
-	else return false;
+        if (_maxKeepAliveRequests > 0)
+            --_maxKeepAliveRequests;
+        return buffered() > 0 || socket().poll(_keepAliveTimeout, Socket::SELECT_READ);
+    }
+    else
+        return false;
 }
 
 

cmake/autogenerated_versions.txt

@@ -2,11 +2,11 @@
 
 # NOTE: VERSION_REVISION has nothing common with DBMS_TCP_PROTOCOL_VERSION,
 # only DBMS_TCP_PROTOCOL_VERSION should be incremented on protocol changes.
-SET(VERSION_REVISION 54489)
+SET(VERSION_REVISION 54490)
 SET(VERSION_MAJOR 24)
-SET(VERSION_MINOR 8)
+SET(VERSION_MINOR 9)
 SET(VERSION_PATCH 1)
-SET(VERSION_GITHASH 3f8b27d7accd2b5ec4afe7d0dd459115323304af)
-SET(VERSION_DESCRIBE v24.8.1.1-testing)
-SET(VERSION_STRING 24.8.1.1)
+SET(VERSION_GITHASH e02b434d2fc0c4fbee29ca675deab7474d274608)
+SET(VERSION_DESCRIBE v24.9.1.1-testing)
+SET(VERSION_STRING 24.9.1.1)
 # end of autochange

cmake/sanitize.cmake

@@ -57,8 +57,8 @@ option(WITH_COVERAGE "Instrumentation for code coverage with default implementat
 
 if (WITH_COVERAGE)
     message (STATUS "Enabled instrumentation for code coverage")
-    set(COVERAGE_FLAGS "SHELL:-fprofile-instr-generate -fcoverage-mapping")
-    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fprofile-instr-generate -fcoverage-mapping")
+    set (COVERAGE_FLAGS -fprofile-instr-generate -fcoverage-mapping)
+    set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fprofile-instr-generate -fcoverage-mapping")
 endif()
 
 option (SANITIZE_COVERAGE "Instrumentation for code coverage with custom callbacks" OFF)

contrib/CMakeLists.txt

@@ -205,9 +205,8 @@ add_contrib (morton-nd-cmake morton-nd)
 if (ARCH_S390X)
     add_contrib(crc32-s390x-cmake crc32-s390x)
 endif()
-add_contrib (annoy-cmake annoy)
 
-option(ENABLE_USEARCH "Enable USearch (Approximate Neighborhood Search, HNSW) support" ${ENABLE_LIBRARIES})
+option(ENABLE_USEARCH "Enable USearch" ${ENABLE_LIBRARIES})
 if (ENABLE_USEARCH)
     add_contrib (FP16-cmake FP16)
     add_contrib (robin-map-cmake robin-map)

contrib/annoy

@@ -1 +0,0 @@
-Subproject commit f2ac8e7b48f9a9cf676d3b58286e5455aba8e956

contrib/annoy-cmake/CMakeLists.txt

@@ -1,24 +0,0 @@
-option(ENABLE_ANNOY "Enable Annoy index support" ${ENABLE_LIBRARIES})
-
-# Annoy index should be disabled with undefined sanitizer. Because of memory storage optimizations
-# (https://github.com/ClickHouse/annoy/blob/9d8a603a4cd252448589e84c9846f94368d5a289/src/annoylib.h#L442-L463)
-# UBSan fails and leads to crash. Simmilar issue is already opened in Annoy repo
-# https://github.com/spotify/annoy/issues/456
-# Problem with aligment can lead to errors like
-# (https://stackoverflow.com/questions/46790550/c-undefined-behavior-strict-aliasing-rule-or-incorrect-alignment)
-# or will lead to crash on arm https://developer.arm.com/documentation/ka003038/latest
-# This issues should be resolved before annoy became non-experimental (--> setting "allow_experimental_annoy_index")
-if ((NOT ENABLE_ANNOY) OR (SANITIZE STREQUAL "undefined") OR (ARCH_AARCH64))
-    message (STATUS "Not using annoy")
-    return()
-endif()
-
-set(ANNOY_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/annoy")
-set(ANNOY_SOURCE_DIR "${ANNOY_PROJECT_DIR}/src")
-
-add_library(_annoy INTERFACE)
-target_include_directories(_annoy SYSTEM INTERFACE ${ANNOY_SOURCE_DIR})
-
-add_library(ch_contrib::annoy ALIAS _annoy)
-target_compile_definitions(_annoy INTERFACE ENABLE_ANNOY)
-target_compile_definitions(_annoy INTERFACE ANNOYLIB_MULTITHREADED_BUILD)

contrib/libprotobuf-mutator

@@ -1 +1 @@
-Subproject commit 1f95f8083066f5b38fd2db172e7e7f9aa7c49d2d
+Subproject commit b922c8ab9004ef9944982e4f165e2747b13223fa

contrib/libunwind

@@ -1 +1 @@
-Subproject commit a89d904befea07814628c6ce0b44083c4e149c62
+Subproject commit 601db0b0e03018c01710470a37703b618f9cf08b

contrib/usearch-cmake/CMakeLists.txt

@@ -1,9 +1,7 @@
-set(USEARCH_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/usearch")
-set(USEARCH_SOURCE_DIR "${USEARCH_PROJECT_DIR}/include")
-
 set(FP16_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/FP16")
 set(ROBIN_MAP_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/robin-map")
-set(SIMSIMD_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/SimSIMD-map")
+set(SIMSIMD_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/SimSIMD")
+set(USEARCH_PROJECT_DIR "${ClickHouse_SOURCE_DIR}/contrib/usearch")
 
 add_library(_usearch INTERFACE)
 
@@ -11,7 +9,6 @@ target_include_directories(_usearch SYSTEM INTERFACE
     ${FP16_PROJECT_DIR}/include
     ${ROBIN_MAP_PROJECT_DIR}/include
     ${SIMSIMD_PROJECT_DIR}/include
-    ${USEARCH_SOURCE_DIR})
+    ${USEARCH_PROJECT_DIR}/include)
 
 add_library(ch_contrib::usearch ALIAS _usearch)
-target_compile_definitions(_usearch INTERFACE ENABLE_USEARCH)

docker/packager/binary-builder/build.sh

@@ -108,7 +108,8 @@ if [ -n "$MAKE_DEB" ]; then
   bash -x /build/packages/build
 fi
 
-mv ./programs/clickhouse* /output || mv ./programs/*_fuzzer /output
+mv ./programs/clickhouse* /output ||:
+mv ./programs/*_fuzzer /output ||:
 [ -x ./programs/self-extracting/clickhouse ] && mv ./programs/self-extracting/clickhouse /output
 [ -x ./programs/self-extracting/clickhouse-stripped ] && mv ./programs/self-extracting/clickhouse-stripped /output
 [ -x ./programs/self-extracting/clickhouse-keeper ] && mv ./programs/self-extracting/clickhouse-keeper /output

docker/test/stateless/setup_minio.sh

@@ -99,10 +99,9 @@ upload_data() {
   # iterating over globs will cause redundant file variable to be
   # a path to a file, not a filename
   # shellcheck disable=SC2045
-  for file in $(ls "${data_path}"); do
-    echo "${file}";
-    ./mc cp "${data_path}"/"${file}" clickminio/test/"${file}";
-  done
+  if [ -d "${data_path}" ]; then
+    ./mc cp --recursive "${data_path}"/ clickminio/test/
+  fi
 }
 
 setup_aws_credentials() {

docs/changelogs/v24.3.7.30-lts.md

@@ -0,0 +1,29 @@
+---
+sidebar_position: 1
+sidebar_label: 2024
+---
+
+# 2024 Changelog
+
+### ClickHouse release v24.3.7.30-lts (c8a28cf4331) FIXME as compared to v24.3.6.48-lts (b2d33c3c45d)
+
+#### Improvement
+* Backported in [#68103](https://github.com/ClickHouse/ClickHouse/issues/68103): Distinguish booleans and integers while parsing values for custom settings: ``` SET custom_a = true; SET custom_b = 1; ```. [#62206](https://github.com/ClickHouse/ClickHouse/pull/62206) ([Vitaly Baranov](https://github.com/vitlibar)).
+
+#### Bug Fix (user-visible misbehavior in an official stable release)
+* Backported in [#67931](https://github.com/ClickHouse/ClickHouse/issues/67931): Fixing the `Not-ready Set` error after the `PREWHERE` optimization for StorageMerge. [#65057](https://github.com/ClickHouse/ClickHouse/pull/65057) ([Nikolai Kochetov](https://github.com/KochetovNicolai)).
+* Backported in [#68062](https://github.com/ClickHouse/ClickHouse/issues/68062): Fix boolean literals in query sent to external database (for engines like `PostgreSQL`). [#66282](https://github.com/ClickHouse/ClickHouse/pull/66282) ([vdimir](https://github.com/vdimir)).
+* Backported in [#67812](https://github.com/ClickHouse/ClickHouse/issues/67812): Only relevant to the experimental Variant data type. Fix crash with Variant + AggregateFunction type. [#67122](https://github.com/ClickHouse/ClickHouse/pull/67122) ([Kruglov Pavel](https://github.com/Avogar)).
+* Backported in [#67848](https://github.com/ClickHouse/ClickHouse/issues/67848): Fixes [#66026](https://github.com/ClickHouse/ClickHouse/issues/66026). Avoid unresolved table function arguments traversal in `ReplaceTableNodeToDummyVisitor`. [#67522](https://github.com/ClickHouse/ClickHouse/pull/67522) ([Dmitry Novik](https://github.com/novikd)).
+* Backported in [#68271](https://github.com/ClickHouse/ClickHouse/issues/68271): Fix inserting into stream like engines (Kafka, RabbitMQ, NATS) through HTTP interface. [#67554](https://github.com/ClickHouse/ClickHouse/pull/67554) ([János Benjamin Antal](https://github.com/antaljanosbenjamin)).
+* Backported in [#67806](https://github.com/ClickHouse/ClickHouse/issues/67806): Fix reloading SQL UDFs with UNION. Previously, restarting the server could make UDF invalid. [#67665](https://github.com/ClickHouse/ClickHouse/pull/67665) ([Antonio Andelic](https://github.com/antonio2368)).
+* Backported in [#67834](https://github.com/ClickHouse/ClickHouse/issues/67834): Fix potential stack overflow in `JSONMergePatch` function. Renamed this function from `jsonMergePatch` to `JSONMergePatch` because the previous name was wrong. The previous name is still kept for compatibility. Improved diagnostic of errors in the function. This closes [#67304](https://github.com/ClickHouse/ClickHouse/issues/67304). [#67756](https://github.com/ClickHouse/ClickHouse/pull/67756) ([Alexey Milovidov](https://github.com/alexey-milovidov)).
+* Backported in [#68206](https://github.com/ClickHouse/ClickHouse/issues/68206): Fix wrong `count()` result when there is non-deterministic function in predicate. [#67922](https://github.com/ClickHouse/ClickHouse/pull/67922) ([János Benjamin Antal](https://github.com/antaljanosbenjamin)).
+* Backported in [#68089](https://github.com/ClickHouse/ClickHouse/issues/68089): Fixed the calculation of the maximum thread soft limit in containerized environments where the usable CPU count is limited. [#67963](https://github.com/ClickHouse/ClickHouse/pull/67963) ([Robert Schulze](https://github.com/rschu1ze)).
+* Backported in [#68120](https://github.com/ClickHouse/ClickHouse/issues/68120): Fixed skipping of untouched parts in mutations with new analyzer. Previously with enabled analyzer data in part could be rewritten by mutation even if mutation doesn't affect this part according to predicate. [#68052](https://github.com/ClickHouse/ClickHouse/pull/68052) ([Anton Popov](https://github.com/CurtizJ)).
+
+#### NOT FOR CHANGELOG / INSIGNIFICANT
+
+* Update version after release. [#67676](https://github.com/ClickHouse/ClickHouse/pull/67676) ([robot-clickhouse](https://github.com/robot-clickhouse)).
+* Backported in [#68074](https://github.com/ClickHouse/ClickHouse/issues/68074): Add an explicit error for `ALTER MODIFY SQL SECURITY` on non-view tables. [#67953](https://github.com/ClickHouse/ClickHouse/pull/67953) ([pufit](https://github.com/pufit)).
+

docs/en/engines/table-engines/mergetree-family/annindexes.md

@@ -17,7 +17,7 @@ In terms of SQL, the nearest neighborhood problem can be expressed as follows:
 
 ``` sql
 SELECT *
-FROM table_with_ann_index
+FROM table
 ORDER BY Distance(vectors, Point)
 LIMIT N
 ```
@@ -27,75 +27,109 @@ Function `Distance` computes the distance between two vectors. Often, the Euclid
 distance functions](/docs/en/sql-reference/functions/distance-functions.md) are also possible. `Point` is the reference point, e.g. `(0.17,
 0.33, ...)`, and `N` limits the number of search results.
 
-An alternative formulation of the nearest neighborhood search problem looks as follows:
+This query returns the top-`N` closest points to the reference point. Parameter `N` limits the number of returned values which is useful for
+situations where `MaxDistance` is difficult to determine in advance.
 
-``` sql
-SELECT *
-FROM table_with_ann_index
-WHERE Distance(vectors, Point) < MaxDistance
-LIMIT N
-```
-
-While the first query returns the top-`N` closest points to the reference point, the second query returns all points closer to the reference
-point than a maximally allowed radius `MaxDistance`. Parameter `N` limits the number of returned values which is useful for situations where
-`MaxDistance` is difficult to determine in advance.
-
-With brute force search, both queries are expensive (linear in the number of points) because the distance between all points in `vectors` and
+With brute force search, the query is expensive (linear in the number of points) because the distance between all points in `vectors` and
 `Point` must be computed. To speed this process up, Approximate Nearest Neighbor Search Indexes (ANN indexes) store a compact representation
 of the search space (using clustering, search trees, etc.) which allows to compute an approximate answer much quicker (in sub-linear time).
 
-# Creating and Using ANN Indexes {#creating_using_ann_indexes}
+# Creating and Using Vector Similarity Indexes
 
-Syntax to create an ANN index over an [Array(Float32)](../../../sql-reference/data-types/array.md) column:
+Syntax to create a vector similarity index over an [Array(Float32)](../../../sql-reference/data-types/array.md) column:
 
 ```sql
-CREATE TABLE table_with_ann_index
+CREATE TABLE table
 (
-  `id` Int64,
-  `vectors` Array(Float32),
-  INDEX [ann_index_name vectors TYPE [ann_index_type]([ann_index_parameters]) [GRANULARITY [N]]
+  id Int64,
+  vectors Array(Float32),
+  INDEX index_name vectors TYPE vector_similarity(method, distance_function[, quantization, connectivity, expansion_add, expansion_search]) [GRANULARITY N]
 )
 ENGINE = MergeTree
 ORDER BY id;
 ```
 
+Parameters:
+- `method`: Supports currently only `hnsw`.
+- `distance_function`: either `L2Distance` (the [Euclidean distance](https://en.wikipedia.org/wiki/Euclidean_distance) - the length of a
+  line between two points in Euclidean space), or `cosineDistance` (the [cosine
+  distance](https://en.wikipedia.org/wiki/Cosine_similarity#Cosine_distance)- the angle between two non-zero vectors).
+- `quantization`: either `f32`, `f16`, or `i8` for storing the vector with reduced precision (optional, default: `f32`)
+- `m`: the number of neighbors per graph node (optional, default: 16)
+- `ef_construction`: (optional, default: 128)
+- `ef_search`: (optional, default: 64)
+
+Example:
+
+```sql
+CREATE TABLE table
+(
+  id Int64,
+  vectors Array(Float32),
+  INDEX idx vectors TYPE vector_similarity('hnsw', 'L2Distance') -- Alternative syntax: TYPE vector_similarity(hnsw, L2Distance)
+)
+ENGINE = MergeTree
+ORDER BY id;
+```
+
+Vector similarity indexes are based on the [USearch library](https://github.com/unum-cloud/usearch), which implements the [HNSW
+algorithm](https://arxiv.org/abs/1603.09320), i.e., a hierarchical graph where each point represents a vector and the edges represent
+similarity. Such hierarchical structures can be very efficient on large collections. They may often fetch 0.05% or less data from the
+overall dataset, while still providing 99% recall. This is especially useful when working with high-dimensional vectors, that are expensive
+to load and compare. The library also has several hardware-specific SIMD optimizations to accelerate further distance computations on modern
+Arm (NEON and SVE) and x86 (AVX2 and AVX-512) CPUs and OS-specific optimizations to allow efficient navigation around immutable persistent
+files, without loading them into RAM.
+
+USearch indexes are currently experimental, to use them you first need to `SET allow_experimental_vector_similarity_index = 1`.
+
+Vector similarity indexes currently support two distance functions:
+- `L2Distance`, also called Euclidean distance, is the length of a line segment between two points in Euclidean space
+  ([Wikipedia](https://en.wikipedia.org/wiki/Euclidean_distance)).
+- `cosineDistance`, also called cosine similarity, is the cosine of the angle between two (non-zero) vectors
+  ([Wikipedia](https://en.wikipedia.org/wiki/Cosine_similarity)).
+
+Vector similarity indexes allows storing the vectors in reduced precision formats. Supported scalar kinds are `f64`, `f32`, `f16` or `i8`.
+If no scalar kind was specified during index creation, `f16` is used as default.
+
+For normalized data, `L2Distance` is usually a better choice, otherwise `cosineDistance` is recommended to compensate for scale. If no
+distance function was specified during index creation, `L2Distance` is used as default.
+
+:::note
+All arrays must have same length. To avoid errors, you can use a
+[CONSTRAINT](/docs/en/sql-reference/statements/create/table.md#constraints), for example, `CONSTRAINT constraint_name_1 CHECK
+length(vectors) = 256`. Also, empty `Arrays` and unspecified `Array` values in INSERT statements (i.e. default values) are not supported.
+:::
+
+:::note
+The vector similarity index currently does not work with per-table, non-default `index_granularity` settings (see
+[here](https://github.com/ClickHouse/ClickHouse/pull/51325#issuecomment-1605920475)). If necessary, the value must be changed in config.xml.
+:::
+
 ANN indexes are built during column insertion and merge. As a result, `INSERT` and `OPTIMIZE` statements will be slower than for ordinary
 tables. ANNIndexes are ideally used only with immutable or rarely changed data, respectively when are far more read requests than write
 requests.
 
-ANN indexes support two types of queries:
-
-- ORDER BY queries:
+ANN indexes support these queries:
 
   ``` sql
   SELECT *
-  FROM table_with_ann_index
+  FROM table
   [WHERE ...]
   ORDER BY Distance(vectors, Point)
   LIMIT N
   ```
 
-- WHERE queries:
-
-   ``` sql
-   SELECT *
-   FROM table_with_ann_index
-   WHERE Distance(vectors, Point) < MaxDistance
-   LIMIT N
-   ```
-
 :::tip
 To avoid writing out large vectors, you can use [query
 parameters](/docs/en/interfaces/cli.md#queries-with-parameters-cli-queries-with-parameters), e.g.
 
 ```bash
-clickhouse-client --param_vec='hello' --query="SELECT * FROM table_with_ann_index WHERE L2Distance(vectors, {vec: Array(Float32)}) < 1.0"
+clickhouse-client --param_vec='hello' --query="SELECT * FROM table WHERE L2Distance(vectors, {vec: Array(Float32)}) < 1.0"
 ```
 :::
 
-**Restrictions**: Queries that contain both a `WHERE Distance(vectors, Point) < MaxDistance` and an `ORDER BY Distance(vectors, Point)`
-clause cannot use ANN indexes. Also, the approximate algorithms used to determine the nearest neighbors require a limit, hence queries
-without `LIMIT` clause cannot utilize ANN indexes. Also, ANN indexes are only used if the query has a `LIMIT` value smaller than setting
+**Restrictions**: Approximate algorithms used to determine the nearest neighbors require a limit, hence queries without `LIMIT` clause
+cannot utilize ANN indexes. Also, ANN indexes are only used if the query has a `LIMIT` value smaller than setting
 `max_limit_for_ann_queries` (default: 1 million rows). This is a safeguard to prevent large memory allocations by external libraries for
 approximate neighbor search.
 
@@ -122,128 +156,3 @@ brute-force distance calculation over all rows of the granules. With a small `GR
 equally good, only the processing performance differs. It is generally recommended to use a large `GRANULARITY` for ANN indexes and fall
 back to a smaller `GRANULARITY` values only in case of problems like excessive memory consumption of the ANN structures. If no `GRANULARITY`
 was specified for ANN indexes, the default value is 100 million.
-
-
-# Available ANN Indexes {#available_ann_indexes}
-
-- [Annoy](/docs/en/engines/table-engines/mergetree-family/annindexes.md#annoy-annoy)
-
-- [USearch](/docs/en/engines/table-engines/mergetree-family/annindexes.md#usearch-usearch)
-
-## Annoy {#annoy}
-
-Annoy indexes are currently experimental, to use them you first need to `SET allow_experimental_annoy_index = 1`. They are also currently
-disabled on ARM due to memory safety problems with the algorithm.
-
-This type of ANN index is based on the [Annoy library](https://github.com/spotify/annoy) which recursively divides the space into random
-linear surfaces (lines in 2D, planes in 3D etc.).
-
-<div class='vimeo-container'>
-  <iframe src="//www.youtube.com/embed/QkCCyLW0ehU"
-    width="640"
-    height="360"
-    frameborder="0"
-    allow="autoplay;
-    fullscreen;
-    picture-in-picture"
-    allowfullscreen>
-  </iframe>
-</div>
-
-Syntax to create an Annoy index over an [Array(Float32)](../../../sql-reference/data-types/array.md) column:
-
-```sql
-CREATE TABLE table_with_annoy_index
-(
-  id Int64,
-  vectors Array(Float32),
-  INDEX [ann_index_name] vectors TYPE annoy([Distance[, NumTrees]]) [GRANULARITY N]
-)
-ENGINE = MergeTree
-ORDER BY id;
-```
-
-Annoy currently supports two distance functions:
-- `L2Distance`, also called Euclidean distance, is the length of a line segment between two points in Euclidean space
-  ([Wikipedia](https://en.wikipedia.org/wiki/Euclidean_distance)).
-- `cosineDistance`, also called cosine similarity, is the cosine of the angle between two (non-zero) vectors
-  ([Wikipedia](https://en.wikipedia.org/wiki/Cosine_similarity)).
-
-For normalized data, `L2Distance` is usually a better choice, otherwise `cosineDistance` is recommended to compensate for scale. If no
-distance function was specified during index creation, `L2Distance` is used as default.
-
-Parameter `NumTrees` is the number of trees which the algorithm creates (default if not specified: 100). Higher values of `NumTree` mean
-more accurate search results but slower index creation / query times (approximately linearly) as well as larger index sizes.
-
-:::note
-All arrays must have same length. To avoid errors, you can use a
-[CONSTRAINT](/docs/en/sql-reference/statements/create/table.md#constraints), for example, `CONSTRAINT constraint_name_1 CHECK
-length(vectors) = 256`. Also, empty `Arrays` and unspecified `Array` values in INSERT statements (i.e. default values) are not supported.
-:::
-
-The creation of Annoy indexes (whenever a new part is build, e.g. at the end of a merge) is a relatively slow process. You can increase
-setting `max_threads_for_annoy_index_creation` (default: 4) which controls how many threads are used to create an Annoy index. Please be
-careful with this setting, it is possible that multiple indexes are created in parallel in which case there can be overparallelization.
-
-Setting `annoy_index_search_k_nodes` (default: `NumTrees * LIMIT`) determines how many tree nodes are inspected during SELECTs. Larger
-values mean more accurate results at the cost of longer query runtime:
-
-```sql
-SELECT *
-FROM table_name
-ORDER BY L2Distance(vectors, Point)
-LIMIT N
-SETTINGS annoy_index_search_k_nodes=100;
-```
-
-:::note
-The Annoy index currently does not work with per-table, non-default `index_granularity` settings (see
-[here](https://github.com/ClickHouse/ClickHouse/pull/51325#issuecomment-1605920475)). If necessary, the value must be changed in config.xml.
-:::
-
-## USearch {#usearch}
-
-This type of ANN index is based on the [USearch library](https://github.com/unum-cloud/usearch), which implements the [HNSW
-algorithm](https://arxiv.org/abs/1603.09320), i.e., builds a hierarchical graph where each point represents a vector and the edges represent
-similarity. Such hierarchical structures can be very efficient on large collections. They may often fetch 0.05% or less data from the
-overall dataset, while still providing 99% recall. This is especially useful when working with high-dimensional vectors,
-that are expensive to load and compare. The library also has several hardware-specific SIMD optimizations to accelerate further
-distance computations on modern Arm (NEON and SVE) and x86 (AVX2 and AVX-512) CPUs and OS-specific optimizations to allow efficient
-navigation around immutable persistent files, without loading them into RAM.
-
-<div class='vimeo-container'>
-  <iframe src="//www.youtube.com/embed/UMrhB3icP9w"
-    width="640"
-    height="360"
-    frameborder="0"
-    allow="autoplay;
-    fullscreen;
-    picture-in-picture"
-    allowfullscreen>
-  </iframe>
-</div>
-
-Syntax to create an USearch index over an [Array](../../../sql-reference/data-types/array.md) column:
-
-```sql
-CREATE TABLE table_with_usearch_index
-(
-  id Int64,
-  vectors Array(Float32),
-  INDEX [ann_index_name] vectors TYPE usearch([Distance[, ScalarKind]]) [GRANULARITY N]
-)
-ENGINE = MergeTree
-ORDER BY id;
-```
-
-USearch currently supports two distance functions:
-- `L2Distance`, also called Euclidean distance, is the length of a line segment between two points in Euclidean space
-  ([Wikipedia](https://en.wikipedia.org/wiki/Euclidean_distance)).
-- `cosineDistance`, also called cosine similarity, is the cosine of the angle between two (non-zero) vectors
-  ([Wikipedia](https://en.wikipedia.org/wiki/Cosine_similarity)).
-
-USearch allows storing the vectors in reduced precision formats. Supported scalar kinds are `f64`, `f32`, `f16` or `i8`. If no scalar kind
-was specified during index creation, `f16` is used as default.
-
-For normalized data, `L2Distance` is usually a better choice, otherwise `cosineDistance` is recommended to compensate for scale. If no
-distance function was specified during index creation, `L2Distance` is used as default.

docs/en/operations/query-cache.md

@@ -143,6 +143,18 @@ value can be specified at session, profile or query level using setting [query_c
 Entries in the query cache are compressed by default. This reduces the overall memory consumption at the cost of slower writes into / reads
 from the query cache. To disable compression, use setting [query_cache_compress_entries](settings/settings.md#query-cache-compress-entries).
 
+Sometimes it is useful to keep multiple results for the same query cached. This can be achieved using setting
+[query_cache_tag](settings/settings.md#query-cache-tag) that acts as as a label (or namespace) for a query cache entries. The query cache
+considers results of the same query with different tags different.
+
+Example for creating three different query cache entries for the same query:
+
+```sql
+SELECT 1 SETTINGS use_query_cache = true; -- query_cache_tag is implicitly '' (empty string)
+SELECT 1 SETTINGS use_query_cache = true, query_cache_tag = 'tag 1';
+SELECT 1 SETTINGS use_query_cache = true, query_cache_tag = 'tag 2';
+```
+
 ClickHouse reads table data in blocks of [max_block_size](settings/settings.md#setting-max_block_size) rows. Due to filtering, aggregation,
 etc., result blocks are typically much smaller than 'max_block_size' but there are also cases where they are much bigger. Setting
 [query_cache_squash_partial_results](settings/settings.md#query-cache-squash-partial-results) (enabled by default) controls if result blocks

docs/en/operations/server-configuration-parameters/settings.md

@@ -1400,6 +1400,16 @@ The number of seconds that ClickHouse waits for incoming requests before closing
 <keep_alive_timeout>10</keep_alive_timeout>
 ```
 
+## max_keep_alive_requests {#max-keep-alive-requests}
+
+Maximal number of requests through a single keep-alive connection until it will be closed by ClickHouse server. Default to 10000.
+
+**Example**
+
+``` xml
+<max_keep_alive_requests>10</max_keep_alive_requests>
+```
+
 ## listen_host {#listen_host}
 
 Restriction on hosts that requests can come from. If you want the server to answer all of them, specify `::`.

docs/en/operations/settings/merge-tree-settings.md

@@ -1041,3 +1041,14 @@ Compression rates of LZ4 or ZSTD improve on average by 20-40%.
 
 This setting works best for tables with no primary key or a low-cardinality primary key, i.e. a table with only few distinct primary key values.
 High-cardinality primary keys, e.g. involving timestamp columns of type `DateTime64`, are not expected to benefit from this setting.
+
+### deduplicate_merge_projection_mode
+
+Whether to allow create projection for the table with non-classic MergeTree, that is not (Replicated, Shared) MergeTree. If allowed, what is the action when merge projections, either drop or rebuild. So classic MergeTree would ignore this setting.
+It also controls `OPTIMIZE DEDUPLICATE` as well, but has effect on all MergeTree family members.
+
+Possible values:
+
+- throw, drop, rebuild
+
+Default value: throw

docs/en/operations/settings/settings.md

@@ -1800,6 +1800,17 @@ Possible values:
 
 Default value: `0`.
 
+## query_cache_tag {#query-cache-tag}
+
+A string which acts as a label for [query cache](../query-cache.md) entries.
+The same queries with different tags are considered different by the query cache.
+
+Possible values:
+
+- Any string
+
+Default value: `''`
+
 ## query_cache_max_size_in_bytes {#query-cache-max-size-in-bytes}
 
 The maximum amount of memory (in bytes) the current user may allocate in the [query cache](../query-cache.md). 0 means unlimited.
@@ -5627,6 +5638,12 @@ Disable all insert and mutations (alter table update / alter table delete / alte
 
 Default value: `false`.
 
+## use_hive_partitioning
+
+When enabled, ClickHouse will detect Hive-style partitioning in path (`/name=value/`) in file-like table engines [File](../../engines/table-engines/special/file.md#hive-style-partitioning)/[S3](../../engines/table-engines/integrations/s3.md#hive-style-partitioning)/[URL](../../engines/table-engines/special/url.md#hive-style-partitioning)/[HDFS](../../engines/table-engines/integrations/hdfs.md#hive-style-partitioning)/[AzureBlobStorage](../../engines/table-engines/integrations/azureBlobStorage.md#hive-style-partitioning) and will allow to use partition columns as virtual columns in the query. These virtual columns will have the same names as in the partitioned path, but starting with `_`.
+
+Default value: `false`.
+
 ## allow_experimental_time_series_table {#allow-experimental-time-series-table}
 
 Allows creation of tables with the [TimeSeries](../../engines/table-engines/integrations/time-series.md) table engine.

docs/en/operations/system-tables/query_cache.md

@@ -9,6 +9,7 @@ Columns:
 
 - `query` ([String](../../sql-reference/data-types/string.md)) — Query string.
 - `result_size` ([UInt64](../../sql-reference/data-types/int-uint.md#uint-ranges)) — Size of the query cache entry.
+- `tag` ([LowCardinality(String)](../../sql-reference/data-types/lowcardinality.md)) — Tag of the query cache entry.
 - `stale` ([UInt8](../../sql-reference/data-types/int-uint.md)) — If the query cache entry is stale.
 - `shared` ([UInt8](../../sql-reference/data-types/int-uint.md)) — If the query cache entry is shared between multiple users.
 - `compressed` ([UInt8](../../sql-reference/data-types/int-uint.md)) — If the query cache entry is compressed.
@@ -26,6 +27,7 @@ Row 1:
 ──────
 query:       SELECT 1 SETTINGS use_query_cache = 1
 result_size: 128
+tag:
 stale:       0
 shared:      0
 compressed:  1

docs/en/sql-reference/data-types/dynamic.md

@@ -14,7 +14,7 @@ To declare a column of `Dynamic` type, use the following syntax:
 <column_name> Dynamic(max_types=N)
 ```
 
-Where `N` is an optional parameter between `1` and `255` indicating how many different data types can be stored inside a column with type `Dynamic` across single block of data that is stored separately (for example across single data part for MergeTree table). If this limit is exceeded, all new types will be converted to type `String`. Default value of `max_types` is `32`.
+Where `N` is an optional parameter between `0` and `254` indicating how many different data types can be stored as separate subcolumns inside a column with type `Dynamic` across single block of data that is stored separately (for example across single data part for MergeTree table). If this limit is exceeded, all values with new types will be stored together in a special shared data structure in binary form. Default value of `max_types` is `32`.
 
 :::note
 The Dynamic data type is an experimental feature. To use it, set `allow_experimental_dynamic_type = 1`.
@@ -224,41 +224,43 @@ SELECT d::Dynamic(max_types=5) as d2, dynamicType(d2) FROM test;
 └───────┴────────────────┘
 ```
 
-If `K < N`, then the values with the rarest types are converted to `String`:
+If `K < N`, then the values with the rarest types will be inserted into a single special subcolumn, but still will be accessible:
 ```text
 CREATE TABLE test (d Dynamic(max_types=4)) ENGINE = Memory;
 INSERT INTO test VALUES (NULL), (42), (43), ('42.42'), (true), ([1, 2, 3]);
-SELECT d, dynamicType(d), d::Dynamic(max_types=2) as d2, dynamicType(d2) FROM test;
+SELECT d, dynamicType(d), d::Dynamic(max_types=2) as d2, dynamicType(d2), isDynamicElementInSharedData(d2) FROM test;
 ```
 
 ```text
-┌─d───────┬─dynamicType(d)─┬─d2──────┬─dynamicType(d2)─┐
-│ ᴺᵁᴸᴸ    │ None           │ ᴺᵁᴸᴸ    │ None            │
-│ 42      │ Int64          │ 42      │ Int64           │
-│ 43      │ Int64          │ 43      │ Int64           │
-│ 42.42   │ String         │ 42.42   │ String          │
-│ true    │ Bool           │ true    │ String          │
-│ [1,2,3] │ Array(Int64)   │ [1,2,3] │ String          │
-└─────────┴────────────────┴─────────┴─────────────────┘
+┌─d───────┬─dynamicType(d)─┬─d2──────┬─dynamicType(d2)─┬─isDynamicElementInSharedData(d2)─┐
+│ ᴺᵁᴸᴸ    │ None           │ ᴺᵁᴸᴸ    │ None            │ false                            │
+│ 42      │ Int64          │ 42      │ Int64           │ false                            │
+│ 43      │ Int64          │ 43      │ Int64           │ false                            │
+│ 42.42   │ String         │ 42.42   │ String          │ false                            │
+│ true    │ Bool           │ true    │ Bool            │ true                             │
+│ [1,2,3] │ Array(Int64)   │ [1,2,3] │ Array(Int64)    │ true                             │
+└─────────┴────────────────┴─────────┴─────────────────┴──────────────────────────────────┘
 ```
 
-If `K=1`, all types are converted to `String`:
+Functions `isDynamicElementInSharedData` returns `true` for rows that are stored in a special shared data structure inside `Dynamic` and as we can see, resulting column contains only 2 types that are not stored in shared data structure.
+
+If `K=0`, all types will be inserted into single special subcolumn:
 
 ```text
 CREATE TABLE test (d Dynamic(max_types=4)) ENGINE = Memory;
 INSERT INTO test VALUES (NULL), (42), (43), ('42.42'), (true), ([1, 2, 3]);
-SELECT d, dynamicType(d), d::Dynamic(max_types=1) as d2, dynamicType(d2) FROM test;
+SELECT d, dynamicType(d), d::Dynamic(max_types=0) as d2, dynamicType(d2), isDynamicElementInSharedData(d2) FROM test;
 ```
 
 ```text
-┌─d───────┬─dynamicType(d)─┬─d2──────┬─dynamicType(d2)─┐
-│ ᴺᵁᴸᴸ    │ None           │ ᴺᵁᴸᴸ    │ None            │
-│ 42      │ Int64          │ 42      │ String          │
-│ 43      │ Int64          │ 43      │ String          │
-│ 42.42   │ String         │ 42.42   │ String          │
-│ true    │ Bool           │ true    │ String          │
-│ [1,2,3] │ Array(Int64)   │ [1,2,3] │ String          │
-└─────────┴────────────────┴─────────┴─────────────────┘
+┌─d───────┬─dynamicType(d)─┬─d2──────┬─dynamicType(d2)─┬─isDynamicElementInSharedData(d2)─┐
+│ ᴺᵁᴸᴸ    │ None           │ ᴺᵁᴸᴸ    │ None            │ false                            │
+│ 42      │ Int64          │ 42      │ Int64           │ true                             │
+│ 43      │ Int64          │ 43      │ Int64           │ true                             │
+│ 42.42   │ String         │ 42.42   │ String          │ true                             │
+│ true    │ Bool           │ true    │ Bool            │ true                             │
+│ [1,2,3] │ Array(Int64)   │ [1,2,3] │ Array(Int64)    │ true                             │
+└─────────┴────────────────┴─────────┴─────────────────┴──────────────────────────────────┘
 ```
 
 ## Reading Dynamic type from the data
@@ -411,17 +413,17 @@ SELECT d, dynamicType(d) FROM test ORDER by d;
 
 ## Reaching the limit in number of different data types stored inside Dynamic
 
-`Dynamic` data type can store only limited number of different data types inside. By default, this limit is 32, but you can change it in type declaration using syntax `Dynamic(max_types=N)` where N is between 1 and 255 (due to implementation details, it's impossible to have more than 255 different data types inside Dynamic).
-When the limit is reached, all new data types inserted to `Dynamic` column will be casted to `String` and stored as `String` values.
+`Dynamic` data type can store only limited number of different data types as separate subcolumns. By default, this limit is 32, but you can change it in type declaration using syntax `Dynamic(max_types=N)` where N is between 0 and 254 (due to implementation details, it's impossible to have more than 254 different data types that can be stored as separate subcolumns inside Dynamic).
+When the limit is reached, all new data types inserted to `Dynamic` column will be inserted into a single shared data structure that stores values with different data types in binary form.
 
 Let's see what happens when the limit is reached in different scenarios.
 
 ### Reaching the limit during data parsing
 
-During parsing of `Dynamic` values from the data, when the limit is reached for current block of data, all new values will be inserted as `String` values:
+During parsing of `Dynamic` values from the data, when the limit is reached for current block of data, all new values will be inserted into shared data structure:
 
 ```sql
-SELECT d, dynamicType(d) FROM format(JSONEachRow, 'd Dynamic(max_types=3)', '
+SELECT d, dynamicType(d), isDynamicElementInSharedData(d) FROM format(JSONEachRow, 'd Dynamic(max_types=3)', '
 {"d" : 42}
 {"d" : [1, 2, 3]}
 {"d" : "Hello, World!"}
@@ -432,22 +434,22 @@ SELECT d, dynamicType(d) FROM format(JSONEachRow, 'd Dynamic(max_types=3)', '
 ```
 
 ```text
-┌─d──────────────────────────┬─dynamicType(d)─┐
-│ 42                         │ Int64          │
-│ [1,2,3]                    │ Array(Int64)   │
-│ Hello, World!              │ String         │
-│ 2020-01-01                 │ String         │
-│ ["str1", "str2", "str3"]   │ String         │
-│ {"a" : 1, "b" : [1, 2, 3]} │ String         │
-└────────────────────────────┴────────────────┘
+┌─d──────────────────────┬─dynamicType(d)─────────────────┬─isDynamicElementInSharedData(d)─┐
+│ 42                     │ Int64                          │ false                           │
+│ [1,2,3]                │ Array(Int64)                   │ false                           │
+│ Hello, World!          │ String                         │ false                           │
+│ 2020-01-01             │ Date                           │ true                            │
+│ ['str1','str2','str3'] │ Array(String)                  │ true                            │
+│ (1,[1,2,3])            │ Tuple(a Int64, b Array(Int64)) │ true                            │
+└────────────────────────┴────────────────────────────────┴─────────────────────────────────┘
 ```
 
-As we can see, after inserting 3 different data types `Int64`, `Array(Int64)` and `String` all new types were converted to `String`.
+As we can see, after inserting 3 different data types `Int64`, `Array(Int64)` and `String` all new types were inserted into special shared data structure.
 
 ### During merges of data parts in MergeTree table engines
 
-During merge of several data parts in MergeTree table the `Dynamic` column in the resulting data part can reach the limit of different data types inside and won't be able to store all types from source parts.
-In this case ClickHouse chooses what types will remain after merge and what types will be casted to `String`.  In most cases ClickHouse tries to keep the most frequent types and cast the rarest types to `String`, but it depends on the implementation.
+During merge of several data parts in MergeTree table the `Dynamic` column in the resulting data part can reach the limit of different data types that can be stored in separate subcolumns inside and won't be able to store all types as subcolumns from source parts.
+In this case ClickHouse chooses what types will remain as separate subcolumns after merge and what types will be inserted into shared data structure. In most cases ClickHouse tries to keep the most frequent types and store the rarest types in shared data structure, but it depends on the implementation.
 
 Let's see an example of such merge. First, let's create a table with `Dynamic` column, set the limit of different data types to `3` and insert values with `5` different types:
 
@@ -463,17 +465,17 @@ INSERT INTO test SELECT number, 'str_' || toString(number) FROM numbers(1);
 
 Each insert will create a separate data pert with `Dynamic` column containing single type:
 ```sql
-SELECT count(), dynamicType(d), _part FROM test GROUP BY _part, dynamicType(d) ORDER BY _part;
+SELECT count(), dynamicType(d), isDynamicElementInSharedData(d), _part FROM test GROUP BY _part, dynamicType(d), isDynamicElementInSharedData(d) ORDER BY _part, count();
 ```
 
 ```text
-┌─count()─┬─dynamicType(d)──────┬─_part─────┐
-│       5 │ UInt64              │ all_1_1_0 │
-│       4 │ Array(UInt64)       │ all_2_2_0 │
-│       3 │ Date                │ all_3_3_0 │
-│       2 │ Map(UInt64, UInt64) │ all_4_4_0 │
-│       1 │ String              │ all_5_5_0 │
-└─────────┴─────────────────────┴───────────┘
+┌─count()─┬─dynamicType(d)──────┬─isDynamicElementInSharedData(d)─┬─_part─────┐
+│       5 │ UInt64              │ false                           │ all_1_1_0 │
+│       4 │ Array(UInt64)       │ false                           │ all_2_2_0 │
+│       3 │ Date                │ false                           │ all_3_3_0 │
+│       2 │ Map(UInt64, UInt64) │ false                           │ all_4_4_0 │
+│       1 │ String              │ false                           │ all_5_5_0 │
+└─────────┴─────────────────────┴─────────────────────────────────┴───────────┘
 ```
 
 Now, let's merge all parts into one and see what will happen:
@@ -481,18 +483,20 @@ Now, let's merge all parts into one and see what will happen:
 ```sql
 SYSTEM START MERGES test;
 OPTIMIZE TABLE test FINAL;
-SELECT count(), dynamicType(d), _part FROM test GROUP BY _part, dynamicType(d) ORDER BY _part;
+SELECT count(), dynamicType(d), isDynamicElementInSharedData(d), _part FROM test GROUP BY _part, dynamicType(d), isDynamicElementInSharedData(d) ORDER BY _part, count() desc;
 ```
 
 ```text
-┌─count()─┬─dynamicType(d)─┬─_part─────┐
-│       5 │ UInt64         │ all_1_5_2 │
-│       6 │ String         │ all_1_5_2 │
-│       4 │ Array(UInt64)  │ all_1_5_2 │
-└─────────┴────────────────┴───────────┘
+┌─count()─┬─dynamicType(d)──────┬─isDynamicElementInSharedData(d)─┬─_part─────┐
+│       5 │ UInt64              │ false                           │ all_1_5_2 │
+│       4 │ Array(UInt64)       │ false                           │ all_1_5_2 │
+│       3 │ Date                │ false                           │ all_1_5_2 │
+│       2 │ Map(UInt64, UInt64) │ true                            │ all_1_5_2 │
+│       1 │ String              │ true                            │ all_1_5_2 │
+└─────────┴─────────────────────┴─────────────────────────────────┴───────────┘
 ```
 
-As we can see, ClickHouse kept the most frequent types `UInt64` and `Array(UInt64)` and casted all other types to `String`.
+As we can see, ClickHouse kept the most frequent types `UInt64` and `Array(UInt64)` as subcolumns and inserted all other types into shared data.
 
 ## JSONExtract functions with Dynamic
 
@@ -509,22 +513,23 @@ SELECT JSONExtract('{"a" : [1, 2, 3]}', 'a', 'Dynamic') AS dynamic, dynamicType(
 ```
 
 ```sql
-SELECT JSONExtract('{"obj" : {"a" : 42, "b" : "Hello", "c" : [1,2,3]}}', 'obj', 'Map(String, Variant(UInt32, String, Array(UInt32)))') AS map_of_dynamics, mapApply((k, v) -> (k, variantType(v)), map_of_dynamics) AS map_of_dynamic_types```
+SELECT JSONExtract('{"obj" : {"a" : 42, "b" : "Hello", "c" : [1,2,3]}}', 'obj', 'Map(String, Dynamic)') AS map_of_dynamics, mapApply((k, v) -> (k, dynamicType(v)), map_of_dynamics) AS map_of_dynamic_types
+```
 
 ```text
-┌─map_of_dynamics──────────────────┬─map_of_dynamic_types────────────────────────────┐
-│ {'a':42,'b':'Hello','c':[1,2,3]} │ {'a':'UInt32','b':'String','c':'Array(UInt32)'} │
-└──────────────────────────────────┴─────────────────────────────────────────────────┘
+┌─map_of_dynamics──────────────────┬─map_of_dynamic_types────────────────────────────────────┐
+│ {'a':42,'b':'Hello','c':[1,2,3]} │ {'a':'Int64','b':'String','c':'Array(Nullable(Int64))'} │
+└──────────────────────────────────┴─────────────────────────────────────────────────────────┘
 ```
 
 ```sql
-SELECT JSONExtractKeysAndValues('{"a" : 42, "b" : "Hello", "c" : [1,2,3]}', 'Variant(UInt32, String, Array(UInt32))') AS dynamics, arrayMap(x -> (x.1, variantType(x.2)), dynamics) AS dynamic_types```
+SELECT JSONExtractKeysAndValues('{"a" : 42, "b" : "Hello", "c" : [1,2,3]}', 'Dynamic') AS dynamics, arrayMap(x -> (x.1, dynamicType(x.2)), dynamics) AS dynamic_types```
 ```
 
 ```text
-┌─dynamics───────────────────────────────┬─dynamic_types─────────────────────────────────────────┐
-│ [('a',42),('b','Hello'),('c',[1,2,3])] │ [('a','UInt32'),('b','String'),('c','Array(UInt32)')] │
-└────────────────────────────────────────┴───────────────────────────────────────────────────────┘
+┌─dynamics───────────────────────────────┬─dynamic_types─────────────────────────────────────────────────┐
+│ [('a',42),('b','Hello'),('c',[1,2,3])] │ [('a','Int64'),('b','String'),('c','Array(Nullable(Int64))')] │
+└────────────────────────────────────────┴───────────────────────────────────────────────────────────────┘
 ```
 
 ### Binary output format

docs/en/sql-reference/functions/other-functions.md

@@ -4189,3 +4189,94 @@ Result:
 │                          32 │
 └─────────────────────────────┘
 ```
+
+## getSubcolumn
+
+Takes a table expression or identifier and constant string with the name of the sub-column, and returns the requested sub-column extracted from the expression.
+
+**Syntax**
+
+```sql
+getSubcolumn(col_name, subcol_name)
+```
+
+**Arguments**
+
+- `col_name` — Table expression or identifier. [Expression](../syntax.md/#expressions), [Identifier](../syntax.md/#identifiers).
+- `subcol_name` — The name of the sub-column. [String](../data-types/string.md).
+
+**Returned value**
+
+- Returns the extracted sub-column.
+
+**Example**
+
+Query:
+
+```sql
+CREATE TABLE t_arr (arr Array(Tuple(subcolumn1 UInt32, subcolumn2 String))) ENGINE = MergeTree ORDER BY tuple();
+INSERT INTO t_arr VALUES ([(1, 'Hello'), (2, 'World')]), ([(3, 'This'), (4, 'is'), (5, 'subcolumn')]);
+SELECT getSubcolumn(arr, 'subcolumn1'), getSubcolumn(arr, 'subcolumn2') FROM t_arr;
+```
+
+Result:
+
+```response
+   ┌─getSubcolumn(arr, 'subcolumn1')─┬─getSubcolumn(arr, 'subcolumn2')─┐
+1. │ [1,2]                           │ ['Hello','World']               │
+2. │ [3,4,5]                         │ ['This','is','subcolumn']       │
+   └─────────────────────────────────┴─────────────────────────────────┘
+```
+
+## getTypeSerializationStreams
+
+Enumerates stream paths of a data type.
+
+:::note
+This function is intended for use by developers.
+:::
+
+**Syntax**
+
+```sql
+getTypeSerializationStreams(col)
+```
+
+**Arguments**
+
+- `col` — Column or string representation of a data-type from which the data type will be detected.
+
+**Returned value**
+
+- Returns an array with all the serialization sub-stream paths.[Array](../data-types/array.md)([String](../data-types/string.md)).
+
+**Examples**
+
+Query:
+
+```sql
+SELECT getTypeSerializationStreams(tuple('a', 1, 'b', 2));
+```
+
+Result:
+
+```response
+   ┌─getTypeSerializationStreams(('a', 1, 'b', 2))─────────────────────────────────────────────────────────────────────────┐
+1. │ ['{TupleElement(1), Regular}','{TupleElement(2), Regular}','{TupleElement(3), Regular}','{TupleElement(4), Regular}'] │
+   └───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
+```
+
+Query:
+
+```sql
+SELECT getTypeSerializationStreams('Map(String, Int64)');
+```
+
+Result:
+
+```response
+   ┌─getTypeSerializationStreams('Map(String, Int64)')────────────────────────────────────────────────────────────────┐
+1. │ ['{ArraySizes}','{ArrayElements, TupleElement(keys), Regular}','{ArrayElements, TupleElement(values), Regular}'] │
+   └──────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
+```
+

docs/en/sql-reference/table-functions/azureBlobStorage.md

@@ -77,3 +77,16 @@ SELECT count(*) FROM azureBlobStorage('DefaultEndpointsProtocol=https;AccountNam
 **See Also**
 
 - [AzureBlobStorage Table Engine](/docs/en/engines/table-engines/integrations/azureBlobStorage.md)
+
+## Hive-style partitioning {#hive-style-partitioning}
+
+When setting `use_hive_partitioning` is set to 1, ClickHouse will detect Hive-style partitioning in the path (`/name=value/`) and will allow to use partition columns as virtual columns in the query. These virtual columns will have the same names as in the partitioned path, but starting with `_`.
+
+**Example**
+
+Use virtual column, created with Hive-style partitioning
+
+``` sql
+SET use_hive_partitioning = 1;
+SELECT * from azureBlobStorage(config, storage_account_url='...', container='...', blob_path='http://data/path/date=*/country=*/code=*/*.parquet') where _date > '2020-01-01' and _country = 'Netherlands' and _code = 42;
+```

docs/en/sql-reference/table-functions/file.md

@@ -103,7 +103,7 @@ LIMIT 2;
 └─────────┴─────────┴─────────┘
 ```
 
-### Inserting data from a file into a table:
+### Inserting data from a file into a table
 
 ``` sql
 INSERT INTO FUNCTION
@@ -206,6 +206,19 @@ SELECT count(*) FROM file('big_dir/**/file002', 'CSV', 'name String, value UInt3
 - `_size` — Size of the file in bytes. Type: `Nullable(UInt64)`. If the file size is unknown, the value is `NULL`.
 - `_time` — Last modified time of the file. Type: `Nullable(DateTime)`. If the time is unknown, the value is `NULL`.
 
+## Hive-style partitioning {#hive-style-partitioning}
+
+When setting `use_hive_partitioning` is set to 1, ClickHouse will detect Hive-style partitioning in the path (`/name=value/`) and will allow to use partition columns as virtual columns in the query. These virtual columns will have the same names as in the partitioned path, but starting with `_`.
+
+**Example**
+
+Use virtual column, created with Hive-style partitioning
+
+``` sql
+SET use_hive_partitioning = 1;
+SELECT * from file('data/path/date=*/country=*/code=*/*.parquet') where _date > '2020-01-01' and _country = 'Netherlands' and _code = 42;
+```
+
 ## Settings {#settings}
 
 - [engine_file_empty_if_not_exists](/docs/en/operations/settings/settings.md#engine-file-empty_if-not-exists) - allows to select empty data from a file that doesn't exist. Disabled by default.

docs/en/sql-reference/table-functions/hdfs.md

@@ -100,6 +100,19 @@ FROM hdfs('hdfs://hdfs1:9000/big_dir/file{0..9}{0..9}{0..9}', 'CSV', 'name Strin
 - `_size` — Size of the file in bytes. Type: `Nullable(UInt64)`. If the size is unknown, the value is `NULL`.
 - `_time` — Last modified time of the file. Type: `Nullable(DateTime)`. If the time is unknown, the value is `NULL`.
 
+## Hive-style partitioning {#hive-style-partitioning}
+
+When setting `use_hive_partitioning` is set to 1, ClickHouse will detect Hive-style partitioning in the path (`/name=value/`) and will allow to use partition columns as virtual columns in the query. These virtual columns will have the same names as in the partitioned path, but starting with `_`.
+
+**Example**
+
+Use virtual column, created with Hive-style partitioning
+
+``` sql
+SET use_hive_partitioning = 1;
+SELECT * from HDFS('hdfs://hdfs1:9000/data/path/date=*/country=*/code=*/*.parquet') where _date > '2020-01-01' and _country = 'Netherlands' and _code = 42;
+```
+
 ## Storage Settings {#storage-settings}
 
 - [hdfs_truncate_on_insert](/docs/en/operations/settings/settings.md#hdfs_truncate_on_insert) - allows to truncate file before insert into it. Disabled by default.

docs/en/sql-reference/table-functions/s3.md

@@ -274,6 +274,19 @@ FROM s3(
 - `_size` — Size of the file in bytes. Type: `Nullable(UInt64)`. If the file size is unknown, the value is `NULL`. In case of archive shows uncompressed file size of the file inside the archive. 
 - `_time` — Last modified time of the file. Type: `Nullable(DateTime)`. If the time is unknown, the value is `NULL`.
 
+## Hive-style partitioning {#hive-style-partitioning}
+
+When setting `use_hive_partitioning` is set to 1, ClickHouse will detect Hive-style partitioning in the path (`/name=value/`) and will allow to use partition columns as virtual columns in the query. These virtual columns will have the same names as in the partitioned path, but starting with `_`.
+
+**Example**
+
+Use virtual column, created with Hive-style partitioning
+
+``` sql
+SET use_hive_partitioning = 1;
+SELECT * from s3('s3://data/path/date=*/country=*/code=*/*.parquet') where _date > '2020-01-01' and _country = 'Netherlands' and _code = 42;
+```
+
 ## Storage Settings {#storage-settings}
 
 - [s3_truncate_on_insert](/docs/en/operations/settings/settings.md#s3_truncate_on_insert) - allows to truncate file before insert into it. Disabled by default.

docs/en/sql-reference/table-functions/url.md

@@ -55,6 +55,19 @@ Character `|` inside patterns is used to specify failover addresses. They are it
 - `_size` — Size of the resource in bytes. Type: `Nullable(UInt64)`. If the size is unknown, the value is `NULL`.
 - `_time` — Last modified time of the file. Type: `Nullable(DateTime)`. If the time is unknown, the value is `NULL`.
 
+## Hive-style partitioning {#hive-style-partitioning}
+
+When setting `use_hive_partitioning` is set to 1, ClickHouse will detect Hive-style partitioning in the path (`/name=value/`) and will allow to use partition columns as virtual columns in the query. These virtual columns will have the same names as in the partitioned path, but starting with `_`.
+
+**Example**
+
+Use virtual column, created with Hive-style partitioning
+
+``` sql
+SET use_hive_partitioning = 1;
+SELECT * from url('http://data/path/date=*/country=*/code=*/*.parquet') where _date > '2020-01-01' and _country = 'Netherlands' and _code = 42;
+```
+
 ## Storage Settings {#storage-settings}
 
 - [engine_url_skip_empty_files](/docs/en/operations/settings/settings.md#engine_url_skip_empty_files) - allows to skip empty files while reading. Disabled by default.

programs/CMakeLists.txt

@@ -1,4 +1,4 @@
-add_compile_options($<$<OR:$<COMPILE_LANGUAGE:C>,$<COMPILE_LANGUAGE:CXX>>:${COVERAGE_FLAGS}>)
+add_compile_options("$<$<OR:$<COMPILE_LANGUAGE:C>,$<COMPILE_LANGUAGE:CXX>>:${COVERAGE_FLAGS}>")
 
 if (USE_CLANG_TIDY)
     set (CMAKE_CXX_CLANG_TIDY "${CLANG_TIDY_PATH}")

programs/benchmark/Benchmark.cpp

@@ -75,6 +75,8 @@ public:
             const String & default_database_,
             const String & user_,
             const String & password_,
+            const String & proto_send_chunked_,
+            const String & proto_recv_chunked_,
             const String & quota_key_,
             const String & stage,
             bool randomize_,
@@ -128,7 +130,9 @@ public:
             connections.emplace_back(std::make_unique<ConnectionPool>(
                 concurrency,
                 cur_host, cur_port,
-                default_database_, user_, password_, quota_key_,
+                default_database_, user_, password_,
+                proto_send_chunked_, proto_recv_chunked_,
+                quota_key_,
                 /* cluster_= */ "",
                 /* cluster_secret_= */ "",
                 /* client_name_= */ std::string(DEFAULT_CLIENT_NAME),
@@ -662,6 +666,50 @@ int mainEntryClickHouseBenchmark(int argc, char ** argv)
 
         Strings hosts = options.count("host") ? options["host"].as<Strings>() : Strings({"localhost"});
 
+        String proto_send_chunked {"notchunked"};
+        String proto_recv_chunked {"notchunked"};
+
+        if (options.count("proto_caps"))
+        {
+            std::string proto_caps_str = options["proto_caps"].as<std::string>();
+
+            std::vector<std::string_view> proto_caps;
+            splitInto<','>(proto_caps, proto_caps_str);
+
+            for (auto cap_str : proto_caps)
+            {
+                std::string direction;
+
+                if (cap_str.starts_with("send_"))
+                {
+                    direction = "send";
+                    cap_str = cap_str.substr(std::string_view("send_").size());
+                }
+                else if (cap_str.starts_with("recv_"))
+                {
+                    direction = "recv";
+                    cap_str = cap_str.substr(std::string_view("recv_").size());
+                }
+
+                if (cap_str != "chunked" && cap_str != "notchunked" && cap_str != "chunked_optional" && cap_str != "notchunked_optional")
+                    throw Exception(ErrorCodes::BAD_ARGUMENTS, "proto_caps option is incorrect ({})", proto_caps_str);
+
+                if (direction.empty())
+                {
+                    proto_send_chunked = cap_str;
+                    proto_recv_chunked = cap_str;
+                }
+                else
+                {
+                    if (direction == "send")
+                        proto_send_chunked = cap_str;
+                    else
+                        proto_recv_chunked = cap_str;
+                }
+            }
+        }
+
+
         Benchmark benchmark(
             options["concurrency"].as<unsigned>(),
             options["delay"].as<double>(),
@@ -673,6 +721,8 @@ int mainEntryClickHouseBenchmark(int argc, char ** argv)
             options["database"].as<std::string>(),
             options["user"].as<std::string>(),
             options["password"].as<std::string>(),
+            proto_send_chunked,
+            proto_recv_chunked,
             options["quota_key"].as<std::string>(),
             options["stage"].as<std::string>(),
             options.count("randomize"),

programs/client/clickhouse-client.xml

@@ -38,6 +38,21 @@
         <production>{display_name} \e[1;31m:)\e[0m </production> <!-- if it matched to the substring "production" in the server display name -->
     </prompt_by_server_display_name>
 
+    <!-- Chunked capabilities for native protocol by client.
+         Can be enabled separately for send and receive channels.
+         Supported modes:
+         - chunked - client will only work with server supporting chunked protocol;
+         - chunked_optional - client prefer server to enable chunked protocol, but can switch to notchunked if server does not support this;
+         - notchunked - client will only work with server supporting notchunked protocol (current default);
+         - notchunked_optional - client prefer server notchunked protocol, but can switch to chunked if server does not support this.
+     -->
+    <!--
+    <proto_caps>
+        <send>chunked_optional</send>
+        <recv>chunked_optional</recv>
+    </proto_caps>
+    -->
+
     <!--
         Settings adjustable via command-line parameters
         can take their defaults from that config file, see examples:

programs/library-bridge/LibraryBridge.cpp

@@ -27,7 +27,7 @@ std::string LibraryBridge::bridgeName() const
 
 LibraryBridge::HandlerFactoryPtr LibraryBridge::getHandlerFactoryPtr(ContextPtr context) const
 {
-    return std::make_shared<LibraryBridgeHandlerFactory>("LibraryRequestHandlerFactory", keep_alive_timeout, context);
+    return std::make_shared<LibraryBridgeHandlerFactory>("LibraryRequestHandlerFactory", context);
 }
 
 }

programs/library-bridge/LibraryBridgeHandlerFactory.cpp

@@ -9,12 +9,10 @@ namespace DB
 {
 LibraryBridgeHandlerFactory::LibraryBridgeHandlerFactory(
     const std::string & name_,
-    size_t keep_alive_timeout_,
     ContextPtr context_)
     : WithContext(context_)
     , log(getLogger(name_))
     , name(name_)
-    , keep_alive_timeout(keep_alive_timeout_)
 {
 }
 
@@ -26,17 +24,17 @@ std::unique_ptr<HTTPRequestHandler> LibraryBridgeHandlerFactory::createRequestHa
     if (request.getMethod() == Poco::Net::HTTPRequest::HTTP_GET)
     {
         if (uri.getPath() == "/extdict_ping")
-            return std::make_unique<ExternalDictionaryLibraryBridgeExistsHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<ExternalDictionaryLibraryBridgeExistsHandler>(getContext());
         else if (uri.getPath() == "/catboost_ping")
-            return std::make_unique<CatBoostLibraryBridgeExistsHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<CatBoostLibraryBridgeExistsHandler>(getContext());
     }
 
     if (request.getMethod() == Poco::Net::HTTPRequest::HTTP_POST)
     {
         if (uri.getPath() == "/extdict_request")
-            return std::make_unique<ExternalDictionaryLibraryBridgeRequestHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<ExternalDictionaryLibraryBridgeRequestHandler>(getContext());
         else if (uri.getPath() == "/catboost_request")
-            return std::make_unique<CatBoostLibraryBridgeRequestHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<CatBoostLibraryBridgeRequestHandler>(getContext());
     }
 
     return nullptr;

programs/library-bridge/LibraryBridgeHandlerFactory.h

@@ -13,7 +13,6 @@ class LibraryBridgeHandlerFactory : public HTTPRequestHandlerFactory, WithContex
 public:
     LibraryBridgeHandlerFactory(
         const std::string & name_,
-        size_t keep_alive_timeout_,
         ContextPtr context_);
 
     std::unique_ptr<HTTPRequestHandler> createRequestHandler(const HTTPServerRequest & request) override;
@@ -21,7 +20,6 @@ public:
 private:
     LoggerPtr log;
     const std::string name;
-    const size_t keep_alive_timeout;
 };
 
 }

programs/library-bridge/LibraryBridgeHandlers.cpp

@@ -87,10 +87,8 @@ static void writeData(Block data, OutputFormatPtr format)
 }
 
 
-ExternalDictionaryLibraryBridgeRequestHandler::ExternalDictionaryLibraryBridgeRequestHandler(size_t keep_alive_timeout_, ContextPtr context_)
-    : WithContext(context_)
-    , keep_alive_timeout(keep_alive_timeout_)
-    , log(getLogger("ExternalDictionaryLibraryBridgeRequestHandler"))
+ExternalDictionaryLibraryBridgeRequestHandler::ExternalDictionaryLibraryBridgeRequestHandler(ContextPtr context_)
+    : WithContext(context_), log(getLogger("ExternalDictionaryLibraryBridgeRequestHandler"))
 {
 }
 
@@ -137,7 +135,7 @@ void ExternalDictionaryLibraryBridgeRequestHandler::handleRequest(HTTPServerRequ
     const String & dictionary_id = params.get("dictionary_id");
 
     LOG_TRACE(log, "Library method: '{}', dictionary id: {}", method, dictionary_id);
-    WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD, keep_alive_timeout);
+    WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD);
 
     try
     {
@@ -374,10 +372,8 @@ void ExternalDictionaryLibraryBridgeRequestHandler::handleRequest(HTTPServerRequ
 }
 
 
-ExternalDictionaryLibraryBridgeExistsHandler::ExternalDictionaryLibraryBridgeExistsHandler(size_t keep_alive_timeout_, ContextPtr context_)
-    : WithContext(context_)
-    , keep_alive_timeout(keep_alive_timeout_)
-    , log(getLogger("ExternalDictionaryLibraryBridgeExistsHandler"))
+ExternalDictionaryLibraryBridgeExistsHandler::ExternalDictionaryLibraryBridgeExistsHandler(ContextPtr context_)
+    : WithContext(context_), log(getLogger("ExternalDictionaryLibraryBridgeExistsHandler"))
 {
 }
 
@@ -401,7 +397,7 @@ void ExternalDictionaryLibraryBridgeExistsHandler::handleRequest(HTTPServerReque
 
         String res = library_handler ? "1" : "0";
 
-        setResponseDefaultHeaders(response, keep_alive_timeout);
+        setResponseDefaultHeaders(response);
         LOG_TRACE(log, "Sending ping response: {} (dictionary id: {})", res, dictionary_id);
         response.sendBuffer(res.data(), res.size());
     }
@@ -412,11 +408,8 @@ void ExternalDictionaryLibraryBridgeExistsHandler::handleRequest(HTTPServerReque
 }
 
 
-CatBoostLibraryBridgeRequestHandler::CatBoostLibraryBridgeRequestHandler(
-    size_t keep_alive_timeout_, ContextPtr context_)
-    : WithContext(context_)
-    , keep_alive_timeout(keep_alive_timeout_)
-    , log(getLogger("CatBoostLibraryBridgeRequestHandler"))
+CatBoostLibraryBridgeRequestHandler::CatBoostLibraryBridgeRequestHandler(ContextPtr context_)
+    : WithContext(context_), log(getLogger("CatBoostLibraryBridgeRequestHandler"))
 {
 }
 
@@ -455,7 +448,7 @@ void CatBoostLibraryBridgeRequestHandler::handleRequest(HTTPServerRequest & requ
     const String & method = params.get("method");
 
     LOG_TRACE(log, "Library method: '{}'", method);
-    WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD, keep_alive_timeout);
+    WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD);
 
     try
     {
@@ -617,10 +610,8 @@ void CatBoostLibraryBridgeRequestHandler::handleRequest(HTTPServerRequest & requ
 }
 
 
-CatBoostLibraryBridgeExistsHandler::CatBoostLibraryBridgeExistsHandler(size_t keep_alive_timeout_, ContextPtr context_)
-    : WithContext(context_)
-    , keep_alive_timeout(keep_alive_timeout_)
-    , log(getLogger("CatBoostLibraryBridgeExistsHandler"))
+CatBoostLibraryBridgeExistsHandler::CatBoostLibraryBridgeExistsHandler(ContextPtr context_)
+    : WithContext(context_), log(getLogger("CatBoostLibraryBridgeExistsHandler"))
 {
 }
 
@@ -634,7 +625,7 @@ void CatBoostLibraryBridgeExistsHandler::handleRequest(HTTPServerRequest & reque
 
         String res = "1";
 
-        setResponseDefaultHeaders(response, keep_alive_timeout);
+        setResponseDefaultHeaders(response);
         LOG_TRACE(log, "Sending ping response: {}", res);
         response.sendBuffer(res.data(), res.size());
     }

programs/library-bridge/LibraryBridgeHandlers.h

@@ -18,14 +18,13 @@ namespace DB
 class ExternalDictionaryLibraryBridgeRequestHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    ExternalDictionaryLibraryBridgeRequestHandler(size_t keep_alive_timeout_, ContextPtr context_);
+    explicit ExternalDictionaryLibraryBridgeRequestHandler(ContextPtr context_);
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
     static constexpr auto FORMAT = "RowBinary";
 
-    const size_t keep_alive_timeout;
     LoggerPtr log;
 };
 
@@ -34,12 +33,11 @@ private:
 class ExternalDictionaryLibraryBridgeExistsHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    ExternalDictionaryLibraryBridgeExistsHandler(size_t keep_alive_timeout_, ContextPtr context_);
+    explicit ExternalDictionaryLibraryBridgeExistsHandler(ContextPtr context_);
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
-    const size_t keep_alive_timeout;
     LoggerPtr log;
 };
 
@@ -63,12 +61,11 @@ private:
 class CatBoostLibraryBridgeRequestHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    CatBoostLibraryBridgeRequestHandler(size_t keep_alive_timeout_, ContextPtr context_);
+    explicit CatBoostLibraryBridgeRequestHandler(ContextPtr context_);
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
-    const size_t keep_alive_timeout;
     LoggerPtr log;
 };
 
@@ -77,12 +74,11 @@ private:
 class CatBoostLibraryBridgeExistsHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    CatBoostLibraryBridgeExistsHandler(size_t keep_alive_timeout_, ContextPtr context_);
+    explicit CatBoostLibraryBridgeExistsHandler(ContextPtr context_);
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
-    const size_t keep_alive_timeout;
     LoggerPtr log;
 };
 

programs/obfuscator/Obfuscator.cpp

@@ -1307,6 +1307,7 @@ try
             throw ErrnoException(ErrorCodes::CANNOT_SEEK_THROUGH_FILE, "Input must be seekable file (it will be read twice)");
 
         SingleReadBufferIterator read_buffer_iterator(std::move(file));
+
         schema_columns = readSchemaFromFormat(input_format, {}, read_buffer_iterator, context_const);
     }
     else

programs/odbc-bridge/ColumnInfoHandler.cpp

@@ -202,10 +202,7 @@ void ODBCColumnsInfoHandler::handleRequest(HTTPServerRequest & request, HTTPServ
         if (columns.empty())
             throw Exception(ErrorCodes::UNKNOWN_TABLE, "Columns definition was not returned");
 
-        WriteBufferFromHTTPServerResponse out(
-            response,
-            request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD,
-            keep_alive_timeout);
+        WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD);
         try
         {
             writeStringBinary(columns.toString(), out);

programs/odbc-bridge/ColumnInfoHandler.h

@@ -15,18 +15,12 @@ namespace DB
 class ODBCColumnsInfoHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    ODBCColumnsInfoHandler(size_t keep_alive_timeout_, ContextPtr context_)
-        : WithContext(context_)
-        , log(getLogger("ODBCColumnsInfoHandler"))
-        , keep_alive_timeout(keep_alive_timeout_)
-    {
-    }
+    explicit ODBCColumnsInfoHandler(ContextPtr context_) : WithContext(context_), log(getLogger("ODBCColumnsInfoHandler")) { }
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
     LoggerPtr log;
-    size_t keep_alive_timeout;
 };
 
 }

programs/odbc-bridge/IdentifierQuoteHandler.cpp

@@ -74,7 +74,7 @@ void IdentifierQuoteHandler::handleRequest(HTTPServerRequest & request, HTTPServ
 
         auto identifier = getIdentifierQuote(std::move(connection));
 
-        WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD, keep_alive_timeout);
+        WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD);
         try
         {
             writeStringBinary(identifier, out);

programs/odbc-bridge/IdentifierQuoteHandler.h

@@ -14,18 +14,12 @@ namespace DB
 class IdentifierQuoteHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    IdentifierQuoteHandler(size_t keep_alive_timeout_, ContextPtr context_)
-        : WithContext(context_)
-        , log(getLogger("IdentifierQuoteHandler"))
-        , keep_alive_timeout(keep_alive_timeout_)
-    {
-    }
+    explicit IdentifierQuoteHandler(ContextPtr context_) : WithContext(context_), log(getLogger("IdentifierQuoteHandler")) { }
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
     LoggerPtr log;
-    size_t keep_alive_timeout;
 };
 
 }

programs/odbc-bridge/MainHandler.cpp

@@ -132,7 +132,7 @@ void ODBCHandler::handleRequest(HTTPServerRequest & request, HTTPServerResponse
         return;
     }
 
-    WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD, keep_alive_timeout);
+    WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD);
 
     try
     {

programs/odbc-bridge/MainHandler.h

@@ -20,12 +20,10 @@ class ODBCHandler : public HTTPRequestHandler, WithContext
 {
 public:
     ODBCHandler(
-        size_t keep_alive_timeout_,
         ContextPtr context_,
         const String & mode_)
         : WithContext(context_)
         , log(getLogger("ODBCHandler"))
-        , keep_alive_timeout(keep_alive_timeout_)
         , mode(mode_)
     {
     }
@@ -35,7 +33,6 @@ public:
 private:
     LoggerPtr log;
 
-    size_t keep_alive_timeout;
     String mode;
 
     static inline std::mutex mutex;

programs/odbc-bridge/ODBCBridge.cpp

@@ -27,7 +27,7 @@ std::string ODBCBridge::bridgeName() const
 
 ODBCBridge::HandlerFactoryPtr ODBCBridge::getHandlerFactoryPtr(ContextPtr context) const
 {
-    return std::make_shared<ODBCBridgeHandlerFactory>("ODBCRequestHandlerFactory-factory", keep_alive_timeout, context);
+    return std::make_shared<ODBCBridgeHandlerFactory>("ODBCRequestHandlerFactory-factory", context);
 }
 
 }

programs/odbc-bridge/ODBCHandlerFactory.cpp

@@ -9,11 +9,8 @@
 namespace DB
 {
 
-ODBCBridgeHandlerFactory::ODBCBridgeHandlerFactory(const std::string & name_, size_t keep_alive_timeout_, ContextPtr context_)
-    : WithContext(context_)
-    , log(getLogger(name_))
-    , name(name_)
-    , keep_alive_timeout(keep_alive_timeout_)
+ODBCBridgeHandlerFactory::ODBCBridgeHandlerFactory(const std::string & name_, ContextPtr context_)
+    : WithContext(context_), log(getLogger(name_)), name(name_)
 {
 }
 
@@ -23,33 +20,33 @@ std::unique_ptr<HTTPRequestHandler> ODBCBridgeHandlerFactory::createRequestHandl
     LOG_TRACE(log, "Request URI: {}", uri.toString());
 
     if (uri.getPath() == "/ping" && request.getMethod() == Poco::Net::HTTPRequest::HTTP_GET)
-        return std::make_unique<PingHandler>(keep_alive_timeout);
+        return std::make_unique<PingHandler>();
 
     if (request.getMethod() == Poco::Net::HTTPRequest::HTTP_POST)
     {
 
         if (uri.getPath() == "/columns_info")
 #if USE_ODBC
-            return std::make_unique<ODBCColumnsInfoHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<ODBCColumnsInfoHandler>(getContext());
 #else
             return nullptr;
 #endif
         else if (uri.getPath() == "/identifier_quote")
 #if USE_ODBC
-            return std::make_unique<IdentifierQuoteHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<IdentifierQuoteHandler>(getContext());
 #else
             return nullptr;
 #endif
         else if (uri.getPath() == "/schema_allowed")
 #if USE_ODBC
-            return std::make_unique<SchemaAllowedHandler>(keep_alive_timeout, getContext());
+            return std::make_unique<SchemaAllowedHandler>(getContext());
 #else
             return nullptr;
 #endif
         else if (uri.getPath() == "/write")
-            return std::make_unique<ODBCHandler>(keep_alive_timeout, getContext(), "write");
+            return std::make_unique<ODBCHandler>(getContext(), "write");
         else
-            return std::make_unique<ODBCHandler>(keep_alive_timeout, getContext(), "read");
+            return std::make_unique<ODBCHandler>(getContext(), "read");
     }
     return nullptr;
 }

programs/odbc-bridge/ODBCHandlerFactory.h

@@ -17,14 +17,13 @@ namespace DB
 class ODBCBridgeHandlerFactory : public HTTPRequestHandlerFactory, WithContext
 {
 public:
-    ODBCBridgeHandlerFactory(const std::string & name_, size_t keep_alive_timeout_, ContextPtr context_);
+    ODBCBridgeHandlerFactory(const std::string & name_, ContextPtr context_);
 
     std::unique_ptr<HTTPRequestHandler> createRequestHandler(const HTTPServerRequest & request) override;
 
 private:
     LoggerPtr log;
     std::string name;
-    size_t keep_alive_timeout;
 };
 
 }

programs/odbc-bridge/PingHandler.cpp

@@ -10,7 +10,7 @@ void PingHandler::handleRequest(HTTPServerRequest & /* request */, HTTPServerRes
 {
     try
     {
-        setResponseDefaultHeaders(response, keep_alive_timeout);
+        setResponseDefaultHeaders(response);
         const char * data = "Ok.\n";
         response.sendBuffer(data, strlen(data));
     }

programs/odbc-bridge/PingHandler.h

@@ -9,11 +9,7 @@ namespace DB
 class PingHandler : public HTTPRequestHandler
 {
 public:
-    explicit PingHandler(size_t keep_alive_timeout_) : keep_alive_timeout(keep_alive_timeout_) {}
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
-
-private:
-    size_t keep_alive_timeout;
 };
 
 }

programs/odbc-bridge/SchemaAllowedHandler.cpp

@@ -88,7 +88,7 @@ void SchemaAllowedHandler::handleRequest(HTTPServerRequest & request, HTTPServer
 
         bool result = isSchemaAllowed(std::move(connection));
 
-        WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD, keep_alive_timeout);
+        WriteBufferFromHTTPServerResponse out(response, request.getMethod() == Poco::Net::HTTPRequest::HTTP_HEAD);
         try
         {
             writeBoolText(result, out);

programs/odbc-bridge/SchemaAllowedHandler.h

@@ -17,18 +17,12 @@ class Context;
 class SchemaAllowedHandler : public HTTPRequestHandler, WithContext
 {
 public:
-    SchemaAllowedHandler(size_t keep_alive_timeout_, ContextPtr context_)
-        : WithContext(context_)
-        , log(getLogger("SchemaAllowedHandler"))
-        , keep_alive_timeout(keep_alive_timeout_)
-    {
-    }
+    explicit SchemaAllowedHandler(ContextPtr context_) : WithContext(context_), log(getLogger("SchemaAllowedHandler")) { }
 
     void handleRequest(HTTPServerRequest & request, HTTPServerResponse & response, const ProfileEvents::Event & write_event) override;
 
 private:
     LoggerPtr log;
-    size_t keep_alive_timeout;
 };
 
 }

programs/server/Server.cpp

@@ -2428,6 +2428,7 @@ void Server::createServers(
     Poco::Net::HTTPServerParams::Ptr http_params = new Poco::Net::HTTPServerParams;
     http_params->setTimeout(settings.http_receive_timeout);
     http_params->setKeepAliveTimeout(global_context->getServerSettings().keep_alive_timeout);
+    http_params->setMaxKeepAliveRequests(static_cast<int>(global_context->getServerSettings().max_keep_alive_requests));
 
     Poco::Util::AbstractConfiguration::Keys protocols;
     config.keys("protocols", protocols);

programs/server/config.xml

@@ -150,6 +150,21 @@
     -->
     <tcp_port>9000</tcp_port>
 
+    <!-- Chunked capabilities for native protocol by server.
+         Can be enabled separately for send and receive channels.
+         Supported modes:
+         - chunked - server requires from client to have chunked enabled;
+         - chunked_optional - server supports both chunked and notchunked protocol;
+         - notchunked - server requires from client notchunked protocol (current default);
+         - notchunked_optional - server supports both chunked and notchunked protocol.
+     -->
+    <!--
+    <proto_caps>
+        <send>notchunked_optional</send>
+        <recv>notchunked_optional</recv>
+    </proto_caps>
+    -->
+
     <!-- Compatibility with MySQL protocol.
          ClickHouse will pretend to be MySQL for applications connecting to this port.
     -->

programs/server/fuzzers/tcp_protocol_fuzzer.cpp

@@ -10,6 +10,7 @@
 #include <Poco/Net/SocketAddress.h>
 #include <Poco/Net/StreamSocket.h>
 
+#include <Daemon/BaseDaemon.h>
 #include <Interpreters/Context.h>
 
 
@@ -25,6 +26,12 @@ static int64_t port = 9000;
 
 using namespace std::chrono_literals;
 
+void on_exit()
+{
+    BaseDaemon::terminate();
+    main_app.wait();
+}
+
 extern "C"
 int LLVMFuzzerInitialize(int * argc, char ***argv)
 {
@@ -60,6 +67,8 @@ int LLVMFuzzerInitialize(int * argc, char ***argv)
             exit(-1);
     }
 
+    atexit(on_exit);
+
     return 0;
 }
 

src/AggregateFunctions/fuzzers/CMakeLists.txt

@@ -1,2 +1,2 @@
 clickhouse_add_executable(aggregate_function_state_deserialization_fuzzer aggregate_function_state_deserialization_fuzzer.cpp ${SRCS})
-target_link_libraries(aggregate_function_state_deserialization_fuzzer PRIVATE dbms clickhouse_aggregate_functions clickhouse_functions)
+target_link_libraries(aggregate_function_state_deserialization_fuzzer PRIVATE clickhouse_functions clickhouse_aggregate_functions)

src/CMakeLists.txt

@@ -1,4 +1,4 @@
-add_compile_options($<$<OR:$<COMPILE_LANGUAGE:C>,$<COMPILE_LANGUAGE:CXX>>:${COVERAGE_FLAGS}>)
+add_compile_options("$<$<OR:$<COMPILE_LANGUAGE:C>,$<COMPILE_LANGUAGE:CXX>>:${COVERAGE_FLAGS}>")
 
 if (USE_INCLUDE_WHAT_YOU_USE)
     set (CMAKE_CXX_INCLUDE_WHAT_YOU_USE ${IWYU_PATH})
@@ -601,10 +601,6 @@ endif()
 
 dbms_target_link_libraries(PUBLIC ch_contrib::consistent_hashing)
 
-if (TARGET ch_contrib::annoy)
-    dbms_target_link_libraries(PUBLIC ch_contrib::annoy)
-endif()
-
 if (TARGET ch_contrib::usearch)
     dbms_target_link_libraries(PUBLIC ch_contrib::usearch)
 endif()

src/Client/ClientApplicationBase.cpp

@@ -158,6 +158,8 @@ void ClientApplicationBase::init(int argc, char ** argv)
 
         ("config-file,C", po::value<std::string>(), "config-file path")
 
+        ("proto_caps", po::value<std::string>(), "enable/disable chunked protocol: chunked_optional, notchunked, notchunked_optional, send_chunked, send_chunked_optional, send_notchunked, send_notchunked_optional, recv_chunked, recv_chunked_optional, recv_notchunked, recv_notchunked_optional")
+
         ("query,q", po::value<std::vector<std::string>>()->multitoken(), R"(Query. Can be specified multiple times (--query "SELECT 1" --query "SELECT 2") or once with multiple comma-separated queries (--query "SELECT 1; SELECT 2;"). In the latter case, INSERT queries with non-VALUE format must be separated by empty lines.)")
         ("queries-file", po::value<std::vector<std::string>>()->multitoken(), "file path with queries to execute; multiple files can be specified (--queries-file file1 file2...)")
         ("multiquery,n", "Obsolete, does nothing")
@@ -337,6 +339,41 @@ void ClientApplicationBase::init(int argc, char ** argv)
     if (options.count("server_logs_file"))
         server_logs_file = options["server_logs_file"].as<std::string>();
 
+    if (options.count("proto_caps"))
+    {
+        std::string proto_caps_str = options["proto_caps"].as<std::string>();
+
+        std::vector<std::string_view> proto_caps;
+        splitInto<','>(proto_caps, proto_caps_str);
+
+        for (auto cap_str : proto_caps)
+        {
+            std::string direction;
+
+            if (cap_str.starts_with("send_"))
+            {
+                direction = "send";
+                cap_str = cap_str.substr(std::string_view("send_").size());
+            }
+            else if (cap_str.starts_with("recv_"))
+            {
+                direction = "recv";
+                cap_str = cap_str.substr(std::string_view("recv_").size());
+            }
+
+            if (cap_str != "chunked" && cap_str != "notchunked" && cap_str != "chunked_optional" && cap_str != "notchunked_optional")
+                throw Exception(ErrorCodes::BAD_ARGUMENTS, "proto_caps option is incorrect ({})", proto_caps_str);
+
+            if (direction.empty())
+            {
+                config().setString("proto_caps.send", std::string(cap_str));
+                config().setString("proto_caps.recv", std::string(cap_str));
+            }
+            else
+                config().setString("proto_caps." + direction, std::string(cap_str));
+        }
+    }
+
     query_processing_stage = QueryProcessingStage::fromString(options["stage"].as<std::string>());
     query_kind = parseQueryKind(options["query_kind"].as<std::string>());
     profile_events.print = options.count("print-profile-events");

src/Client/ClientBase.cpp

@@ -73,9 +73,11 @@
 #include <limits>
 #include <map>
 #include <memory>
+#include <string_view>
 #include <unordered_map>
 
 #include <Common/config_version.h>
+#include <base/find_symbols.h>
 #include "config.h"
 #include <IO/ReadHelpers.h>
 #include <Processors/Formats/Impl/ValuesBlockInputFormat.h>
@@ -914,6 +916,8 @@ void ClientBase::processTextAsSingleQuery(const String & full_query)
     }
     catch (Exception & e)
     {
+        if (server_exception)
+            server_exception->rethrow();
         if (!is_interactive)
             e.addMessage("(in query: {})", full_query);
         throw;
@@ -1032,19 +1036,28 @@ void ClientBase::processOrdinaryQuery(const String & query_to_execute, ASTPtr pa
             query_interrupt_handler.start(signals_before_stop);
             SCOPE_EXIT({ query_interrupt_handler.stop(); });
 
-            connection->sendQuery(
-                connection_parameters.timeouts,
-                query,
-                query_parameters,
-                client_context->getCurrentQueryId(),
-                query_processing_stage,
-                &client_context->getSettingsRef(),
-                &client_context->getClientInfo(),
-                true,
-                [&](const Progress & progress) { onProgress(progress); });
+            try {
+                connection->sendQuery(
+                    connection_parameters.timeouts,
+                    query,
+                    query_parameters,
+                    client_context->getCurrentQueryId(),
+                    query_processing_stage,
+                    &client_context->getSettingsRef(),
+                    &client_context->getClientInfo(),
+                    true,
+                    [&](const Progress & progress) { onProgress(progress); });
+
+                if (send_external_tables)
+                    sendExternalTables(parsed_query);
+            }
+            catch (const NetException &)
+            {
+                // We still want to attempt to process whatever we already received or can receive (socket receive buffer can be not empty)
+                receiveResult(parsed_query, signals_before_stop, settings.partial_result_on_first_cancel);
+                throw;
+            }
 
-            if (send_external_tables)
-                sendExternalTables(parsed_query);
             receiveResult(parsed_query, signals_before_stop, settings.partial_result_on_first_cancel);
 
             break;
@@ -2751,7 +2764,7 @@ void ClientBase::runLibFuzzer()
     for (auto & arg : fuzzer_args_holder)
         fuzzer_args.emplace_back(arg.data());
 
-    int fuzzer_argc = fuzzer_args.size();
+    int fuzzer_argc = static_cast<int>(fuzzer_args.size());
     char ** fuzzer_argv = fuzzer_args.data();
 
     LLVMFuzzerRunDriver(&fuzzer_argc, &fuzzer_argv, [](const uint8_t * data, size_t size)

src/Client/Connection.cpp

@@ -5,8 +5,6 @@
 #include <Core/Settings.h>
 #include <Compression/CompressedReadBuffer.h>
 #include <Compression/CompressedWriteBuffer.h>
-#include <IO/ReadBufferFromPocoSocket.h>
-#include <IO/WriteBufferFromPocoSocket.h>
 #include <IO/ReadHelpers.h>
 #include <IO/WriteHelpers.h>
 #include <IO/copyData.h>
@@ -85,6 +83,7 @@ Connection::~Connection()
 Connection::Connection(const String & host_, UInt16 port_,
     const String & default_database_,
     const String & user_, const String & password_,
+    const String & proto_send_chunked_, const String & proto_recv_chunked_,
     [[maybe_unused]] const SSHKey & ssh_private_key_,
     const String & jwt_,
     const String & quota_key_,
@@ -95,6 +94,7 @@ Connection::Connection(const String & host_, UInt16 port_,
     Protocol::Secure secure_)
     : host(host_), port(port_), default_database(default_database_)
     , user(user_), password(password_)
+    , proto_send_chunked(proto_send_chunked_), proto_recv_chunked(proto_recv_chunked_)
 #if USE_SSH
     , ssh_private_key(ssh_private_key_)
 #endif
@@ -211,10 +211,10 @@ void Connection::connect(const ConnectionTimeouts & timeouts)
                 , tcp_keep_alive_timeout_in_sec);
         }
 
-        in = std::make_shared<ReadBufferFromPocoSocket>(*socket);
+        in = std::make_shared<ReadBufferFromPocoSocketChunked>(*socket);
         in->setAsyncCallback(async_callback);
 
-        out = std::make_shared<WriteBufferFromPocoSocket>(*socket);
+        out = std::make_shared<WriteBufferFromPocoSocketChunked>(*socket);
         out->setAsyncCallback(async_callback);
         connected = true;
         setDescription();
@@ -222,9 +222,61 @@ void Connection::connect(const ConnectionTimeouts & timeouts)
         sendHello();
         receiveHello(timeouts.handshake_timeout);
 
+        if (server_revision >= DBMS_MIN_PROTOCOL_VERSION_WITH_CHUNKED_PACKETS)
+        {
+            /// Client side of chunked protocol negotiation.
+            /// Server advertises its protocol capabilities (separate for send and receive channels) by sending
+            /// in its 'Hello' response one of four types - chunked, notchunked, chunked_optional, notchunked_optional.
+            /// Not optional types are strict meaning that server only supports this type, optional means that
+            /// server prefer this type but capable to work in opposite.
+            /// Client selects which type it is going to communicate based on the settings from config or arguments,
+            /// and sends either "chunked" or "notchunked" protocol request in addendum section of handshake.
+            /// Client can detect if server's protocol capabilities are not compatible with client's settings (for example
+            /// server strictly requires chunked protocol but client's settings only allows notchunked protocol) - in such case
+            /// client should interrupt this connection. However if client continues with incompatible protocol type request, server
+            /// will send appropriate exception and disconnect client.
+
+            auto is_chunked = [](const String & chunked_srv_str, const String & chunked_cl_str, const String & direction)
+            {
+                bool chunked_srv = chunked_srv_str.starts_with("chunked");
+                bool optional_srv = chunked_srv_str.ends_with("_optional");
+                bool chunked_cl = chunked_cl_str.starts_with("chunked");
+                bool optional_cl = chunked_cl_str.ends_with("_optional");
+
+                if (optional_srv)
+                    return chunked_cl;
+                if (optional_cl)
+                    return chunked_srv;
+                if (chunked_cl != chunked_srv)
+                    throw NetException(
+                        ErrorCodes::NETWORK_ERROR,
+                        "Incompatible protocol: {} set to {}, server requires {}",
+                        direction,
+                        chunked_cl ? "chunked" : "notchunked",
+                        chunked_srv ? "chunked" : "notchunked");
+
+                return chunked_srv;
+            };
+
+            proto_send_chunked = is_chunked(proto_recv_chunked_srv, proto_send_chunked, "send") ? "chunked" : "notchunked";
+            proto_recv_chunked = is_chunked(proto_send_chunked_srv, proto_recv_chunked, "recv") ? "chunked" : "notchunked";
+        }
+        else
+        {
+            if (proto_send_chunked == "chunked" || proto_recv_chunked == "chunked")
+                throw NetException(
+                        ErrorCodes::NETWORK_ERROR,
+                        "Incompatible protocol: server's version is too old and doesn't support chunked protocol while client settings require it.");
+        }
+
         if (server_revision >= DBMS_MIN_PROTOCOL_VERSION_WITH_ADDENDUM)
             sendAddendum();
 
+        if (proto_send_chunked == "chunked")
+            out->enableChunked();
+        if (proto_recv_chunked == "chunked")
+            in->enableChunked();
+
         LOG_TRACE(log_wrapper.get(), "Connected to {} server version {}.{}.{}.",
             server_name, server_version_major, server_version_minor, server_version_patch);
     }
@@ -393,6 +445,13 @@ void Connection::sendAddendum()
 {
     if (server_revision >= DBMS_MIN_PROTOCOL_VERSION_WITH_QUOTA_KEY)
         writeStringBinary(quota_key, *out);
+
+    if (server_revision >= DBMS_MIN_PROTOCOL_VERSION_WITH_CHUNKED_PACKETS)
+    {
+        writeStringBinary(proto_send_chunked, *out);
+        writeStringBinary(proto_recv_chunked, *out);
+    }
+
     out->next();
 }
 
@@ -472,6 +531,12 @@ void Connection::receiveHello(const Poco::Timespan & handshake_timeout)
         else
             server_version_patch = server_revision;
 
+        if (server_revision >= DBMS_MIN_PROTOCOL_VERSION_WITH_CHUNKED_PACKETS)
+        {
+            readStringBinary(proto_send_chunked_srv, *in);
+            readStringBinary(proto_recv_chunked_srv, *in);
+        }
+
         if (server_revision >= DBMS_MIN_PROTOCOL_VERSION_WITH_PASSWORD_COMPLEXITY_RULES)
         {
             UInt64 rules_size;
@@ -611,6 +676,7 @@ bool Connection::ping(const ConnectionTimeouts & timeouts)
 
         UInt64 pong = 0;
         writeVarUInt(Protocol::Client::Ping, *out);
+        out->finishChunk();
         out->next();
 
         if (in->eof())
@@ -660,6 +726,7 @@ TablesStatusResponse Connection::getTablesStatus(const ConnectionTimeouts & time
 
     writeVarUInt(Protocol::Client::TablesStatusRequest, *out);
     request.write(*out, server_revision);
+    out->finishChunk();
     out->next();
 
     UInt64 response_type = 0;
@@ -813,6 +880,8 @@ void Connection::sendQuery(
     block_profile_events_in.reset();
     block_out.reset();
 
+    out->finishChunk();
+
     /// Send empty block which means end of data.
     if (!with_pending_data)
     {
@@ -829,6 +898,7 @@ void Connection::sendCancel()
         return;
 
     writeVarUInt(Protocol::Client::Cancel, *out);
+    out->finishChunk();
     out->next();
 }
 
@@ -854,7 +924,10 @@ void Connection::sendData(const Block & block, const String & name, bool scalar)
     size_t prev_bytes = out->count();
 
     block_out->write(block);
-    maybe_compressed_out->next();
+    if (maybe_compressed_out != out)
+        maybe_compressed_out->next();
+    if (!block)
+        out->finishChunk();
     out->next();
 
     if (throttler)
@@ -865,6 +938,7 @@ void Connection::sendIgnoredPartUUIDs(const std::vector<UUID> & uuids)
 {
     writeVarUInt(Protocol::Client::IgnoredPartUUIDs, *out);
     writeVectorBinary(uuids, *out);
+    out->finishChunk();
     out->next();
 }
 
@@ -874,6 +948,7 @@ void Connection::sendReadTaskResponse(const String & response)
     writeVarUInt(Protocol::Client::ReadTaskResponse, *out);
     writeVarUInt(DBMS_CLUSTER_PROCESSING_PROTOCOL_VERSION, *out);
     writeStringBinary(response, *out);
+    out->finishChunk();
     out->next();
 }
 
@@ -882,6 +957,7 @@ void Connection::sendMergeTreeReadTaskResponse(const ParallelReadResponse & resp
 {
     writeVarUInt(Protocol::Client::MergeTreeReadTaskResponse, *out);
     response.serialize(*out);
+    out->finishChunk();
     out->next();
 }
 
@@ -899,6 +975,8 @@ void Connection::sendPreparedData(ReadBuffer & input, size_t size, const String
         copyData(input, *out);
     else
         copyData(input, *out, size);
+
+    out->finishChunk();
     out->next();
 }
 
@@ -927,6 +1005,8 @@ void Connection::sendScalarsData(Scalars & data)
         sendData(elem.second, elem.first, true /* scalar */);
     }
 
+    out->finishChunk();
+
     out_bytes = out->count() - out_bytes;
     maybe_compressed_out_bytes = maybe_compressed_out->count() - maybe_compressed_out_bytes;
     double elapsed = watch.elapsedSeconds();
@@ -1069,13 +1149,13 @@ std::optional<Poco::Net::SocketAddress> Connection::getResolvedAddress() const
 
 bool Connection::poll(size_t timeout_microseconds)
 {
-    return static_cast<ReadBufferFromPocoSocket &>(*in).poll(timeout_microseconds);
+    return in->poll(timeout_microseconds);
 }
 
 
 bool Connection::hasReadPendingData() const
 {
-    return last_input_packet_type.has_value() || static_cast<const ReadBufferFromPocoSocket &>(*in).hasPendingData();
+    return last_input_packet_type.has_value() || in->hasBufferedData();
 }
 
 
@@ -1349,6 +1429,8 @@ ServerConnectionPtr Connection::createConnection(const ConnectionParameters & pa
         parameters.default_database,
         parameters.user,
         parameters.password,
+        parameters.proto_send_chunked,
+        parameters.proto_recv_chunked,
         parameters.ssh_private_key,
         parameters.jwt,
         parameters.quota_key,

src/Client/Connection.h

@@ -8,8 +8,8 @@
 #include <Core/Defines.h>
 
 
-#include <IO/ReadBufferFromPocoSocket.h>
-#include <IO/WriteBufferFromPocoSocket.h>
+#include <IO/ReadBufferFromPocoSocketChunked.h>
+#include <IO/WriteBufferFromPocoSocketChunked.h>
 
 #include <Interpreters/TablesStatus.h>
 #include <Interpreters/Context_fwd.h>
@@ -52,6 +52,7 @@ public:
     Connection(const String & host_, UInt16 port_,
         const String & default_database_,
         const String & user_, const String & password_,
+        const String & proto_send_chunked_, const String & proto_recv_chunked_,
         const SSHKey & ssh_private_key_,
         const String & jwt_,
         const String & quota_key_,
@@ -170,6 +171,10 @@ private:
     String default_database;
     String user;
     String password;
+    String proto_send_chunked;
+    String proto_recv_chunked;
+    String proto_send_chunked_srv;
+    String proto_recv_chunked_srv;
 #if USE_SSH
     SSHKey ssh_private_key;
 #endif
@@ -209,8 +214,8 @@ private:
     String server_display_name;
 
     std::unique_ptr<Poco::Net::StreamSocket> socket;
-    std::shared_ptr<ReadBufferFromPocoSocket> in;
-    std::shared_ptr<WriteBufferFromPocoSocket> out;
+    std::shared_ptr<ReadBufferFromPocoSocketChunked> in;
+    std::shared_ptr<WriteBufferFromPocoSocketChunked> out;
     std::optional<UInt64> last_input_packet_type;
 
     String query_id;

src/Client/ConnectionParameters.cpp

@@ -107,6 +107,9 @@ ConnectionParameters::ConnectionParameters(const Poco::Util::AbstractConfigurati
         }
     }
 
+    proto_send_chunked = config.getString("proto_caps.send", "notchunked");
+    proto_recv_chunked = config.getString("proto_caps.recv", "notchunked");
+
     quota_key = config.getString("quota_key", "");
 
     /// By default compression is disabled if address looks like localhost.

src/Client/ConnectionParameters.h

@@ -20,6 +20,8 @@ struct ConnectionParameters
     std::string default_database;
     std::string user;
     std::string password;
+    std::string proto_send_chunked = "notchunked";
+    std::string proto_recv_chunked = "notchunked";
     std::string quota_key;
     SSHKey ssh_private_key;
     std::string jwt;

src/Client/ConnectionPool.cpp

@@ -13,6 +13,8 @@ ConnectionPoolPtr ConnectionPoolFactory::get(
     String default_database,
     String user,
     String password,
+    String proto_send_chunked,
+    String proto_recv_chunked,
     String quota_key,
     String cluster,
     String cluster_secret,
@@ -22,7 +24,7 @@ ConnectionPoolPtr ConnectionPoolFactory::get(
     Priority priority)
 {
     Key key{
-        max_connections, host, port, default_database, user, password, quota_key, cluster, cluster_secret, client_name, compression, secure, priority};
+        max_connections, host, port, default_database, user, password, proto_send_chunked, proto_recv_chunked, quota_key, cluster, cluster_secret, client_name, compression, secure, priority};
 
     std::lock_guard lock(mutex);
     auto [it, inserted] = pools.emplace(key, ConnectionPoolPtr{});
@@ -39,6 +41,8 @@ ConnectionPoolPtr ConnectionPoolFactory::get(
             default_database,
             user,
             password,
+            proto_send_chunked,
+            proto_recv_chunked,
             quota_key,
             cluster,
             cluster_secret,

src/Client/ConnectionPool.h

@@ -73,6 +73,8 @@ public:
         const String & default_database_,
         const String & user_,
         const String & password_,
+        const String & proto_send_chunked_,
+        const String & proto_recv_chunked_,
         const String & quota_key_,
         const String & cluster_,
         const String & cluster_secret_,
@@ -85,6 +87,8 @@ public:
         , default_database(default_database_)
         , user(user_)
         , password(password_)
+        , proto_send_chunked(proto_send_chunked_)
+        , proto_recv_chunked(proto_recv_chunked_)
         , quota_key(quota_key_)
         , cluster(cluster_)
         , cluster_secret(cluster_secret_)
@@ -116,7 +120,9 @@ protected:
     {
         return std::make_shared<Connection>(
             host, port,
-            default_database, user, password, SSHKey(), /*jwt*/ "", quota_key,
+            default_database, user, password,
+            proto_send_chunked, proto_recv_chunked,
+            SSHKey(), /*jwt*/ "", quota_key,
             cluster, cluster_secret,
             client_name, compression, secure);
     }
@@ -125,6 +131,8 @@ private:
     String default_database;
     String user;
     String password;
+    String proto_send_chunked;
+    String proto_recv_chunked;
     String quota_key;
 
     /// For inter-server authorization
@@ -150,6 +158,8 @@ public:
         String default_database;
         String user;
         String password;
+        String proto_send_chunked;
+        String proto_recv_chunked;
         String quota_key;
         String cluster;
         String cluster_secret;
@@ -173,6 +183,8 @@ public:
         String default_database,
         String user,
         String password,
+        String proto_send_chunked,
+        String proto_recv_chunked,
         String quota_key,
         String cluster,
         String cluster_secret,
@@ -190,6 +202,7 @@ inline bool operator==(const ConnectionPoolFactory::Key & lhs, const ConnectionP
 {
     return lhs.max_connections == rhs.max_connections && lhs.host == rhs.host && lhs.port == rhs.port
         && lhs.default_database == rhs.default_database && lhs.user == rhs.user && lhs.password == rhs.password
+        && lhs.proto_send_chunked == rhs.proto_send_chunked && lhs.proto_recv_chunked == rhs.proto_recv_chunked
         && lhs.quota_key == rhs.quota_key
         && lhs.cluster == rhs.cluster && lhs.cluster_secret == rhs.cluster_secret && lhs.client_name == rhs.client_name
         && lhs.compression == rhs.compression && lhs.secure == rhs.secure && lhs.priority == rhs.priority;

src/Columns/ColumnDynamic.h

@@ -3,6 +3,7 @@
 #include <Columns/IColumn.h>
 #include <Columns/ColumnVector.h>
 #include <Columns/ColumnVariant.h>
+#include <Columns/ColumnString.h>
 #include <DataTypes/IDataType.h>
 #include <Common/WeakHash.h>
 
@@ -19,11 +20,19 @@ namespace DB
  *
  * When new values are inserted into Dynamic column, the internal Variant
  * type and column are extended if the inserted value has new type.
+ * When the limit on number of dynamic types is exceeded, all values
+ * with new types are inserted into special shared variant with type String
+ * that contains values and their types in binary format.
  */
 class ColumnDynamic final : public COWHelper<IColumnHelper<ColumnDynamic>, ColumnDynamic>
 {
 public:
-    ///
+    /// Maximum limit on dynamic types. We use ColumnVariant to store all the types,
+    /// so the limit cannot be greater then ColumnVariant::MAX_NESTED_COLUMNS.
+    /// We also always have reserved variant for shared variant.
+    static constexpr size_t MAX_DYNAMIC_TYPES_LIMIT = ColumnVariant::MAX_NESTED_COLUMNS - 1;
+    static constexpr const char * SHARED_VARIANT_TYPE_NAME = "SharedVariant";
+
     struct Statistics
     {
         enum class Source
@@ -32,12 +41,27 @@ public:
             MERGE, /// Statistics were calculated during merge of several MergeTree parts.
         };
 
+        explicit Statistics(Source source_) : source(source_) {}
+
         /// Source of the statistics.
         Source source;
-        /// Statistics data: (variant name) -> (total variant size in data part).
-        std::unordered_map<String, size_t> data;
+        /// Statistics data for usual variants: (variant name) -> (total variant size in data part).
+        std::unordered_map<String, size_t> variants_statistics;
+        /// Statistics data for variants from shared variant: (variant name) -> (total variant size in data part).
+        /// For shared variant we store statistics only for first 256 variants (should cover almost all cases and it's not expensive).
+        static constexpr const size_t MAX_SHARED_VARIANT_STATISTICS_SIZE = 256;
+        std::unordered_map<String, size_t> shared_variants_statistics;
     };
 
+    using StatisticsPtr = std::shared_ptr<const Statistics>;
+
+    struct ComparatorBase;
+    using ComparatorAscendingUnstable = ComparatorAscendingUnstableImpl<ComparatorBase>;
+    using ComparatorAscendingStable = ComparatorAscendingStableImpl<ComparatorBase>;
+    using ComparatorDescendingUnstable = ComparatorDescendingUnstableImpl<ComparatorBase>;
+    using ComparatorDescendingStable = ComparatorDescendingStableImpl<ComparatorBase>;
+    using ComparatorEqual = ComparatorEqualImpl<ComparatorBase>;
+
 private:
     friend class COWHelper<IColumnHelper<ColumnDynamic>, ColumnDynamic>;
 
@@ -54,28 +78,32 @@ private:
     };
 
     explicit ColumnDynamic(size_t max_dynamic_types_);
-    ColumnDynamic(MutableColumnPtr variant_column_, const VariantInfo & variant_info_, size_t max_dynamic_types_, const Statistics & statistics_ = {});
+    ColumnDynamic(MutableColumnPtr variant_column_, const DataTypePtr & variant_type_, size_t max_dynamic_types_, size_t global_max_dynamic_types_, const StatisticsPtr & statistics_ = {});
+    ColumnDynamic(MutableColumnPtr variant_column_, const VariantInfo & variant_info_, size_t max_dynamic_types_, size_t global_max_dynamic_types_, const StatisticsPtr & statistics_ = {});
 
 public:
     /** Create immutable column using immutable arguments. This arguments may be shared with other columns.
       * Use IColumn::mutate in order to make mutable column and mutate shared nested columns.
       */
     using Base = COWHelper<IColumnHelper<ColumnDynamic>, ColumnDynamic>;
-    static Ptr create(const ColumnPtr & variant_column_, const VariantInfo & variant_info_, size_t max_dynamic_types_, const Statistics & statistics_ = {})
+    static Ptr create(const ColumnPtr & variant_column_, const VariantInfo & variant_info_, size_t max_dynamic_types_, size_t global_max_dynamic_types_, const StatisticsPtr & statistics_ = {})
     {
-        return ColumnDynamic::create(variant_column_->assumeMutable(), variant_info_, max_dynamic_types_, statistics_);
+        return ColumnDynamic::create(variant_column_->assumeMutable(), variant_info_, max_dynamic_types_, global_max_dynamic_types_, statistics_);
     }
 
-    static MutablePtr create(MutableColumnPtr variant_column_, const VariantInfo & variant_info_, size_t max_dynamic_types_, const Statistics & statistics_ = {})
+    static MutablePtr create(MutableColumnPtr variant_column_, const VariantInfo & variant_info_, size_t max_dynamic_types_, size_t global_max_dynamic_types_, const StatisticsPtr & statistics_ = {})
     {
-        return Base::create(std::move(variant_column_), variant_info_, max_dynamic_types_, statistics_);
+        return Base::create(std::move(variant_column_), variant_info_, max_dynamic_types_, global_max_dynamic_types_, statistics_);
     }
 
-    static MutablePtr create(MutableColumnPtr variant_column_, const DataTypePtr & variant_type, size_t max_dynamic_types_, const Statistics & statistics_ = {});
-
-    static ColumnPtr create(ColumnPtr variant_column_, const DataTypePtr & variant_type, size_t max_dynamic_types_, const Statistics & statistics_ = {})
+    static MutablePtr create(MutableColumnPtr variant_column_, const DataTypePtr & variant_type_, size_t max_dynamic_types_, size_t global_max_dynamic_types_, const StatisticsPtr & statistics_ = {})
     {
-        return create(variant_column_->assumeMutable(), variant_type, max_dynamic_types_, statistics_);
+        return Base::create(std::move(variant_column_), variant_type_, max_dynamic_types_, global_max_dynamic_types_, statistics_);
+    }
+
+    static ColumnPtr create(ColumnPtr variant_column_, const DataTypePtr & variant_type, size_t max_dynamic_types_, size_t global_max_dynamic_types_, const StatisticsPtr & statistics_ = {})
+    {
+        return create(variant_column_->assumeMutable(), variant_type, max_dynamic_types_, global_max_dynamic_types_, statistics_);
     }
 
     static MutablePtr create(size_t max_dynamic_types_)
@@ -83,7 +111,7 @@ public:
         return Base::create(max_dynamic_types_);
     }
 
-    std::string getName() const override { return "Dynamic(max_types=" + std::to_string(max_dynamic_types) + ")"; }
+    std::string getName() const override { return "Dynamic(max_types=" + std::to_string(global_max_dynamic_types) + ")"; }
 
     const char * getFamilyName() const override
     {
@@ -98,12 +126,12 @@ public:
     MutableColumnPtr cloneEmpty() const override
     {
         /// Keep current dynamic structure
-        return Base::create(variant_column->cloneEmpty(), variant_info, max_dynamic_types, statistics);
+        return Base::create(variant_column->cloneEmpty(), variant_info, max_dynamic_types, global_max_dynamic_types, statistics);
     }
 
     MutableColumnPtr cloneResized(size_t size) const override
     {
-        return Base::create(variant_column->cloneResized(size), variant_info, max_dynamic_types, statistics);
+        return Base::create(variant_column->cloneResized(size), variant_info, max_dynamic_types, global_max_dynamic_types, statistics);
     }
 
     size_t size() const override
@@ -111,15 +139,9 @@ public:
         return variant_column->size();
     }
 
-    Field operator[](size_t n) const override
-    {
-        return (*variant_column)[n];
-    }
+    Field operator[](size_t n) const override;
 
-    void get(size_t n, Field & res) const override
-    {
-        variant_column->get(n, res);
-    }
+    void get(size_t n, Field & res) const override;
 
     bool isDefaultAt(size_t n) const override
     {
@@ -187,7 +209,7 @@ public:
 
     ColumnPtr filter(const Filter & filt, ssize_t result_size_hint) const override
     {
-        return create(variant_column->filter(filt, result_size_hint), variant_info, max_dynamic_types);
+        return create(variant_column->filter(filt, result_size_hint), variant_info, max_dynamic_types, global_max_dynamic_types);
     }
 
     void expand(const Filter & mask, bool inverted) override
@@ -197,17 +219,17 @@ public:
 
     ColumnPtr permute(const Permutation & perm, size_t limit) const override
     {
-        return create(variant_column->permute(perm, limit), variant_info, max_dynamic_types);
+        return create(variant_column->permute(perm, limit), variant_info, max_dynamic_types, global_max_dynamic_types);
     }
 
     ColumnPtr index(const IColumn & indexes, size_t limit) const override
     {
-        return create(variant_column->index(indexes, limit), variant_info, max_dynamic_types);
+        return create(variant_column->index(indexes, limit), variant_info, max_dynamic_types, global_max_dynamic_types);
     }
 
     ColumnPtr replicate(const Offsets & replicate_offsets) const override
     {
-        return create(variant_column->replicate(replicate_offsets), variant_info, max_dynamic_types);
+        return create(variant_column->replicate(replicate_offsets), variant_info, max_dynamic_types, global_max_dynamic_types);
     }
 
     MutableColumns scatter(ColumnIndex num_columns, const Selector & selector) const override
@@ -216,7 +238,7 @@ public:
         MutableColumns scattered_columns;
         scattered_columns.reserve(num_columns);
         for (auto & scattered_variant_column : scattered_variant_columns)
-            scattered_columns.emplace_back(create(std::move(scattered_variant_column), variant_info, max_dynamic_types));
+            scattered_columns.emplace_back(create(std::move(scattered_variant_column), variant_info, max_dynamic_types, global_max_dynamic_types));
 
         return scattered_columns;
     }
@@ -238,16 +260,10 @@ public:
     }
 
     void getPermutation(IColumn::PermutationSortDirection direction, IColumn::PermutationSortStability stability,
-                        size_t limit, int nan_direction_hint, IColumn::Permutation & res) const override
-    {
-        variant_column->getPermutation(direction, stability, limit, nan_direction_hint, res);
-    }
+                        size_t limit, int nan_direction_hint, IColumn::Permutation & res) const override;
 
     void updatePermutation(IColumn::PermutationSortDirection direction, IColumn::PermutationSortStability stability,
-                           size_t limit, int nan_direction_hint, IColumn::Permutation & res, EqualRanges & equal_ranges) const override
-    {
-        variant_column->updatePermutation(direction, stability, limit, nan_direction_hint, res, equal_ranges);
-    }
+                           size_t limit, int nan_direction_hint, IColumn::Permutation & res, EqualRanges & equal_ranges) const override;
 
     void reserve(size_t n) override
     {
@@ -295,7 +311,7 @@ public:
     bool structureEquals(const IColumn & rhs) const override
     {
         if (const auto * rhs_concrete = typeid_cast<const ColumnDynamic *>(&rhs))
-            return max_dynamic_types == rhs_concrete->max_dynamic_types;
+            return global_max_dynamic_types == rhs_concrete->global_max_dynamic_types;
         return false;
     }
 
@@ -338,17 +354,75 @@ public:
     const ColumnVariant & getVariantColumn() const { return assert_cast<const ColumnVariant &>(*variant_column); }
     ColumnVariant & getVariantColumn() { return assert_cast<ColumnVariant &>(*variant_column); }
 
-    bool addNewVariant(const DataTypePtr & new_variant);
-    void addStringVariant();
+    bool addNewVariant(const DataTypePtr & new_variant, const String & new_variant_name);
+    bool addNewVariant(const DataTypePtr & new_variant) { return addNewVariant(new_variant, new_variant->getName()); }
 
     bool hasDynamicStructure() const override { return true; }
     void takeDynamicStructureFromSourceColumns(const Columns & source_columns) override;
 
-    const Statistics & getStatistics() const { return statistics; }
+    const StatisticsPtr & getStatistics() const { return statistics; }
+    void setStatistics(const StatisticsPtr & statistics_) { statistics = statistics_; }
 
     size_t getMaxDynamicTypes() const { return max_dynamic_types; }
 
+    /// Check if we can add new variant types.
+    /// Shared variant doesn't count in the limit but always presents,
+    /// so we should subtract 1 from the total types count.
+    bool canAddNewVariants(size_t current_variants_count, size_t new_variants_count) const { return current_variants_count + new_variants_count - 1 <= max_dynamic_types; }
+    bool canAddNewVariant(size_t current_variants_count) const { return canAddNewVariants(current_variants_count, 1); }
+    bool canAddNewVariants(size_t new_variants_count) const { return canAddNewVariants(variant_info.variant_names.size(), new_variants_count); }
+    bool canAddNewVariant() const { return canAddNewVariants(variant_info.variant_names.size(), 1); }
+
+    void setVariantType(const DataTypePtr & variant_type);
+    void setMaxDynamicPaths(size_t max_dynamic_type_);
+
+    static const String & getSharedVariantTypeName()
+    {
+        static const String name = SHARED_VARIANT_TYPE_NAME;
+        return name;
+    }
+
+    static DataTypePtr getSharedVariantDataType();
+
+    ColumnVariant::Discriminator getSharedVariantDiscriminator() const
+    {
+        return variant_info.variant_name_to_discriminator.at(getSharedVariantTypeName());
+    }
+
+    ColumnString & getSharedVariant()
+    {
+        return assert_cast<ColumnString &>(getVariantColumn().getVariantByGlobalDiscriminator(getSharedVariantDiscriminator()));
+    }
+
+    const ColumnString & getSharedVariant() const
+    {
+        return assert_cast<const ColumnString &>(getVariantColumn().getVariantByGlobalDiscriminator(getSharedVariantDiscriminator()));
+    }
+
+    /// Serializes type and value in binary format into provided shared variant. Doesn't update Variant discriminators and offsets.
+    static void serializeValueIntoSharedVariant(ColumnString & shared_variant, const IColumn & src, const DataTypePtr & type, const SerializationPtr & serialization, size_t n);
+
+    /// Insert value into shared variant. Also updates Variant discriminators and offsets.
+    void insertValueIntoSharedVariant(const IColumn & src, const DataTypePtr & type, const String & type_name, size_t n);
+
+    const SerializationPtr & getVariantSerialization(const DataTypePtr & variant_type, const String & variant_name) const
+    {
+        /// Get serialization for provided data type.
+        /// To avoid calling type->getDefaultSerialization() every time we use simple cache with max size.
+        /// When max size is reached, just clear the cache.
+        if (serialization_cache.size() == SERIALIZATION_CACHE_MAX_SIZE)
+            serialization_cache.clear();
+
+        if (auto it = serialization_cache.find(variant_name); it != serialization_cache.end())
+            return it->second;
+
+        return serialization_cache.emplace(variant_name, variant_type->getDefaultSerialization()).first->second;
+    }
+
+    const SerializationPtr & getVariantSerialization(const DataTypePtr & variant_type) const { return getVariantSerialization(variant_type, variant_type->getName()); }
 private:
+    void createVariantInfo(const DataTypePtr & variant_type);
+
     /// Combine current variant with the other variant and return global discriminators mapping
     /// from other variant to the combined one. It's used for inserting from
     /// different variants.
@@ -361,12 +435,19 @@ private:
     /// Store the type of current variant with some additional information.
     VariantInfo variant_info;
     /// The maximum number of different types that can be stored in this Dynamic column.
-    /// If exceeded, all new variants will be converted to String.
+    /// If exceeded, all new variants will be added to a special shared variant with type String
+    /// in binary format. This limit can be different for different instances of Dynamic column.
+    /// When max_dynamic_types = 0, we will have only shared variant and insert all values into it.
     size_t max_dynamic_types;
+    /// The types limit specified in the data type by the user Dynamic(max_types=N).
+    /// max_dynamic_types in all column instances of this Dynamic type can be only smaller
+    /// (for example, max_dynamic_types can be reduced in takeDynamicStructureFromSourceColumns
+    /// before merge of different Dynamic columns).
+    size_t global_max_dynamic_types;
 
     /// Size statistics of each variants from MergeTree data part.
     /// Used in takeDynamicStructureFromSourceColumns and set during deserialization.
-    Statistics statistics;
+    StatisticsPtr statistics;
 
     /// Cache (Variant name) -> (global discriminators mapping from this variant to current variant in Dynamic column).
     /// Used to avoid mappings recalculation in combineVariants for the same Variant types.
@@ -374,6 +455,17 @@ private:
     /// Cache of Variant types that couldn't be combined with current variant in Dynamic column.
     /// Used to avoid checking if combination is possible for the same Variant types.
     std::unordered_set<String> variants_with_failed_combination;
+
+    /// We can use serializations of different data types to serialize values into shared variant.
+    /// To avoid creating the same serialization multiple times, use simple cache.
+    static const size_t SERIALIZATION_CACHE_MAX_SIZE = 256;
+    mutable std::unordered_map<String, SerializationPtr> serialization_cache;
 };
 
+void extendVariantColumn(
+    IColumn & variant_column,
+    const DataTypePtr & old_variant_type,
+    const DataTypePtr & new_variant_type,
+    std::unordered_map<String, UInt8> old_variant_name_to_discriminator);
+
 }

src/Columns/ColumnVariant.cpp

@@ -476,7 +476,7 @@ void ColumnVariant::insertFromImpl(const DB::IColumn & src_, size_t n, const std
     }
 }
 
-void ColumnVariant::insertRangeFromImpl(const DB::IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> * global_discriminators_mapping)
+void ColumnVariant::insertRangeFromImpl(const DB::IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> * global_discriminators_mapping, const Discriminator * skip_discriminator)
 {
     const size_t num_variants = variants.size();
     const auto & src = assert_cast<const ColumnVariant &>(src_);
@@ -557,9 +557,12 @@ void ColumnVariant::insertRangeFromImpl(const DB::IColumn & src_, size_t start,
         Discriminator global_discr = src_global_discr;
         if (global_discriminators_mapping && src_global_discr != NULL_DISCRIMINATOR)
             global_discr = (*global_discriminators_mapping)[src_global_discr];
-        Discriminator local_discr = localDiscriminatorByGlobal(global_discr);
-        if (nested_length)
-            variants[local_discr]->insertRangeFrom(*src.variants[src_local_discr], nested_start, nested_length);
+        if (!skip_discriminator || global_discr != *skip_discriminator)
+        {
+            Discriminator local_discr = localDiscriminatorByGlobal(global_discr);
+            if (nested_length)
+                variants[local_discr]->insertRangeFrom(*src.variants[src_local_discr], nested_start, nested_length);
+        }
     }
 }
 
@@ -610,7 +613,7 @@ void ColumnVariant::insertRangeFrom(const IColumn & src_, size_t start, size_t l
 void ColumnVariant::doInsertRangeFrom(const IColumn & src_, size_t start, size_t length)
 #endif
 {
-    insertRangeFromImpl(src_, start, length, nullptr);
+    insertRangeFromImpl(src_, start, length, nullptr, nullptr);
 }
 
 #if !defined(DEBUG_OR_SANITIZER_BUILD)
@@ -627,9 +630,9 @@ void ColumnVariant::insertFrom(const DB::IColumn & src_, size_t n, const std::ve
     insertFromImpl(src_, n, &global_discriminators_mapping);
 }
 
-void ColumnVariant::insertRangeFrom(const IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping)
+void ColumnVariant::insertRangeFrom(const IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping, Discriminator skip_discriminator)
 {
-    insertRangeFromImpl(src_, start, length, &global_discriminators_mapping);
+    insertRangeFromImpl(src_, start, length, &global_discriminators_mapping, &skip_discriminator);
 }
 
 void ColumnVariant::insertManyFrom(const DB::IColumn & src_, size_t position, size_t length, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping)
@@ -673,6 +676,14 @@ void ColumnVariant::insertManyIntoVariantFrom(DB::ColumnVariant::Discriminator g
     variants[local_discr]->insertManyFrom(src_, position, length);
 }
 
+void ColumnVariant::deserializeBinaryIntoVariant(ColumnVariant::Discriminator global_discr, const SerializationPtr & serialization, ReadBuffer & buf, const FormatSettings & format_settings)
+{
+    auto local_discr = localDiscriminatorByGlobal(global_discr);
+    serialization->deserializeBinary(*variants[local_discr], buf, format_settings);
+    getLocalDiscriminators().push_back(local_discr);
+    getOffsets().push_back(variants[local_discr]->size() - 1);
+}
+
 void ColumnVariant::insertDefault()
 {
     getLocalDiscriminators().push_back(NULL_DISCRIMINATOR);
@@ -1213,9 +1224,7 @@ struct ColumnVariant::ComparatorBase
 
     ALWAYS_INLINE int compare(size_t lhs, size_t rhs) const
     {
-        int res = parent.compareAt(lhs, rhs, parent, nan_direction_hint);
-
-        return res;
+        return parent.compareAt(lhs, rhs, parent, nan_direction_hint);
     }
 };
 

src/Columns/ColumnVariant.h

@@ -2,6 +2,8 @@
 
 #include <Columns/IColumn.h>
 #include <Columns/ColumnVector.h>
+#include <Formats/FormatSettings.h>
+#include <DataTypes/Serializations/ISerialization.h>
 
 
 namespace DB
@@ -196,13 +198,15 @@ public:
 
     /// Methods for insertion from another Variant but with known mapping between global discriminators.
     void insertFrom(const IColumn & src_, size_t n, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping);
-    void insertRangeFrom(const IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping);
+    /// Don't insert data into variant with skip_discriminator global discriminator, it will be processed separately.
+    void insertRangeFrom(const IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping, Discriminator skip_discriminator);
     void insertManyFrom(const IColumn & src_, size_t position, size_t length, const std::vector<ColumnVariant::Discriminator> & global_discriminators_mapping);
 
     /// Methods for insertion into a specific variant.
     void insertIntoVariantFrom(Discriminator global_discr, const IColumn & src_, size_t n);
     void insertRangeIntoVariantFrom(Discriminator global_discr, const IColumn & src_, size_t start, size_t length);
     void insertManyIntoVariantFrom(Discriminator global_discr, const IColumn & src_, size_t position, size_t length);
+    void deserializeBinaryIntoVariant(Discriminator global_discr, const SerializationPtr & serialization, ReadBuffer & buf, const FormatSettings & format_settings);
 
     void insertDefault() override;
     void insertManyDefaults(size_t length) override;
@@ -264,6 +268,7 @@ public:
     ColumnPtr & getVariantPtrByGlobalDiscriminator(size_t discr) { return variants[global_to_local_discriminators.at(discr)]; }
 
     const NestedColumns & getVariants() const { return variants; }
+    NestedColumns & getVariants() { return variants; }
 
     const IColumn & getLocalDiscriminatorsColumn() const { return *local_discriminators; }
     IColumn & getLocalDiscriminatorsColumn() { return *local_discriminators; }
@@ -303,6 +308,8 @@ public:
         return true;
     }
 
+    std::vector<Discriminator> getLocalToGlobalDiscriminatorsMapping() const { return local_to_global_discriminators; }
+
     /// Check if we have only 1 non-empty variant and no NULL values,
     /// and if so, return the discriminator of this non-empty column.
     std::optional<Discriminator> getLocalDiscriminatorOfOneNoneEmptyVariantNoNulls() const;
@@ -323,7 +330,7 @@ public:
 
 private:
     void insertFromImpl(const IColumn & src_, size_t n, const std::vector<ColumnVariant::Discriminator> * global_discriminators_mapping);
-    void insertRangeFromImpl(const IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> * global_discriminators_mapping);
+    void insertRangeFromImpl(const IColumn & src_, size_t start, size_t length, const std::vector<ColumnVariant::Discriminator> * global_discriminators_mapping, const Discriminator * skip_discriminator);
     void insertManyFromImpl(const IColumn & src_, size_t position, size_t length, const std::vector<ColumnVariant::Discriminator> * global_discriminators_mapping);
 
     void initIdentityGlobalToLocalDiscriminatorsMapping();

src/Columns/tests/gtest_column_dynamic.cpp

@@ -7,28 +7,34 @@ using namespace DB;
 
 TEST(ColumnDynamic, CreateEmpty)
 {
-    auto column = ColumnDynamic::create(255);
+    auto column = ColumnDynamic::create(254);
     ASSERT_TRUE(column->empty());
-    ASSERT_EQ(column->getVariantInfo().variant_type->getName(), "Variant()");
-    ASSERT_TRUE(column->getVariantInfo().variant_names.empty());
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.empty());
+    ASSERT_EQ(column->getVariantInfo().variant_type->getName(), "Variant(SharedVariant)");
+    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 1);
+    ASSERT_EQ(column->getVariantInfo().variant_names[0], "SharedVariant");
+    ASSERT_EQ(column->getVariantInfo().variant_name_to_discriminator.size(), 1);
+    ASSERT_EQ(column->getVariantInfo().variant_name_to_discriminator.at("SharedVariant"), 0);
+    ASSERT_TRUE(column->getVariantColumn().getVariantByGlobalDiscriminator(0).empty());
 }
 
 TEST(ColumnDynamic, InsertDefault)
 {
-    auto column = ColumnDynamic::create(255);
+    auto column = ColumnDynamic::create(254);
     column->insertDefault();
     ASSERT_TRUE(column->size() == 1);
-    ASSERT_EQ(column->getVariantInfo().variant_type->getName(), "Variant()");
-    ASSERT_TRUE(column->getVariantInfo().variant_names.empty());
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.empty());
+    ASSERT_EQ(column->getVariantInfo().variant_type->getName(), "Variant(SharedVariant)");
+    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 1);
+    ASSERT_EQ(column->getVariantInfo().variant_names[0], "SharedVariant");
+    ASSERT_EQ(column->getVariantInfo().variant_name_to_discriminator.size(), 1);
+    ASSERT_EQ(column->getVariantInfo().variant_name_to_discriminator.at("SharedVariant"), 0);
+    ASSERT_TRUE(column->getVariantColumn().getVariantByGlobalDiscriminator(0).empty());
     ASSERT_TRUE(column->isNullAt(0));
     ASSERT_EQ((*column)[0], Field(Null()));
 }
 
 TEST(ColumnDynamic, InsertFields)
 {
-    auto column = ColumnDynamic::create(255);
+    auto column = ColumnDynamic::create(254);
     column->insert(Field(42));
     column->insert(Field(-42));
     column->insert(Field("str1"));
@@ -41,16 +47,16 @@ TEST(ColumnDynamic, InsertFields)
     column->insert(Field(43.43));
     ASSERT_TRUE(column->size() == 10);
 
-    ASSERT_EQ(column->getVariantInfo().variant_type->getName(), "Variant(Float64, Int8, String)");
-    std::vector<String> expected_names = {"Float64", "Int8", "String"};
+    ASSERT_EQ(column->getVariantInfo().variant_type->getName(), "Variant(Float64, Int8, SharedVariant, String)");
+    std::vector<String> expected_names = {"Float64", "Int8", "SharedVariant", "String"};
     ASSERT_EQ(column->getVariantInfo().variant_names, expected_names);
-    std::unordered_map<String, UInt8> expected_variant_name_to_discriminator = {{"Float64", 0}, {"Int8", 1}, {"String", 2}};
+    std::unordered_map<String, UInt8> expected_variant_name_to_discriminator = {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}};
     ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator == expected_variant_name_to_discriminator);
 }
 
 ColumnDynamic::MutablePtr getDynamicWithManyVariants(size_t num_variants, Field tuple_element = Field(42))
 {
-    auto column = ColumnDynamic::create(255);
+    auto column = ColumnDynamic::create(254);
     for (size_t i = 0; i != num_variants; ++i)
     {
         Tuple tuple;
@@ -66,61 +72,71 @@ TEST(ColumnDynamic, InsertFieldsOverflow1)
 {
     auto column = getDynamicWithManyVariants(253);
 
-    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 253);
+    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 254);
 
     column->insert(Field(42.42));
-    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 254);
+    ASSERT_EQ(column->size(), 254);
+    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
     ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
 
     column->insert(Field(42));
+    ASSERT_EQ(column->size(), 255);
     ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column->getSharedVariant().size(), 1);
     Field field = (*column)[column->size() - 1];
-    ASSERT_EQ(field, "42");
+    ASSERT_EQ(field, 42);
 
     column->insert(Field(43));
+    ASSERT_EQ(column->size(), 256);
     ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column->getSharedVariant().size(), 2);
     field = (*column)[column->size() - 1];
-    ASSERT_EQ(field, "43");
+    ASSERT_EQ(field, 43);
 
     column->insert(Field("str1"));
+    ASSERT_EQ(column->size(), 257);
     ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column->getSharedVariant().size(), 3);
     field = (*column)[column->size() - 1];
     ASSERT_EQ(field, "str1");
 
     column->insert(Field(Array({Field(42), Field(43)})));
     ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("Array(Int8)"));
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column->getSharedVariant().size(), 4);
     field = (*column)[column->size() - 1];
-    ASSERT_EQ(field, "[42, 43]");
+    ASSERT_EQ(field, Field(Array({Field(42), Field(43)})));
 }
 
 TEST(ColumnDynamic, InsertFieldsOverflow2)
 {
     auto column = getDynamicWithManyVariants(254);
-    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 254);
+    ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
 
     column->insert(Field("str1"));
     ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column->getSharedVariant().size(), 1);
+    Field field = (*column)[column->size() - 1];
+    ASSERT_EQ(field, "str1");
 
     column->insert(Field(42));
     ASSERT_EQ(column->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
-    Field field = (*column)[column->size() - 1];
-    ASSERT_EQ(field, "42");
+    ASSERT_FALSE(column->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column->getSharedVariant().size(), 2);
+    field = (*column)[column->size() - 1];
+    ASSERT_EQ(field, 42);
 }
 
 ColumnDynamic::MutablePtr getInsertFromColumn(size_t num = 1)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     for (size_t i = 0; i != num; ++i)
     {
         column_from->insert(Field(42));
@@ -154,41 +170,41 @@ void checkInsertFrom(const ColumnDynamic::MutablePtr & column_from, ColumnDynami
 
 TEST(ColumnDynamic, InsertFrom1)
 {
-    auto column_to = ColumnDynamic::create(255);
-    checkInsertFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, String)", {"Float64", "Int8", "String"}, {{"Float64", 0}, {"Int8", 1}, {"String", 2}});
+    auto column_to = ColumnDynamic::create(254);
+    checkInsertFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, SharedVariant, String)", {"Float64", "Int8", "SharedVariant", "String"}, {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}});
 }
 
 TEST(ColumnDynamic, InsertFrom2)
 {
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     column_to->insert(Field(42));
     column_to->insert(Field(42.42));
     column_to->insert(Field("str"));
 
-    checkInsertFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, String)", {"Float64", "Int8", "String"}, {{"Float64", 0}, {"Int8", 1}, {"String", 2}});
+    checkInsertFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, SharedVariant, String)", {"Float64", "Int8", "SharedVariant", "String"}, {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}});
 }
 
 TEST(ColumnDynamic, InsertFrom3)
 {
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     column_to->insert(Field(42));
     column_to->insert(Field(42.42));
     column_to->insert(Field("str"));
     column_to->insert(Array({Field(42)}));
 
-    checkInsertFrom(getInsertFromColumn(), column_to, "Variant(Array(Int8), Float64, Int8, String)", {"Array(Int8)", "Float64", "Int8", "String"}, {{"Array(Int8)", 0}, {"Float64", 1}, {"Int8", 2}, {"String", 3}});
+    checkInsertFrom(getInsertFromColumn(), column_to, "Variant(Array(Int8), Float64, Int8, SharedVariant, String)", {"Array(Int8)", "Float64", "Int8", "SharedVariant", "String"}, {{"Array(Int8)", 0}, {"Float64", 1}, {"Int8", 2}, {"SharedVariant", 3}, {"String", 4}});
 }
 
 TEST(ColumnDynamic, InsertFromOverflow1)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
     column_from->insert(Field("str"));
 
     auto column_to = getDynamicWithManyVariants(253);
     column_to->insertFrom(*column_from, 0);
-    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 254);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
     auto field = (*column_to)[column_to->size() - 1];
     ASSERT_EQ(field, 42);
@@ -196,20 +212,22 @@ TEST(ColumnDynamic, InsertFromOverflow1)
     column_to->insertFrom(*column_from, 1);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 1);
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, "42.42");
+    ASSERT_EQ(field, 42.42);
 
     column_to->insertFrom(*column_from, 2);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
     field = (*column_to)[column_to->size() - 1];
     ASSERT_EQ(field, "str");
 }
 
 TEST(ColumnDynamic, InsertFromOverflow2)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
 
@@ -221,9 +239,32 @@ TEST(ColumnDynamic, InsertFromOverflow2)
 
     column_to->insertFrom(*column_from, 1);
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 1);
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, "42.42");
+    ASSERT_EQ(field, 42.42);
+}
+
+TEST(ColumnDynamic, InsertFromOverflow3)
+{
+    auto column_from = ColumnDynamic::create(1);
+    column_from->insert(Field(42));
+    column_from->insert(Field(42.42));
+
+    auto column_to = ColumnDynamic::create(254);
+    column_to->insert(Field(41));
+
+    column_to->insertFrom(*column_from, 0);
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 0);
+    auto field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, 42);
+
+    column_to->insertFrom(*column_from, 1);
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 1);
+    field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, 42.42);
 }
 
 void checkInsertManyFrom(const ColumnDynamic::MutablePtr & column_from, ColumnDynamic::MutablePtr & column_to, const std::string & expected_variant, const std::vector<String> & expected_names, const std::unordered_map<String, UInt8> & expected_variant_name_to_discriminator)
@@ -256,42 +297,43 @@ void checkInsertManyFrom(const ColumnDynamic::MutablePtr & column_from, ColumnDy
 
 TEST(ColumnDynamic, InsertManyFrom1)
 {
-    auto column_to = ColumnDynamic::create(255);
-    checkInsertManyFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, String)", {"Float64", "Int8", "String"}, {{"Float64", 0}, {"Int8", 1}, {"String", 2}});
+    auto column_to = ColumnDynamic::create(254);
+    checkInsertManyFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, SharedVariant, String)", {"Float64", "Int8", "SharedVariant", "String"}, {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}});
 }
 
 TEST(ColumnDynamic, InsertManyFrom2)
 {
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     column_to->insert(Field(42));
     column_to->insert(Field(42.42));
     column_to->insert(Field("str"));
 
-    checkInsertManyFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, String)", {"Float64", "Int8", "String"}, {{"Float64", 0}, {"Int8", 1}, {"String", 2}});
+    checkInsertManyFrom(getInsertFromColumn(), column_to, "Variant(Float64, Int8, SharedVariant, String)", {"Float64", "Int8", "SharedVariant", "String"}, {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}});
 }
 
 TEST(ColumnDynamic, InsertManyFrom3)
 {
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     column_to->insert(Field(42));
     column_to->insert(Field(42.42));
     column_to->insert(Field("str"));
     column_to->insert(Array({Field(42)}));
 
-    checkInsertManyFrom(getInsertFromColumn(), column_to, "Variant(Array(Int8), Float64, Int8, String)", {"Array(Int8)", "Float64", "Int8", "String"}, {{"Array(Int8)", 0}, {"Float64", 1}, {"Int8", 2}, {"String", 3}});
+    checkInsertManyFrom(getInsertFromColumn(), column_to, "Variant(Array(Int8), Float64, Int8, SharedVariant, String)", {"Array(Int8)", "Float64", "Int8", "SharedVariant", "String"}, {{"Array(Int8)", 0}, {"Float64", 1}, {"Int8", 2}, {"SharedVariant", 3}, {"String", 4}});
 }
 
 TEST(ColumnDynamic, InsertManyFromOverflow1)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
     column_from->insert(Field("str"));
 
     auto column_to = getDynamicWithManyVariants(253);
     column_to->insertManyFrom(*column_from, 0, 2);
-    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 254);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 0);
     auto field = (*column_to)[column_to->size() - 2];
     ASSERT_EQ(field, 42);
     field = (*column_to)[column_to->size() - 1];
@@ -300,15 +342,17 @@ TEST(ColumnDynamic, InsertManyFromOverflow1)
     column_to->insertManyFrom(*column_from, 1, 2);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
     field = (*column_to)[column_to->size() - 2];
-    ASSERT_EQ(field, "42.42");
+    ASSERT_EQ(field, 42.42);
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, "42.42");
+    ASSERT_EQ(field, 42.42);
 
     column_to->insertManyFrom(*column_from, 2, 2);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 4);
     field = (*column_to)[column_to->size() - 1];
     ASSERT_EQ(field, "str");
     field = (*column_to)[column_to->size() - 2];
@@ -317,14 +361,15 @@ TEST(ColumnDynamic, InsertManyFromOverflow1)
 
 TEST(ColumnDynamic, InsertManyFromOverflow2)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
 
     auto column_to = getDynamicWithManyVariants(253);
     column_to->insertManyFrom(*column_from, 0, 2);
-    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 254);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 0);
     auto field = (*column_to)[column_to->size() - 2];
     ASSERT_EQ(field, 42);
     field = (*column_to)[column_to->size() - 1];
@@ -333,11 +378,39 @@ TEST(ColumnDynamic, InsertManyFromOverflow2)
     column_to->insertManyFrom(*column_from, 1, 2);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
     field = (*column_to)[column_to->size() - 2];
-    ASSERT_EQ(field, "42.42");
+    ASSERT_EQ(field, 42.42);
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, "42.42");
+    ASSERT_EQ(field, 42.42);
+}
+
+
+TEST(ColumnDynamic, InsertManyFromOverflow3)
+{
+    auto column_from = ColumnDynamic::create(1);
+    column_from->insert(Field(42));
+    column_from->insert(Field(42.42));
+
+    auto column_to = ColumnDynamic::create(254);
+    column_to->insert(Field(41));
+
+    column_to->insertManyFrom(*column_from, 0, 2);
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 0);
+    auto field = (*column_to)[column_to->size() - 2];
+    ASSERT_EQ(field, 42);
+    field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, 42);
+
+    column_to->insertManyFrom(*column_from, 1, 2);
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
+    field = (*column_to)[column_to->size() - 2];
+    ASSERT_EQ(field, 42.42);
+    field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, 42.42);
 }
 
 void checkInsertRangeFrom(const ColumnDynamic::MutablePtr & column_from, ColumnDynamic::MutablePtr & column_to, const std::string & expected_variant, const std::vector<String> & expected_names, const std::unordered_map<String, UInt8> & expected_variant_name_to_discriminator)
@@ -368,34 +441,34 @@ void checkInsertRangeFrom(const ColumnDynamic::MutablePtr & column_from, ColumnD
 
 TEST(ColumnDynamic, InsertRangeFrom1)
 {
-    auto column_to = ColumnDynamic::create(255);
-    checkInsertRangeFrom(getInsertFromColumn(2), column_to, "Variant(Float64, Int8, String)", {"Float64", "Int8", "String"}, {{"Float64", 0}, {"Int8", 1}, {"String", 2}});
+    auto column_to = ColumnDynamic::create(254);
+    checkInsertRangeFrom(getInsertFromColumn(2), column_to, "Variant(Float64, Int8, SharedVariant, String)", {"Float64", "Int8", "SharedVariant", "String"}, {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}});
 }
 
 TEST(ColumnDynamic, InsertRangeFrom2)
 {
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     column_to->insert(Field(42));
     column_to->insert(Field(42.42));
     column_to->insert(Field("str1"));
 
-    checkInsertRangeFrom(getInsertFromColumn(2), column_to, "Variant(Float64, Int8, String)", {"Float64", "Int8", "String"}, {{"Float64", 0}, {"Int8", 1}, {"String", 2}});
+    checkInsertRangeFrom(getInsertFromColumn(2), column_to, "Variant(Float64, Int8, SharedVariant, String)", {"Float64", "Int8", "SharedVariant", "String"}, {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}});
 }
 
 TEST(ColumnDynamic, InsertRangeFrom3)
 {
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     column_to->insert(Field(42));
     column_to->insert(Field(42.42));
     column_to->insert(Field("str1"));
     column_to->insert(Array({Field(42)}));
 
-    checkInsertRangeFrom(getInsertFromColumn(2), column_to, "Variant(Array(Int8), Float64, Int8, String)", {"Array(Int8)", "Float64", "Int8", "String"}, {{"Array(Int8)", 0}, {"Float64", 1}, {"Int8", 2}, {"String", 3}});
+    checkInsertRangeFrom(getInsertFromColumn(2), column_to, "Variant(Array(Int8), Float64, Int8, SharedVariant, String)", {"Array(Int8)", "Float64", "Int8", "SharedVariant", "String"}, {{"Array(Int8)", 0}, {"Float64", 1}, {"Int8", 2}, {"SharedVariant", 3}, {"String", 4}});
 }
 
 TEST(ColumnDynamic, InsertRangeFromOverflow1)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(43));
     column_from->insert(Field(42.42));
@@ -403,23 +476,25 @@ TEST(ColumnDynamic, InsertRangeFromOverflow1)
 
     auto column_to = getDynamicWithManyVariants(253);
     column_to->insertRangeFrom(*column_from, 0, 4);
+    ASSERT_EQ(column_to->size(), 257);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
     auto field = (*column_to)[column_to->size() - 4];
     ASSERT_EQ(field, Field(42));
     field = (*column_to)[column_to->size() - 3];
     ASSERT_EQ(field, Field(43));
     field = (*column_to)[column_to->size() - 2];
-    ASSERT_EQ(field, Field("42.42"));
+    ASSERT_EQ(field, Field(42.42));
     field = (*column_to)[column_to->size() - 1];
     ASSERT_EQ(field, Field("str"));
 }
 
 TEST(ColumnDynamic, InsertRangeFromOverflow2)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(43));
     column_from->insert(Field(42.42));
@@ -428,19 +503,20 @@ TEST(ColumnDynamic, InsertRangeFromOverflow2)
     column_to->insertRangeFrom(*column_from, 0, 3);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 1);
     auto field = (*column_to)[column_to->size() - 3];
     ASSERT_EQ(field, Field(42));
     field = (*column_to)[column_to->size() - 2];
     ASSERT_EQ(field, Field(43));
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, Field("42.42"));
+    ASSERT_EQ(field, Field(42.42));
 }
 
 TEST(ColumnDynamic, InsertRangeFromOverflow3)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(43));
     column_from->insert(Field(42.42));
@@ -449,20 +525,21 @@ TEST(ColumnDynamic, InsertRangeFromOverflow3)
     column_to->insert(Field("Str"));
     column_to->insertRangeFrom(*column_from, 0, 3);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 3);
     auto field = (*column_to)[column_to->size() - 3];
     ASSERT_EQ(field, Field(42));
     field = (*column_to)[column_to->size() - 2];
     ASSERT_EQ(field, Field(43));
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, Field("42.42"));
+    ASSERT_EQ(field, Field(42.42));
 }
 
 TEST(ColumnDynamic, InsertRangeFromOverflow4)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
     column_from->insert(Field("str"));
@@ -471,19 +548,20 @@ TEST(ColumnDynamic, InsertRangeFromOverflow4)
     column_to->insertRangeFrom(*column_from, 0, 3);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 3);
     auto field = (*column_to)[column_to->size() - 3];
-    ASSERT_EQ(field, Field("42"));
+    ASSERT_EQ(field, Field(42));
     field = (*column_to)[column_to->size() - 2];
-    ASSERT_EQ(field, Field("42.42"));
+    ASSERT_EQ(field, Field(42.42));
     field = (*column_to)[column_to->size() - 1];
     ASSERT_EQ(field, Field("str"));
 }
 
 TEST(ColumnDynamic, InsertRangeFromOverflow5)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(43));
     column_from->insert(Field(42.42));
@@ -493,22 +571,23 @@ TEST(ColumnDynamic, InsertRangeFromOverflow5)
     column_to->insert(Field("str"));
     column_to->insertRangeFrom(*column_from, 0, 4);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 3);
     auto field = (*column_to)[column_to->size() - 4];
     ASSERT_EQ(field, Field(42));
     field = (*column_to)[column_to->size() - 3];
     ASSERT_EQ(field, Field(43));
     field = (*column_to)[column_to->size() - 2];
-    ASSERT_EQ(field, Field("42.42"));
+    ASSERT_EQ(field, Field(42.42));
     field = (*column_to)[column_to->size() - 1];
     ASSERT_EQ(field, Field("str"));
 }
 
 TEST(ColumnDynamic, InsertRangeFromOverflow6)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(43));
     column_from->insert(Field(44));
@@ -520,13 +599,14 @@ TEST(ColumnDynamic, InsertRangeFromOverflow6)
     auto column_to = getDynamicWithManyVariants(253);
     column_to->insertRangeFrom(*column_from, 2, 5);
     ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 255);
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
-    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Array(Int8)"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 4);
     auto field = (*column_to)[column_to->size() - 5];
 
-    ASSERT_EQ(field, Field("44"));
+    ASSERT_EQ(field, Field(44));
     field = (*column_to)[column_to->size() - 4];
     ASSERT_EQ(field, Field(42.42));
     field = (*column_to)[column_to->size() - 3];
@@ -534,12 +614,136 @@ TEST(ColumnDynamic, InsertRangeFromOverflow6)
     field = (*column_to)[column_to->size() - 2];
     ASSERT_EQ(field, Field("str"));
     field = (*column_to)[column_to->size() - 1];
-    ASSERT_EQ(field, Field("[42]"));
+    ASSERT_EQ(field, Field(Array({Field(42)})));
+}
+
+TEST(ColumnDynamic, InsertRangeFromOverflow7)
+{
+    auto column_from = ColumnDynamic::create(2);
+    column_from->insert(Field(42.42));
+    column_from->insert(Field("str1"));
+    column_from->insert(Field(42));
+    column_from->insert(Field(43.43));
+    column_from->insert(Field(Array({Field(41)})));
+    column_from->insert(Field(43));
+    column_from->insert(Field("str2"));
+    column_from->insert(Field(Array({Field(42)})));
+
+    auto column_to = ColumnDynamic::create(254);
+    column_to->insert(Field(42));
+
+    column_to->insertRangeFrom(*column_from, 0, 8);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 4);
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Array(Int8)"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
+    auto field = (*column_to)[column_to->size() - 8];
+    ASSERT_EQ(field, Field(42.42));
+    field = (*column_to)[column_to->size() - 7];
+    ASSERT_EQ(field, Field("str1"));
+    field = (*column_to)[column_to->size() - 6];
+    ASSERT_EQ(field, Field(42));
+    field = (*column_to)[column_to->size() - 5];
+    ASSERT_EQ(field, Field(43.43));
+    field = (*column_to)[column_to->size() - 4];
+    ASSERT_EQ(field, Field(Array({Field(41)})));
+    field = (*column_to)[column_to->size() - 3];
+    ASSERT_EQ(field, Field(43));
+    field = (*column_to)[column_to->size() - 2];
+    ASSERT_EQ(field, Field("str2"));
+    field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, Field(Array({Field(42)})));
+}
+
+TEST(ColumnDynamic, InsertRangeFromOverflow8)
+{
+    auto column_from = ColumnDynamic::create(2);
+    column_from->insert(Field(42.42));
+    column_from->insert(Field("str1"));
+    column_from->insert(Field(42));
+    column_from->insert(Field(43.43));
+    column_from->insert(Field(Array({Field(41)})));
+    column_from->insert(Field(43));
+    column_from->insert(Field("str2"));
+    column_from->insert(Field(Array({Field(42)})));
+
+    auto column_to = ColumnDynamic::create(2);
+    column_to->insert(Field(42));
+    column_from->insert(Field("str1"));
+
+    column_to->insertRangeFrom(*column_from, 0, 8);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 3);
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Array(Int8)"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 4);
+    auto field = (*column_to)[column_to->size() - 8];
+    ASSERT_EQ(field, Field(42.42));
+    field = (*column_to)[column_to->size() - 7];
+    ASSERT_EQ(field, Field("str1"));
+    field = (*column_to)[column_to->size() - 6];
+    ASSERT_EQ(field, Field(42));
+    field = (*column_to)[column_to->size() - 5];
+    ASSERT_EQ(field, Field(43.43));
+    field = (*column_to)[column_to->size() - 4];
+    ASSERT_EQ(field, Field(Array({Field(41)})));
+    field = (*column_to)[column_to->size() - 3];
+    ASSERT_EQ(field, Field(43));
+    field = (*column_to)[column_to->size() - 2];
+    ASSERT_EQ(field, Field("str2"));
+    field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, Field(Array({Field(42)})));
+}
+
+TEST(ColumnDynamic, InsertRangeFromOverflow9)
+{
+    auto column_from = ColumnDynamic::create(3);
+    column_from->insert(Field("str1"));
+    column_from->insert(Field(42.42));
+    column_from->insert(Field("str2"));
+    column_from->insert(Field(42));
+    column_from->insert(Field(43.43));
+    column_from->insert(Field(Array({Field(41)})));
+    column_from->insert(Field(43));
+    column_from->insert(Field("str2"));
+    column_from->insert(Field(Array({Field(42)})));
+
+    auto column_to = ColumnDynamic::create(2);
+    column_to->insert(Field(42));
+
+    column_to->insertRangeFrom(*column_from, 0, 9);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names.size(), 3);
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Array(Int8)"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 4);
+    auto field = (*column_to)[column_to->size() - 9];
+    ASSERT_EQ(field, Field("str1"));
+    field = (*column_to)[column_to->size() - 8];
+    ASSERT_EQ(field, Field(42.42));
+    field = (*column_to)[column_to->size() - 7];
+    ASSERT_EQ(field, Field("str2"));
+    field = (*column_to)[column_to->size() - 6];
+    ASSERT_EQ(field, Field(42));
+    field = (*column_to)[column_to->size() - 5];
+    ASSERT_EQ(field, Field(43.43));
+    field = (*column_to)[column_to->size() - 4];
+    ASSERT_EQ(field, Field(Array({Field(41)})));
+    field = (*column_to)[column_to->size() - 3];
+    ASSERT_EQ(field, Field(43));
+    field = (*column_to)[column_to->size() - 2];
+    ASSERT_EQ(field, Field("str2"));
+    field = (*column_to)[column_to->size() - 1];
+    ASSERT_EQ(field, Field(Array({Field(42)})));
 }
 
 TEST(ColumnDynamic, SerializeDeserializeFromArena1)
 {
-    auto column = ColumnDynamic::create(255);
+    auto column = ColumnDynamic::create(254);
     column->insert(Field(42));
     column->insert(Field(42.42));
     column->insert(Field("str"));
@@ -564,7 +768,7 @@ TEST(ColumnDynamic, SerializeDeserializeFromArena1)
 
 TEST(ColumnDynamic, SerializeDeserializeFromArena2)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
     column_from->insert(Field("str"));
@@ -577,26 +781,26 @@ TEST(ColumnDynamic, SerializeDeserializeFromArena2)
     column_from->serializeValueIntoArena(2, arena, pos);
     column_from->serializeValueIntoArena(3, arena, pos);
 
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     pos = column_to->deserializeAndInsertFromArena(ref1.data);
     pos = column_to->deserializeAndInsertFromArena(pos);
     pos = column_to->deserializeAndInsertFromArena(pos);
     column_to->deserializeAndInsertFromArena(pos);
 
-    ASSERT_EQ((*column_from)[column_from->size() - 4], 42);
-    ASSERT_EQ((*column_from)[column_from->size() - 3], 42.42);
-    ASSERT_EQ((*column_from)[column_from->size() - 2], "str");
-    ASSERT_EQ((*column_from)[column_from->size() - 1], Null());
-    ASSERT_EQ(column_to->getVariantInfo().variant_type->getName(), "Variant(Float64, Int8, String)");
-    std::vector<String> expected_names = {"Float64", "Int8", "String"};
+    ASSERT_EQ((*column_to)[column_to->size() - 4], 42);
+    ASSERT_EQ((*column_to)[column_to->size() - 3], 42.42);
+    ASSERT_EQ((*column_to)[column_to->size() - 2], "str");
+    ASSERT_EQ((*column_to)[column_to->size() - 1], Null());
+    ASSERT_EQ(column_to->getVariantInfo().variant_type->getName(), "Variant(Float64, Int8, SharedVariant, String)");
+    std::vector<String> expected_names = {"Float64", "Int8", "SharedVariant", "String"};
     ASSERT_EQ(column_to->getVariantInfo().variant_names, expected_names);
-    std::unordered_map<String, UInt8> expected_variant_name_to_discriminator = {{"Float64", 0}, {"Int8", 1}, {"String", 2}};
+    std::unordered_map<String, UInt8> expected_variant_name_to_discriminator = {{"Float64", 0}, {"Int8", 1}, {"SharedVariant", 2}, {"String", 3}};
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator == expected_variant_name_to_discriminator);
 }
 
-TEST(ColumnDynamic, SerializeDeserializeFromArenaOverflow)
+TEST(ColumnDynamic, SerializeDeserializeFromArenaOverflow1)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(254);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
     column_from->insert(Field("str"));
@@ -615,18 +819,56 @@ TEST(ColumnDynamic, SerializeDeserializeFromArenaOverflow)
     pos = column_to->deserializeAndInsertFromArena(pos);
     column_to->deserializeAndInsertFromArena(pos);
 
-    ASSERT_EQ((*column_from)[column_from->size() - 4], 42);
-    ASSERT_EQ((*column_from)[column_from->size() - 3], 42.42);
-    ASSERT_EQ((*column_from)[column_from->size() - 2], "str");
-    ASSERT_EQ((*column_from)[column_from->size() - 1], Null());
+    ASSERT_EQ((*column_to)[column_to->size() - 4], 42);
+    ASSERT_EQ((*column_to)[column_to->size() - 3], 42.42);
+    ASSERT_EQ((*column_to)[column_to->size() - 2], "str");
+    ASSERT_EQ((*column_to)[column_to->size() - 1], Null());
     ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
     ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
+}
+
+TEST(ColumnDynamic, SerializeDeserializeFromArenaOverflow2)
+{
+    auto column_from = ColumnDynamic::create(2);
+    column_from->insert(Field(42));
+    column_from->insert(Field(42.42));
+    column_from->insert(Field("str"));
+    column_from->insert(Field(Null()));
+    column_from->insert(Field(Array({Field(42)})));
+
+    Arena arena;
+    const char * pos = nullptr;
+    auto ref1 = column_from->serializeValueIntoArena(0, arena, pos);
+    column_from->serializeValueIntoArena(1, arena, pos);
+    column_from->serializeValueIntoArena(2, arena, pos);
+    column_from->serializeValueIntoArena(3, arena, pos);
+    column_from->serializeValueIntoArena(4, arena, pos);
+
+    auto column_to = ColumnDynamic::create(2);
+    column_to->insert(Field(42.42));
+    pos = column_to->deserializeAndInsertFromArena(ref1.data);
+    pos = column_to->deserializeAndInsertFromArena(pos);
+    pos = column_to->deserializeAndInsertFromArena(pos);
+    pos = column_to->deserializeAndInsertFromArena(pos);
+    column_to->deserializeAndInsertFromArena(pos);
+
+    ASSERT_EQ((*column_to)[column_to->size() - 5], 42);
+    ASSERT_EQ((*column_to)[column_to->size() - 4], 42.42);
+    ASSERT_EQ((*column_to)[column_to->size() - 3], "str");
+    ASSERT_EQ((*column_to)[column_to->size() - 2], Null());
+    ASSERT_EQ((*column_to)[column_to->size() - 1], Field(Array({Field(42)})));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Int8"));
+    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Float64"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("String"));
+    ASSERT_FALSE(column_to->getVariantInfo().variant_name_to_discriminator.contains("Array(Int8)"));
+    ASSERT_EQ(column_to->getSharedVariant().size(), 2);
 }
 
 TEST(ColumnDynamic, skipSerializedInArena)
 {
-    auto column_from = ColumnDynamic::create(255);
+    auto column_from = ColumnDynamic::create(3);
     column_from->insert(Field(42));
     column_from->insert(Field(42.42));
     column_from->insert(Field("str"));
@@ -640,13 +882,41 @@ TEST(ColumnDynamic, skipSerializedInArena)
     auto ref4 = column_from->serializeValueIntoArena(3, arena, pos);
 
     const char * end = ref4.data + ref4.size;
-    auto column_to = ColumnDynamic::create(255);
+    auto column_to = ColumnDynamic::create(254);
     pos = column_to->skipSerializedInArena(ref1.data);
     pos = column_to->skipSerializedInArena(pos);
     pos = column_to->skipSerializedInArena(pos);
     pos = column_to->skipSerializedInArena(pos);
 
     ASSERT_EQ(pos, end);
-    ASSERT_TRUE(column_to->getVariantInfo().variant_name_to_discriminator.empty());
-    ASSERT_TRUE(column_to->getVariantInfo().variant_names.empty());
+    ASSERT_EQ(column_to->getVariantInfo().variant_name_to_discriminator.at("SharedVariant"), 0);
+    ASSERT_EQ(column_to->getVariantInfo().variant_names, Names{"SharedVariant"});
+}
+
+TEST(ColumnDynamic, compare)
+{
+    auto column_from = ColumnDynamic::create(3);
+    column_from->insert(Field(42));
+    column_from->insert(Field(42.42));
+    column_from->insert(Field("str"));
+    column_from->insert(Field(Null()));
+    column_from->insert(Field(Array({Field(42)})));
+
+    ASSERT_EQ(column_from->compareAt(0, 0, *column_from, -1), 0);
+    ASSERT_EQ(column_from->compareAt(0, 1, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(1, 1, *column_from, -1), 0);
+    ASSERT_EQ(column_from->compareAt(0, 2, *column_from, -1), -1);
+    ASSERT_EQ(column_from->compareAt(2, 0, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(2, 4, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(4, 2, *column_from, -1), -1);
+    ASSERT_EQ(column_from->compareAt(4, 4, *column_from, -1), 0);
+    ASSERT_EQ(column_from->compareAt(0, 3, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(1, 3, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(2, 3, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(3, 3, *column_from, -1), 0);
+    ASSERT_EQ(column_from->compareAt(4, 3, *column_from, -1), 1);
+    ASSERT_EQ(column_from->compareAt(3, 0, *column_from, -1), -1);
+    ASSERT_EQ(column_from->compareAt(3, 1, *column_from, -1), -1);
+    ASSERT_EQ(column_from->compareAt(3, 2, *column_from, -1), -1);
+    ASSERT_EQ(column_from->compareAt(3, 4, *column_from, -1), -1);
 }

src/Common/AsynchronousMetrics.cpp

@@ -1,18 +1,24 @@
-#include <Common/formatReadable.h>
 #include <Common/AsynchronousMetrics.h>
-#include <Common/Exception.h>
-#include <Common/setThreadName.h>
-#include <Common/CurrentMetrics.h>
-#include <Common/filesystemHelpers.h>
-#include <Common/logger_useful.h>
-#include <IO/UncompressedCache.h>
+
 #include <IO/MMappedFileCache.h>
 #include <IO/ReadHelpers.h>
+#include <IO/UncompressedCache.h>
+#include <base/cgroupsv2.h>
 #include <base/errnoToString.h>
 #include <base/find_symbols.h>
 #include <base/getPageSize.h>
 #include <sys/resource.h>
+#include <Common/CurrentMetrics.h>
+#include <Common/Exception.h>
+#include <Common/filesystemHelpers.h>
+#include <Common/formatReadable.h>
+#include <Common/logger_useful.h>
+#include <Common/setThreadName.h>
+
+#include <boost/locale/date_time_facet.hpp>
+
 #include <chrono>
+#include <string_view>
 
 #include "config.h"
 
@@ -52,6 +58,12 @@ static std::unique_ptr<ReadBufferFromFilePRead> openFileIfExists(const std::stri
     return {};
 }
 
+static void openCgroupv2MetricFile(const std::string & filename, std::optional<ReadBufferFromFilePRead> & out)
+{
+    if (auto path = getCgroupsV2PathContainingFile(filename))
+        openFileIfExists((path.value() + filename).c_str(), out);
+};
+
 #endif
 
 
@@ -63,21 +75,15 @@ AsynchronousMetrics::AsynchronousMetrics(
     , protocol_server_metrics_func(protocol_server_metrics_func_)
 {
 #if defined(OS_LINUX)
-    openFileIfExists("/proc/meminfo", meminfo);
-    openFileIfExists("/proc/loadavg", loadavg);
-    openFileIfExists("/proc/stat", proc_stat);
     openFileIfExists("/proc/cpuinfo", cpuinfo);
     openFileIfExists("/proc/sys/fs/file-nr", file_nr);
-    openFileIfExists("/proc/uptime", uptime);
     openFileIfExists("/proc/net/dev", net_dev);
 
     /// CGroups v2
-    openFileIfExists("/sys/fs/cgroup/memory.max", cgroupmem_limit_in_bytes);
-    if (cgroupmem_limit_in_bytes)
-    {
-        openFileIfExists("/sys/fs/cgroup/memory.current", cgroupmem_usage_in_bytes);
-    }
-    openFileIfExists("/sys/fs/cgroup/cpu.max", cgroupcpu_max);
+    openCgroupv2MetricFile("memory.max", cgroupmem_limit_in_bytes);
+    openCgroupv2MetricFile("memory.current", cgroupmem_usage_in_bytes);
+    openCgroupv2MetricFile("cpu.max", cgroupcpu_max);
+    openCgroupv2MetricFile("cpu.stat", cgroupcpu_stat);
 
     /// CGroups v1
     if (!cgroupmem_limit_in_bytes)
@@ -90,6 +96,21 @@ AsynchronousMetrics::AsynchronousMetrics(
         openFileIfExists("/sys/fs/cgroup/cpu/cpu.cfs_period_us", cgroupcpu_cfs_period);
         openFileIfExists("/sys/fs/cgroup/cpu/cpu.cfs_quota_us", cgroupcpu_cfs_quota);
     }
+    if (!cgroupcpu_stat)
+        openFileIfExists("/sys/fs/cgroup/cpuacct/cpuacct.stat", cgroupcpuacct_stat);
+
+    if (!cgroupcpu_stat && !cgroupcpuacct_stat)
+    {
+        /// The following metrics are not cgroup-aware and we've found cgroup-specific metric files for the similar metrics,
+        /// so we're better not reporting them at all to avoid confusion
+        openFileIfExists("/proc/loadavg", loadavg);
+        openFileIfExists("/proc/stat", proc_stat);
+        openFileIfExists("/proc/uptime", uptime);
+    }
+
+    /// The same story for memory metrics
+    if (!cgroupmem_limit_in_bytes)
+        openFileIfExists("/proc/meminfo", meminfo);
 
     openFileIfExists("/proc/sys/vm/max_map_count", vm_max_map_count);
     openFileIfExists("/proc/self/maps", vm_maps);
@@ -570,6 +591,151 @@ AsynchronousMetrics::NetworkInterfaceStatValues::operator-(const AsynchronousMet
 #endif
 
 
+#if defined(OS_LINUX)
+void AsynchronousMetrics::applyCPUMetricsUpdate(
+    AsynchronousMetricValues & new_values, const std::string & cpu_suffix, const ProcStatValuesCPU & delta_values, double multiplier)
+{
+    new_values["OSUserTime" + cpu_suffix]
+        = {delta_values.user * multiplier,
+           "The ratio of time the CPU core was running userspace code. This is a system-wide metric, it includes all the processes on the "
+           "host machine, not just clickhouse-server."
+           " This includes also the time when the CPU was under-utilized due to the reasons internal to the CPU (memory loads, pipeline "
+           "stalls, branch mispredictions, running another SMT core)."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSNiceTime" + cpu_suffix]
+        = {delta_values.nice * multiplier,
+           "The ratio of time the CPU core was running userspace code with higher priority. This is a system-wide metric, it includes all "
+           "the processes on the host machine, not just clickhouse-server."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSSystemTime" + cpu_suffix]
+        = {delta_values.system * multiplier,
+           "The ratio of time the CPU core was running OS kernel (system) code. This is a system-wide metric, it includes all the "
+           "processes on the host machine, not just clickhouse-server."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSIdleTime" + cpu_suffix]
+        = {delta_values.idle * multiplier,
+           "The ratio of time the CPU core was idle (not even ready to run a process waiting for IO) from the OS kernel standpoint. This "
+           "is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
+           " This does not include the time when the CPU was under-utilized due to the reasons internal to the CPU (memory loads, pipeline "
+           "stalls, branch mispredictions, running another SMT core)."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSIOWaitTime" + cpu_suffix]
+        = {delta_values.iowait * multiplier,
+           "The ratio of time the CPU core was not running the code but when the OS kernel did not run any other process on this CPU as "
+           "the processes were waiting for IO. This is a system-wide metric, it includes all the processes on the host machine, not just "
+           "clickhouse-server."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSIrqTime" + cpu_suffix]
+        = {delta_values.irq * multiplier,
+           "The ratio of time spent for running hardware interrupt requests on the CPU. This is a system-wide metric, it includes all the "
+           "processes on the host machine, not just clickhouse-server."
+           " A high number of this metric may indicate hardware misconfiguration or a very high network load."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSSoftIrqTime" + cpu_suffix]
+        = {delta_values.softirq * multiplier,
+           "The ratio of time spent for running software interrupt requests on the CPU. This is a system-wide metric, it includes all the "
+           "processes on the host machine, not just clickhouse-server."
+           " A high number of this metric may indicate inefficient software running on the system."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSStealTime" + cpu_suffix]
+        = {delta_values.steal * multiplier,
+           "The ratio of time spent in other operating systems by the CPU when running in a virtualized environment. This is a system-wide "
+           "metric, it includes all the processes on the host machine, not just clickhouse-server."
+           " Not every virtualized environments present this metric, and most of them don't."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSGuestTime" + cpu_suffix]
+        = {delta_values.guest * multiplier,
+           "The ratio of time spent running a virtual CPU for guest operating systems under the control of the Linux kernel (See `man "
+           "procfs`). This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
+           " This metric is irrelevant for ClickHouse, but still exists for completeness."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+    new_values["OSGuestNiceTime" + cpu_suffix]
+        = {delta_values.guest_nice * multiplier,
+           "The ratio of time spent running a virtual CPU for guest operating systems under the control of the Linux kernel, when a guest "
+           "was set to a higher priority (See `man procfs`). This is a system-wide metric, it includes all the processes on the host "
+           "machine, not just clickhouse-server."
+           " This metric is irrelevant for ClickHouse, but still exists for completeness."
+           " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across "
+           "them [0..num cores]."};
+}
+
+void AsynchronousMetrics::applyNormalizedCPUMetricsUpdate(
+    AsynchronousMetricValues & new_values, double num_cpus_to_normalize, const ProcStatValuesCPU & delta_values_all_cpus, double multiplier)
+{
+    chassert(num_cpus_to_normalize);
+
+    new_values["OSUserTimeNormalized"]
+        = {delta_values_all_cpus.user * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSUserTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSNiceTimeNormalized"]
+        = {delta_values_all_cpus.nice * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSNiceTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSSystemTimeNormalized"]
+        = {delta_values_all_cpus.system * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSSystemTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSIdleTimeNormalized"]
+        = {delta_values_all_cpus.idle * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSIdleTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSIOWaitTimeNormalized"]
+        = {delta_values_all_cpus.iowait * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSIOWaitTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSIrqTimeNormalized"]
+        = {delta_values_all_cpus.irq * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSIrqTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of "
+           "the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSSoftIrqTimeNormalized"]
+        = {delta_values_all_cpus.softirq * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSSoftIrqTime` but divided to the number of CPU cores to be measured in the [0..1] interval "
+           "regardless of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSStealTimeNormalized"]
+        = {delta_values_all_cpus.steal * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSStealTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSGuestTimeNormalized"]
+        = {delta_values_all_cpus.guest * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSGuestTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless "
+           "of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+    new_values["OSGuestNiceTimeNormalized"]
+        = {delta_values_all_cpus.guest_nice * multiplier / num_cpus_to_normalize,
+           "The value is similar to `OSGuestNiceTime` but divided to the number of CPU cores to be measured in the [0..1] interval "
+           "regardless of the number of cores."
+           " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is "
+           "non-uniform, and still get the average resource utilization metric."};
+}
+#endif
+
 void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
 {
     Stopwatch watch;
@@ -831,7 +997,68 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
         new_values["CGroupMaxCPU"] = { max_cpu_cgroups, "The maximum number of CPU cores according to CGroups."};
     }
 
-    if (proc_stat)
+    if (cgroupcpu_stat || cgroupcpuacct_stat)
+    {
+        try
+        {
+            ReadBufferFromFilePRead & in = cgroupcpu_stat ? *cgroupcpu_stat : *cgroupcpuacct_stat;
+            ProcStatValuesCPU current_values{};
+
+            /// We re-read the file from the beginning each time
+            in.rewind();
+
+            while (!in.eof())
+            {
+                String name;
+                readStringUntilWhitespace(name, in);
+                skipWhitespaceIfAny(in);
+
+                /// `user_usec` for cgroup v2 and `user` for cgroup v1
+                if (name.starts_with("user"))
+                {
+                    readText(current_values.user, in);
+                    skipToNextLineOrEOF(in);
+                }
+                /// `system_usec` for cgroup v2 and `system` for cgroup v1
+                else if (name.starts_with("system"))
+                {
+                    readText(current_values.system, in);
+                    skipToNextLineOrEOF(in);
+                }
+                else
+                    skipToNextLineOrEOF(in);
+            }
+
+            if (!first_run)
+            {
+                auto get_clock_ticks = [&]()
+                {
+                    if (auto hz = sysconf(_SC_CLK_TCK); hz != -1)
+                        return hz;
+                    else
+                        throw ErrnoException(ErrorCodes::CANNOT_SYSCONF, "Cannot call 'sysconf' to obtain system HZ");
+                };
+                const auto cgroup_version_specific_divisor = cgroupcpu_stat ? 1e6 : get_clock_ticks();
+                const double multiplier = 1.0 / cgroup_version_specific_divisor
+                    / (std::chrono::duration_cast<std::chrono::nanoseconds>(time_since_previous_update).count() / 1e9);
+
+                const ProcStatValuesCPU delta_values = current_values - proc_stat_values_all_cpus;
+                applyCPUMetricsUpdate(new_values, /*cpu_suffix=*/"", delta_values, multiplier);
+                if (max_cpu_cgroups > 0)
+                    applyNormalizedCPUMetricsUpdate(new_values, max_cpu_cgroups, delta_values, multiplier);
+            }
+
+            proc_stat_values_all_cpus = current_values;
+        }
+        catch (...)
+        {
+            tryLogCurrentException(__PRETTY_FUNCTION__);
+            openCgroupv2MetricFile("cpu.stat", cgroupcpu_stat);
+            if (!cgroupcpu_stat)
+                openFileIfExists("/sys/fs/cgroup/cpuacct/cpuacct.stat", cgroupcpuacct_stat);
+        }
+    }
+    else if (proc_stat)
     {
         try
         {
@@ -886,43 +1113,7 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
                         else
                             delta_values_all_cpus = delta_values;
 
-                        new_values["OSUserTime" + cpu_suffix] = { delta_values.user * multiplier,
-                            "The ratio of time the CPU core was running userspace code. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " This includes also the time when the CPU was under-utilized due to the reasons internal to the CPU (memory loads, pipeline stalls, branch mispredictions, running another SMT core)."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSNiceTime" + cpu_suffix] = { delta_values.nice * multiplier,
-                            "The ratio of time the CPU core was running userspace code with higher priority. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSSystemTime" + cpu_suffix] = { delta_values.system * multiplier,
-                            "The ratio of time the CPU core was running OS kernel (system) code. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSIdleTime" + cpu_suffix] = { delta_values.idle * multiplier,
-                            "The ratio of time the CPU core was idle (not even ready to run a process waiting for IO) from the OS kernel standpoint. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " This does not include the time when the CPU was under-utilized due to the reasons internal to the CPU (memory loads, pipeline stalls, branch mispredictions, running another SMT core)."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSIOWaitTime" + cpu_suffix] = { delta_values.iowait * multiplier,
-                            "The ratio of time the CPU core was not running the code but when the OS kernel did not run any other process on this CPU as the processes were waiting for IO. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSIrqTime" + cpu_suffix] = { delta_values.irq * multiplier,
-                            "The ratio of time spent for running hardware interrupt requests on the CPU. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " A high number of this metric may indicate hardware misconfiguration or a very high network load."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSSoftIrqTime" + cpu_suffix] = { delta_values.softirq * multiplier,
-                            "The ratio of time spent for running software interrupt requests on the CPU. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " A high number of this metric may indicate inefficient software running on the system."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSStealTime" + cpu_suffix] = { delta_values.steal * multiplier,
-                            "The ratio of time spent in other operating systems by the CPU when running in a virtualized environment. This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " Not every virtualized environments present this metric, and most of them don't."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSGuestTime" + cpu_suffix] = { delta_values.guest * multiplier,
-                            "The ratio of time spent running a virtual CPU for guest operating systems under the control of the Linux kernel (See `man procfs`). This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " This metric is irrelevant for ClickHouse, but still exists for completeness."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
-                        new_values["OSGuestNiceTime" + cpu_suffix] = { delta_values.guest_nice * multiplier,
-                            "The ratio of time spent running a virtual CPU for guest operating systems under the control of the Linux kernel, when a guest was set to a higher priority (See `man procfs`). This is a system-wide metric, it includes all the processes on the host machine, not just clickhouse-server."
-                            " This metric is irrelevant for ClickHouse, but still exists for completeness."
-                            " The value for a single CPU core will be in the interval [0..1]. The value for all CPU cores is calculated as a sum across them [0..num cores]."};
+                        applyCPUMetricsUpdate(new_values, cpu_suffix, delta_values, multiplier);
                     }
 
                     prev_values = current_values;
@@ -978,38 +1169,7 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
                 Float64 num_cpus_to_normalize = max_cpu_cgroups > 0 ? max_cpu_cgroups : num_cpus;
 
                 if (num_cpus_to_normalize > 0)
-                {
-                    new_values["OSUserTimeNormalized"] = { delta_values_all_cpus.user * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSUserTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSNiceTimeNormalized"] = { delta_values_all_cpus.nice * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSNiceTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSSystemTimeNormalized"] = { delta_values_all_cpus.system * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSSystemTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSIdleTimeNormalized"] = { delta_values_all_cpus.idle * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSIdleTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSIOWaitTimeNormalized"] = { delta_values_all_cpus.iowait * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSIOWaitTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSIrqTimeNormalized"] = { delta_values_all_cpus.irq * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSIrqTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSSoftIrqTimeNormalized"] = { delta_values_all_cpus.softirq * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSSoftIrqTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSStealTimeNormalized"] = { delta_values_all_cpus.steal * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSStealTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSGuestTimeNormalized"] = { delta_values_all_cpus.guest * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSGuestTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                    new_values["OSGuestNiceTimeNormalized"] = { delta_values_all_cpus.guest_nice * multiplier / num_cpus_to_normalize,
-                        "The value is similar to `OSGuestNiceTime` but divided to the number of CPU cores to be measured in the [0..1] interval regardless of the number of cores."
-                        " This allows you to average the values of this metric across multiple servers in a cluster even if the number of cores is non-uniform, and still get the average resource utilization metric."};
-                }
+                    applyNormalizedCPUMetricsUpdate(new_values, num_cpus_to_normalize, delta_values_all_cpus, multiplier);
             }
 
             proc_stat_values_other = current_other_values;
@@ -1042,8 +1202,7 @@ void AsynchronousMetrics::update(TimePoint update_time, bool force_update)
             tryLogCurrentException(__PRETTY_FUNCTION__);
         }
     }
-
-    if (meminfo)
+    else if (meminfo)
     {
         try
         {

src/Common/AsynchronousMetrics.h

@@ -126,6 +126,8 @@ private:
     std::optional<ReadBufferFromFilePRead> cgroupcpu_cfs_period TSA_GUARDED_BY(data_mutex);
     std::optional<ReadBufferFromFilePRead> cgroupcpu_cfs_quota TSA_GUARDED_BY(data_mutex);
     std::optional<ReadBufferFromFilePRead> cgroupcpu_max TSA_GUARDED_BY(data_mutex);
+    std::optional<ReadBufferFromFilePRead> cgroupcpu_stat TSA_GUARDED_BY(data_mutex);
+    std::optional<ReadBufferFromFilePRead> cgroupcpuacct_stat TSA_GUARDED_BY(data_mutex);
 
     std::optional<ReadBufferFromFilePRead> vm_max_map_count TSA_GUARDED_BY(data_mutex);
     std::optional<ReadBufferFromFilePRead> vm_maps TSA_GUARDED_BY(data_mutex);
@@ -221,6 +223,16 @@ private:
     void openBlockDevices();
     void openSensorsChips();
     void openEDAC();
+
+    void applyCPUMetricsUpdate(
+        AsynchronousMetricValues & new_values, const std::string & cpu_suffix, const ProcStatValuesCPU & delta_values, double multiplier);
+
+    void applyNormalizedCPUMetricsUpdate(
+        AsynchronousMetricValues & new_values,
+        double num_cpus_to_normalize,
+        const ProcStatValuesCPU & delta_values_all_cpus,
+        double multiplier);
+
 #endif
 
     void run();

src/Common/CgroupsMemoryUsageObserver.cpp

@@ -144,31 +144,6 @@ private:
 /// - 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> getCgroupsV2Path()
-{
-    if (!cgroupsV2Enabled())
-        return {};
-
-    if (!cgroupsV2MemoryControllerEnabled())
-        return {};
-
-    fs::path current_cgroup = cgroupV2PathOfProcess();
-    if (current_cgroup.empty())
-        return {};
-
-    /// Return the bottom-most nested current memory file. If there is no such file at the current
-    /// level, try again at the parent level as memory settings are inherited.
-    while (current_cgroup != default_cgroups_mount.parent_path())
-    {
-        const auto current_path = current_cgroup / "memory.current";
-        const auto stat_path = current_cgroup / "memory.stat";
-        if (fs::exists(current_path) && fs::exists(stat_path))
-            return {current_cgroup};
-        current_cgroup = current_cgroup.parent_path();
-    }
-    return {};
-}
-
 std::optional<std::string> getCgroupsV1Path()
 {
     auto path = default_cgroups_mount / "memory/memory.stat";
@@ -179,7 +154,7 @@ std::optional<std::string> getCgroupsV1Path()
 
 std::pair<std::string, CgroupsMemoryUsageObserver::CgroupsVersion> getCgroupsPath()
 {
-    auto v2_path = getCgroupsV2Path();
+    auto v2_path = getCgroupsV2PathContainingFile("memory.current");
     if (v2_path.has_value())
         return {*v2_path, CgroupsMemoryUsageObserver::CgroupsVersion::V2};
 

src/Common/CurrentMetrics.cpp

@@ -307,7 +307,7 @@
     M(FilteringMarksWithPrimaryKey, "Number of threads currently doing filtering of mark ranges by the primary key") \
     M(FilteringMarksWithSecondaryKeys, "Number of threads currently doing filtering of mark ranges by secondary keys") \
     \
-    M(S3DiskNoKeyErrors, "The number of `NoSuchKey` errors that occur when reading data from S3 cloud storage through ClickHouse disks.") \
+    M(DiskS3NoSuchKeyErrors, "The number of `NoSuchKey` errors that occur when reading data from S3 cloud storage through ClickHouse disks.") \
 
 #ifdef APPLY_FOR_EXTERNAL_METRICS
     #define APPLY_FOR_METRICS(M) APPLY_FOR_BUILTIN_METRICS(M) APPLY_FOR_EXTERNAL_METRICS(M)

src/Common/ProfileEvents.cpp

@@ -209,8 +209,35 @@
     \
     M(Merge, "Number of launched background merges.") \
     M(MergedRows, "Rows read for background merges. This is the number of rows before merge.") \
+    M(MergedColumns, "Number of columns merged during the horizontal stage of merges.") \
+    M(GatheredColumns, "Number of columns gathered during the vertical stage of merges.") \
     M(MergedUncompressedBytes, "Uncompressed bytes (for columns as they stored in memory) that was read for background merges. This is the number before merge.") \
-    M(MergesTimeMilliseconds, "Total time spent for background merges.")\
+    M(MergeTotalMilliseconds, "Total time spent for background merges") \
+    M(MergeExecuteMilliseconds, "Total busy time spent for execution of background merges") \
+    M(MergeHorizontalStageTotalMilliseconds, "Total time spent for horizontal stage of background merges") \
+    M(MergeHorizontalStageExecuteMilliseconds, "Total busy time spent for execution of horizontal stage of background merges") \
+    M(MergeVerticalStageTotalMilliseconds, "Total time spent for vertical stage of background merges") \
+    M(MergeVerticalStageExecuteMilliseconds, "Total busy time spent for execution of vertical stage of background merges") \
+    M(MergeProjectionStageTotalMilliseconds, "Total time spent for projection stage of background merges") \
+    M(MergeProjectionStageExecuteMilliseconds, "Total busy time spent for execution of projection stage of background merges") \
+    \
+    M(MergingSortedMilliseconds, "Total time spent while merging sorted columns") \
+    M(AggregatingSortedMilliseconds, "Total time spent while aggregating sorted columns") \
+    M(CollapsingSortedMilliseconds, "Total time spent while collapsing sorted columns") \
+    M(ReplacingSortedMilliseconds, "Total time spent while replacing sorted columns") \
+    M(SummingSortedMilliseconds, "Total time spent while summing sorted columns") \
+    M(VersionedCollapsingSortedMilliseconds, "Total time spent while version collapsing sorted columns") \
+    M(GatheringColumnMilliseconds, "Total time spent while gathering columns for vertical merge") \
+    \
+    M(MutationTotalParts, "Number of total parts for which mutations tried to be applied") \
+    M(MutationUntouchedParts, "Number of total parts for which mutations tried to be applied but which was completely skipped according to predicate") \
+    M(MutatedRows, "Rows read for mutations. This is the number of rows before mutation") \
+    M(MutatedUncompressedBytes, "Uncompressed bytes (for columns as they stored in memory) that was read for mutations. This is the number before mutation.") \
+    M(MutationTotalMilliseconds, "Total time spent for mutations.") \
+    M(MutationExecuteMilliseconds, "Total busy time spent for execution of mutations.") \
+    M(MutationAllPartColumns, "Number of times when task to mutate all columns in part was created") \
+    M(MutationSomePartColumns, "Number of times when task to mutate some columns in part was created") \
+    M(MutateTaskProjectionsCalculationMicroseconds, "Time spent calculating projections in mutations.") \
     \
     M(MergeTreeDataWriterRows, "Number of rows INSERTed to MergeTree tables.") \
     M(MergeTreeDataWriterUncompressedBytes, "Uncompressed bytes (for columns as they stored in memory) INSERTed to MergeTree tables.") \
@@ -225,7 +252,6 @@
     M(MergeTreeDataWriterProjectionsCalculationMicroseconds, "Time spent calculating projections") \
     M(MergeTreeDataProjectionWriterSortingBlocksMicroseconds, "Time spent sorting blocks (for projection it might be a key different from table's sorting key)") \
     M(MergeTreeDataProjectionWriterMergingBlocksMicroseconds, "Time spent merging blocks") \
-    M(MutateTaskProjectionsCalculationMicroseconds, "Time spent calculating projections") \
     \
     M(InsertedWideParts, "Number of parts inserted in Wide format.") \
     M(InsertedCompactParts, "Number of parts inserted in Compact format.") \

src/Common/config.h.in

@@ -58,6 +58,7 @@
 #cmakedefine01 USE_FILELOG
 #cmakedefine01 USE_ODBC
 #cmakedefine01 USE_BLAKE3
+#cmakedefine01 USE_USEARCH
 #cmakedefine01 USE_SKIM
 #cmakedefine01 USE_PRQL
 #cmakedefine01 USE_ULID

src/Core/ProtocolDefines.h

@@ -83,6 +83,9 @@ static constexpr auto DBMS_MIN_REVISION_WITH_SYSTEM_KEYWORDS_TABLE = 54468;
 
 static constexpr auto DBMS_MIN_REVISION_WITH_ROWS_BEFORE_AGGREGATION = 54469;
 
+/// Packets size header
+static constexpr auto DBMS_MIN_PROTOCOL_VERSION_WITH_CHUNKED_PACKETS = 54470;
+
 /// Version of ClickHouse TCP protocol.
 ///
 /// Should be incremented manually on protocol changes.
@@ -90,6 +93,6 @@ static constexpr auto DBMS_MIN_REVISION_WITH_ROWS_BEFORE_AGGREGATION = 54469;
 /// NOTE: DBMS_TCP_PROTOCOL_VERSION has nothing common with VERSION_REVISION,
 /// later is just a number for server version (one number instead of commit SHA)
 /// for simplicity (sometimes it may be more convenient in some use cases).
-static constexpr auto DBMS_TCP_PROTOCOL_VERSION = 54469;
+static constexpr auto DBMS_TCP_PROTOCOL_VERSION = 54470;
 
 }

src/Core/ServerSettings.h

@@ -134,6 +134,7 @@ namespace DB
     M(Bool, async_load_databases, false, "Enable asynchronous loading of databases and tables to speedup server startup. Queries to not yet loaded entity will be blocked until load is finished.", 0) \
     M(Bool, display_secrets_in_show_and_select, false, "Allow showing secrets in SHOW and SELECT queries via a format setting and a grant", 0) \
     M(Seconds, keep_alive_timeout, DEFAULT_HTTP_KEEP_ALIVE_TIMEOUT, "The number of seconds that ClickHouse waits for incoming requests before closing the connection.", 0) \
+    M(UInt64, max_keep_alive_requests, 10000, "The maximum number of requests handled via a single http keepalive connection before the server closes this connection.", 0) \
     M(Seconds, replicated_fetches_http_connection_timeout, 0, "HTTP connection timeout for part fetch requests. Inherited from default profile `http_connection_timeout` if not set explicitly.", 0) \
     M(Seconds, replicated_fetches_http_send_timeout, 0, "HTTP send timeout for part fetch requests. Inherited from default profile `http_send_timeout` if not set explicitly.", 0) \
     M(Seconds, replicated_fetches_http_receive_timeout, 0, "HTTP receive timeout for fetch part requests. Inherited from default profile `http_receive_timeout` if not set explicitly.", 0) \

src/Core/Settings.h

@@ -593,7 +593,6 @@ class IColumn;
     M(UInt64, mutations_sync, 0, "Wait for synchronous execution of ALTER TABLE UPDATE/DELETE queries (mutations). 0 - execute asynchronously. 1 - wait current server. 2 - wait all replicas if they exist.", 0) \
     M(Bool, enable_lightweight_delete, true, "Enable lightweight DELETE mutations for mergetree tables.", 0) ALIAS(allow_experimental_lightweight_delete) \
     M(UInt64, lightweight_deletes_sync, 2, "The same as 'mutation_sync', but controls only execution of lightweight deletes", 0) \
-    M(LightweightMutationProjectionMode, lightweight_mutation_projection_mode, LightweightMutationProjectionMode::THROW, "When lightweight delete happens on a table with projection(s), the possible operations include throw the exception as projection exists, or drop all projection related to this table then do lightweight delete.", 0) \
     M(Bool, apply_deleted_mask, true, "Enables filtering out rows deleted with lightweight DELETE. If disabled, a query will be able to read those rows. This is useful for debugging and \"undelete\" scenarios", 0) \
     M(Bool, optimize_normalize_count_variants, true, "Rewrite aggregate functions that semantically equals to count() as count().", 0) \
     M(Bool, optimize_injective_functions_inside_uniq, true, "Delete injective functions of one argument inside uniq*() functions.", 0) \
@@ -676,6 +675,7 @@ class IColumn;
     M(Bool, query_cache_squash_partial_results, true, "Squash partial result blocks to blocks of size 'max_block_size'. Reduces performance of inserts into the query cache but improves the compressability of cache entries.", 0) \
     M(Seconds, query_cache_ttl, 60, "After this time in seconds entries in the query cache become stale", 0) \
     M(Bool, query_cache_share_between_users, false, "Allow other users to read entry in the query cache", 0) \
+    M(String, query_cache_tag, "", "A string which acts as a label for query cache entries. The same queries with different tags are considered different by the query cache.", 0) \
     M(Bool, enable_sharing_sets_for_mutations, true, "Allow sharing set objects build for IN subqueries between different tasks of the same mutation. This reduces memory usage and CPU consumption", 0) \
     \
     M(Bool, optimize_rewrite_sum_if_to_count_if, true, "Rewrite sumIf() and sum(if()) function countIf() function when logically equivalent", 0) \
@@ -907,20 +907,18 @@ class IColumn;
     M(Bool, allow_experimental_hash_functions, false, "Enable experimental hash functions", 0) \
     M(Bool, allow_experimental_object_type, false, "Allow Object and JSON data types", 0) \
     M(Bool, allow_experimental_time_series_table, false, "Allows experimental TimeSeries table engine", 0) \
+    M(Bool, allow_experimental_vector_similarity_index, false, "Allow experimental vector similarity index", 0) \
     M(Bool, allow_experimental_variant_type, false, "Allow Variant data type", 0) \
     M(Bool, allow_experimental_dynamic_type, false, "Allow Dynamic data type", 0) \
-    M(Bool, allow_experimental_annoy_index, false, "Allows to use Annoy index. Disabled by default because this feature is experimental", 0) \
-    M(Bool, allow_experimental_usearch_index, false, "Allows to use USearch index. Disabled by default because this feature is experimental", 0) \
     M(Bool, allow_experimental_codecs, false, "If it is set to true, allow to specify experimental compression codecs (but we don't have those yet and this option does nothing).", 0) \
     M(UInt64, max_limit_for_ann_queries, 1'000'000, "SELECT queries with LIMIT bigger than this setting cannot use ANN indexes. Helps to prevent memory overflows in ANN search indexes.", 0) \
-    M(UInt64, max_threads_for_annoy_index_creation, 4, "Number of threads used to build Annoy indexes (0 means all cores, not recommended)", 0) \
-    M(Int64, annoy_index_search_k_nodes, -1, "SELECT queries search up to this many nodes in Annoy indexes.", 0) \
     M(Bool, throw_on_unsupported_query_inside_transaction, true, "Throw exception if unsupported query is used inside transaction", 0) \
     M(TransactionsWaitCSNMode, wait_changes_become_visible_after_commit_mode, TransactionsWaitCSNMode::WAIT_UNKNOWN, "Wait for committed changes to become actually visible in the latest snapshot", 0) \
     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(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)\
     \
     M(Bool, allow_statistics_optimize, false, "Allows using statistics to optimize queries", 0) ALIAS(allow_statistic_optimize) \
     M(Bool, allow_experimental_statistics, false, "Allows using statistics", 0) ALIAS(allow_experimental_statistic) \
@@ -1036,6 +1034,10 @@ class IColumn;
     MAKE_OBSOLETE(M, UInt64, parallel_replicas_min_number_of_granules_to_enable, 0) \
     MAKE_OBSOLETE(M, Bool, query_plan_optimize_projection, true) \
     MAKE_OBSOLETE(M, Bool, query_cache_store_results_of_queries_with_nondeterministic_functions, false) \
+    MAKE_OBSOLETE(M, Bool, allow_experimental_annoy_index, false) \
+    MAKE_OBSOLETE(M, UInt64, max_threads_for_annoy_index_creation, 4) \
+    MAKE_OBSOLETE(M, Int64, annoy_index_search_k_nodes, -1) \
+    MAKE_OBSOLETE(M, Bool, allow_experimental_usearch_index, false) \
     MAKE_OBSOLETE(M, Bool, optimize_move_functions_out_of_any, false) \
     MAKE_OBSOLETE(M, Bool, allow_experimental_undrop_table_query, true) \
     MAKE_OBSOLETE(M, Bool, allow_experimental_s3queue, true) \

src/Core/SettingsChangesHistory.cpp

@@ -80,11 +80,14 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
             {"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."},
             {"merge_tree_min_bytes_per_task_for_remote_reading", 4194304, 2097152, "Value is unified with `filesystem_prefetch_min_bytes_for_single_read_task`"},
+            {"use_hive_partitioning", false, false, "Allows to use hive partitioning for File, URL, S3, AzureBlobStorage and HDFS engines."},
             {"allow_experimental_kafka_offsets_storage_in_keeper", false, false, "Allow the usage of experimental Kafka storage engine that stores the committed offsets in ClickHouse Keeper"},
             {"allow_archive_path_syntax", true, true, "Added new setting to allow disabling archive path syntax."},
+            {"query_cache_tag", "", "", "New setting for labeling query cache settings."},
             {"allow_experimental_time_series_table", false, false, "Added new setting to allow the TimeSeries table engine"},
             {"enable_analyzer", 1, 1, "Added an alias to a setting `allow_experimental_analyzer`."},
             {"optimize_functions_to_subcolumns", false, true, "Enabled settings by default"},
+            {"allow_experimental_vector_similarity_index", false, false, "Added new setting to allow experimental vector similarity indexes"},
         }
     },
     {"24.7",
@@ -100,7 +103,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
             {"dictionary_validate_primary_key_type", false, false, "Validate primary key type for dictionaries. By default id type for simple layouts will be implicitly converted to UInt64."},
             {"collect_hash_table_stats_during_joins", false, true, "New setting."},
             {"max_size_to_preallocate_for_joins", 0, 100'000'000, "New setting."},
-            {"input_format_orc_reader_time_zone_name", "GMT", "GMT", "The time zone name for ORC row reader, the default ORC row reader's time zone is GMT."},            {"lightweight_mutation_projection_mode", "throw", "throw", "When lightweight delete happens on a table with projection(s), the possible operations include throw the exception as projection exists, or drop all projection related to this table then do lightweight delete."},
+            {"input_format_orc_reader_time_zone_name", "GMT", "GMT", "The time zone name for ORC row reader, the default ORC row reader's time zone is GMT."},
             {"database_replicated_allow_heavy_create", true, false, "Long-running DDL queries (CREATE AS SELECT and POPULATE) for Replicated database engine was forbidden"},
             {"query_plan_merge_filters", false, false, "Allow to merge filters in the query plan"},
             {"azure_sdk_max_retries", 10, 10, "Maximum number of retries in azure sdk"},

src/Core/SettingsEnums.cpp

@@ -175,7 +175,8 @@ IMPLEMENT_SETTING_ENUM(ParallelReplicasCustomKeyFilterType, ErrorCodes::BAD_ARGU
 
 IMPLEMENT_SETTING_ENUM(LightweightMutationProjectionMode, ErrorCodes::BAD_ARGUMENTS,
     {{"throw", LightweightMutationProjectionMode::THROW},
-     {"drop", LightweightMutationProjectionMode::DROP}})
+     {"drop", LightweightMutationProjectionMode::DROP},
+     {"rebuild", LightweightMutationProjectionMode::REBUILD}})
 
 IMPLEMENT_SETTING_ENUM(DeduplicateMergeProjectionMode, ErrorCodes::BAD_ARGUMENTS,
     {{"throw", DeduplicateMergeProjectionMode::THROW},

src/Core/SettingsEnums.h

@@ -311,6 +311,7 @@ enum class LightweightMutationProjectionMode : uint8_t
 {
     THROW,
     DROP,
+    REBUILD,
 };
 
 DECLARE_SETTING_ENUM(LightweightMutationProjectionMode)

src/Core/fuzzers/CMakeLists.txt

@@ -1,2 +1,2 @@
 clickhouse_add_executable (names_and_types_fuzzer names_and_types_fuzzer.cpp)
-target_link_libraries (names_and_types_fuzzer PRIVATE dbms clickhouse_functions)
+target_link_libraries (names_and_types_fuzzer PRIVATE clickhouse_functions)

src/DataTypes/DataTypeDynamic.cpp

@@ -7,6 +7,7 @@
 #include <DataTypes/NestedUtils.h>
 #include <DataTypes/DataTypeNullable.h>
 #include <DataTypes/DataTypesNumber.h>
+#include <DataTypes/DataTypesBinaryEncoding.h>
 #include <Columns/ColumnDynamic.h>
 #include <Columns/ColumnVariant.h>
 #include <Core/Field.h>
@@ -14,6 +15,7 @@
 #include <Parsers/ASTFunction.h>
 #include <Parsers/ASTIdentifier.h>
 #include <Parsers/ASTLiteral.h>
+#include <IO/ReadBufferFromMemory.h>
 
 namespace DB
 {
@@ -71,8 +73,8 @@ static DataTypePtr create(const ASTPtr & arguments)
 
     auto * literal = argument->arguments->children[1]->as<ASTLiteral>();
 
-    if (!literal || literal->value.getType() != Field::Types::UInt64 || literal->value.safeGet<UInt64>() == 0 || literal->value.safeGet<UInt64>() > ColumnVariant::MAX_NESTED_COLUMNS)
-        throw Exception(ErrorCodes::UNEXPECTED_AST_STRUCTURE, "'max_types' argument for Dynamic type should be a positive integer between 1 and 255");
+    if (!literal || literal->value.getType() != Field::Types::UInt64 || literal->value.safeGet<UInt64>() > ColumnDynamic::MAX_DYNAMIC_TYPES_LIMIT)
+        throw Exception(ErrorCodes::UNEXPECTED_AST_STRUCTURE, "'max_types' argument for Dynamic type should be a positive integer between 0 and {}", ColumnDynamic::MAX_DYNAMIC_TYPES_LIMIT);
 
     return std::make_shared<DataTypeDynamic>(literal->value.safeGet<UInt64>());
 }
@@ -84,30 +86,72 @@ void registerDataTypeDynamic(DataTypeFactory & factory)
 
 std::unique_ptr<IDataType::SubstreamData> DataTypeDynamic::getDynamicSubcolumnData(std::string_view subcolumn_name, const DB::IDataType::SubstreamData & data, bool throw_if_null) const
 {
-    auto [subcolumn_type_name, subcolumn_nested_name] = Nested::splitName(subcolumn_name);
+    auto [type_subcolumn_name, subcolumn_nested_name] = Nested::splitName(subcolumn_name);
     /// Check if requested subcolumn is a valid data type.
-    auto subcolumn_type = DataTypeFactory::instance().tryGet(String(subcolumn_type_name));
+    auto subcolumn_type = DataTypeFactory::instance().tryGet(String(type_subcolumn_name));
     if (!subcolumn_type)
     {
         if (throw_if_null)
-            throw Exception(ErrorCodes::ILLEGAL_COLUMN, "Dynamic type doesn't have subcolumn '{}'", subcolumn_type_name);
+            throw Exception(ErrorCodes::ILLEGAL_COLUMN, "Dynamic type doesn't have subcolumn '{}'", type_subcolumn_name);
         return nullptr;
     }
 
     std::unique_ptr<SubstreamData> res = std::make_unique<SubstreamData>(subcolumn_type->getDefaultSerialization());
     res->type = subcolumn_type;
     std::optional<ColumnVariant::Discriminator> discriminator;
+    ColumnPtr null_map_for_variant_from_shared_variant;
     if (data.column)
     {
         /// If column was provided, we should extract subcolumn from Dynamic column.
         const auto & dynamic_column = assert_cast<const ColumnDynamic &>(*data.column);
         const auto & variant_info = dynamic_column.getVariantInfo();
+        const auto & variant_column = dynamic_column.getVariantColumn();
+        const auto & shared_variant = dynamic_column.getSharedVariant();
         /// Check if provided Dynamic column has subcolumn of this type.
-        auto it = variant_info.variant_name_to_discriminator.find(subcolumn_type->getName());
+        String subcolumn_type_name = subcolumn_type->getName();
+        auto it = variant_info.variant_name_to_discriminator.find(subcolumn_type_name);
         if (it != variant_info.variant_name_to_discriminator.end())
         {
             discriminator = it->second;
-            res->column = dynamic_column.getVariantColumn().getVariantPtrByGlobalDiscriminator(*discriminator);
+            res->column = variant_column.getVariantPtrByGlobalDiscriminator(*discriminator);
+        }
+        /// Otherwise if there is data in shared variant try to find requested type there.
+        else if (!shared_variant.empty())
+        {
+            /// Create null map for resulting subcolumn to make it Nullable.
+            auto null_map_column = ColumnUInt8::create();
+            NullMap & null_map = assert_cast<ColumnUInt8 &>(*null_map_column).getData();
+            null_map.reserve(variant_column.size());
+            auto subcolumn = subcolumn_type->createColumn();
+            auto shared_variant_local_discr = variant_column.localDiscriminatorByGlobal(dynamic_column.getSharedVariantDiscriminator());
+            const auto & local_discriminators = variant_column.getLocalDiscriminators();
+            const auto & offsets = variant_column.getOffsets();
+            const FormatSettings format_settings;
+            for (size_t i = 0; i != local_discriminators.size(); ++i)
+            {
+                if (local_discriminators[i] == shared_variant_local_discr)
+                {
+                    auto value = shared_variant.getDataAt(offsets[i]);
+                    ReadBufferFromMemory buf(value.data, value.size);
+                    auto type = decodeDataType(buf);
+                    if (type->getName() == subcolumn_type_name)
+                    {
+                        dynamic_column.getVariantSerialization(subcolumn_type, subcolumn_type_name)->deserializeBinary(*subcolumn, buf, format_settings);
+                        null_map.push_back(0);
+                    }
+                    else
+                    {
+                        null_map.push_back(1);
+                    }
+                }
+                else
+                {
+                    null_map.push_back(1);
+                }
+            }
+
+            res->column = std::move(subcolumn);
+            null_map_for_variant_from_shared_variant = std::move(null_map_column);
         }
     }
 
@@ -125,7 +169,7 @@ std::unique_ptr<IDataType::SubstreamData> DataTypeDynamic::getDynamicSubcolumnDa
             return nullptr;
     }
 
-    res->serialization = std::make_shared<SerializationDynamicElement>(res->serialization, subcolumn_type->getName(), is_null_map_subcolumn);
+    res->serialization = std::make_shared<SerializationDynamicElement>(res->serialization, subcolumn_type->getName(), String(subcolumn_nested_name), is_null_map_subcolumn);
     /// Make resulting subcolumn Nullable only if type subcolumn can be inside Nullable or can be LowCardinality(Nullable()).
     bool make_subcolumn_nullable = subcolumn_type->canBeInsideNullable() || subcolumn_type->lowCardinality();
     if (!is_null_map_subcolumn && make_subcolumn_nullable)
@@ -133,10 +177,10 @@ std::unique_ptr<IDataType::SubstreamData> DataTypeDynamic::getDynamicSubcolumnDa
 
     if (data.column)
     {
+        /// Check if provided Dynamic column has subcolumn of this type. In this case we should use VariantSubcolumnCreator/VariantNullMapSubcolumnCreator to
+        /// create full subcolumn from variant according to discriminators.
         if (discriminator)
         {
-            /// Provided Dynamic column has subcolumn of this type, we should use VariantSubcolumnCreator/VariantNullMapSubcolumnCreator to
-            /// create full subcolumn from variant according to discriminators.
             const auto & variant_column = assert_cast<const ColumnDynamic &>(*data.column).getVariantColumn();
             std::unique_ptr<ISerialization::ISubcolumnCreator> creator;
             if (is_null_map_subcolumn)
@@ -154,6 +198,21 @@ std::unique_ptr<IDataType::SubstreamData> DataTypeDynamic::getDynamicSubcolumnDa
                     make_subcolumn_nullable);
             res->column = creator->create(res->column);
         }
+        /// Check if requested type was extracted from shared variant. In this case we should use
+        /// VariantSubcolumnCreator to create full subcolumn from variant according to created null map.
+        else if (null_map_for_variant_from_shared_variant)
+        {
+            if (is_null_map_subcolumn)
+            {
+                res->column = null_map_for_variant_from_shared_variant;
+            }
+            else
+            {
+                SerializationVariantElement::VariantSubcolumnCreator creator(
+                    null_map_for_variant_from_shared_variant, "", 0, 0, make_subcolumn_nullable, null_map_for_variant_from_shared_variant);
+                res->column = creator.create(res->column);
+            }
+        }
         /// Provided Dynamic column doesn't have subcolumn of this type, just create column filled with default values.
         else if (is_null_map_subcolumn)
         {

src/DataTypes/DataTypeFactory.cpp

@@ -150,6 +150,12 @@ DataTypePtr DataTypeFactory::getCustom(DataTypeCustomDescPtr customization) cons
     return type;
 }
 
+DataTypePtr DataTypeFactory::getCustom(const String & base_name, DataTypeCustomDescPtr customization) const
+{
+    auto type = get(base_name);
+    type->setCustomization(std::move(customization));
+    return type;
+}
 
 void DataTypeFactory::registerDataType(const String & family_name, Value creator, Case case_sensitiveness)
 {

src/DataTypes/DataTypeFactory.h

@@ -34,6 +34,7 @@ public:
     DataTypePtr get(const String & family_name, const ASTPtr & parameters) const;
     DataTypePtr get(const ASTPtr & ast) const;
     DataTypePtr getCustom(DataTypeCustomDescPtr customization) const;
+    DataTypePtr getCustom(const String & base_name, DataTypeCustomDescPtr customization) const;
 
     /// Return nullptr in case of error.
     DataTypePtr tryGet(const String & full_name) const;

src/DataTypes/DataTypeTuple.cpp

@@ -192,17 +192,12 @@ MutableColumnPtr DataTypeTuple::createColumn() const
 
 MutableColumnPtr DataTypeTuple::createColumn(const ISerialization & serialization) const
 {
-    /// If we read Tuple as Variant subcolumn, it may be wrapped to SerializationVariantElement.
-    /// Here we don't need it, so we drop this wrapper.
-    const auto * current_serialization = &serialization;
-    while (const auto * serialization_variant_element = typeid_cast<const SerializationVariantElement *>(current_serialization))
-        current_serialization = serialization_variant_element->getNested().get();
-
-    /// If we read subcolumn of nested Tuple, it may be wrapped to SerializationNamed
+    /// If we read subcolumn of nested Tuple or this Tuple is a subcolumn, it may be wrapped to SerializationWrapper
     /// several times to allow to reconstruct the substream path name.
     /// Here we don't need substream path name, so we drop first several wrapper serializations.
-    while (const auto * serialization_named = typeid_cast<const SerializationNamed *>(current_serialization))
-        current_serialization = serialization_named->getNested().get();
+    const auto * current_serialization = &serialization;
+    while (const auto * serialization_wrapper = dynamic_cast<const SerializationWrapper *>(current_serialization))
+        current_serialization = serialization_wrapper->getNested().get();
 
     const auto * serialization_tuple = typeid_cast<const SerializationTuple *>(current_serialization);
     if (!serialization_tuple)

src/DataTypes/DataTypeVariant.cpp

@@ -117,7 +117,7 @@ bool DataTypeVariant::equals(const IDataType & rhs) const
 
         /// The same data types with different custom names considered different.
         /// For example, UInt8 and Bool.
-        if ((variants[i]->hasCustomName() || rhs_variant.variants[i]) && variants[i]->getName() != rhs_variant.variants[i]->getName())
+        if ((variants[i]->hasCustomName() || rhs_variant.variants[i]->hasCustomName()) && variants[i]->getName() != rhs_variant.variants[i]->getName())
             return false;
     }
 

src/DataTypes/DataTypesBinaryEncoding.cpp

@@ -444,7 +444,7 @@ void encodeDataType(const DataTypePtr & type, WriteBuffer & buf)
         case BinaryTypeIndex::Dynamic:
         {
             const auto & dynamic_type = assert_cast<const DataTypeDynamic &>(*type);
-            /// Maximum number of dynamic types is 255, we can write it as 1 byte.
+            /// Maximum number of dynamic types is 254, we can write it as 1 byte.
             writeBinary(UInt8(dynamic_type.getMaxDynamicTypes()), buf);
             break;
         }

src/DataTypes/Serializations/SerializationDynamic.cpp

@@ -27,15 +27,21 @@ namespace ErrorCodes
 struct SerializeBinaryBulkStateDynamic : public ISerialization::SerializeBinaryBulkState
 {
     SerializationDynamic::DynamicStructureSerializationVersion structure_version;
+    size_t max_dynamic_types;
     DataTypePtr variant_type;
     Names variant_names;
     SerializationPtr variant_serialization;
     ISerialization::SerializeBinaryBulkStatePtr variant_state;
 
-    /// Variants statistics. Map (Variant name) -> (Variant size).
-    ColumnDynamic::Statistics statistics = { .source = ColumnDynamic::Statistics::Source::READ, .data = {} };
+    /// Variants statistics.
+    ColumnDynamic::Statistics statistics;
+    /// If true, statistics will be recalculated during serialization.
+    bool recalculate_statistics = false;
 
-    explicit SerializeBinaryBulkStateDynamic(UInt64 structure_version_) : structure_version(structure_version_) {}
+    explicit SerializeBinaryBulkStateDynamic(UInt64 structure_version_)
+        : structure_version(structure_version_), statistics(ColumnDynamic::Statistics::Source::READ)
+    {
+    }
 };
 
 struct DeserializeBinaryBulkStateDynamic : public ISerialization::DeserializeBinaryBulkState
@@ -106,20 +112,41 @@ void SerializationDynamic::serializeBinaryBulkStatePrefix(
     writeBinaryLittleEndian(structure_version, *stream);
     auto dynamic_state = std::make_shared<SerializeBinaryBulkStateDynamic>(structure_version);
 
+    dynamic_state->max_dynamic_types = column_dynamic.getMaxDynamicTypes();
+    /// Write max_dynamic_types parameter, because it can differ from the max_dynamic_types
+    /// that is specified in the Dynamic type (we could decrease it before merge).
+    writeBinaryLittleEndian(dynamic_state->max_dynamic_types, *stream);
+
     dynamic_state->variant_type = variant_info.variant_type;
     dynamic_state->variant_names = variant_info.variant_names;
     const auto & variant_column = column_dynamic.getVariantColumn();
 
-    /// Write internal Variant type name.
+    /// Write information about variants.
+    size_t num_variants = dynamic_state->variant_names.size() - 1; /// Don't write shared variant, Dynamic column should always have it.
+    writeBinaryLittleEndian(num_variants, *stream);
     if (settings.data_types_binary_encoding)
-        encodeDataType(dynamic_state->variant_type, *stream);
+    {
+        const auto & variants = assert_cast<const DataTypeVariant &>(*dynamic_state->variant_type).getVariants();
+        for (const auto & variant: variants)
+        {
+            if (variant->getName() != ColumnDynamic::getSharedVariantTypeName())
+                encodeDataType(variant, *stream);
+        }
+    }
     else
-        writeStringBinary(dynamic_state->variant_type->getName(), *stream);
+    {
+        for (const auto & name : dynamic_state->variant_names)
+        {
+            if (name != ColumnDynamic::getSharedVariantTypeName())
+                writeStringBinary(name, *stream);
+        }
+    }
 
     /// Write statistics in prefix if needed.
     if (settings.dynamic_write_statistics == SerializeBinaryBulkSettings::DynamicStatisticsMode::PREFIX)
     {
         const auto & statistics = column_dynamic.getStatistics();
+        /// First, write statistics for usual variants.
         for (size_t i = 0; i != variant_info.variant_names.size(); ++i)
         {
             size_t size = 0;
@@ -129,13 +156,55 @@ void SerializationDynamic::serializeBinaryBulkStatePrefix(
             ///   - statistics read from the data part during deserialization of Dynamic column (Statistics::Source::READ).
             /// We can rely only on statistics calculated during the merge, because column with statistics that was read
             /// during deserialization from some data part could be filtered/limited/transformed/etc and so the statistics can be outdated.
-            if (!statistics.data.empty() && statistics.source == ColumnDynamic::Statistics::Source::MERGE)
-                size = statistics.data.at(variant_info.variant_names[i]);
+            if (statistics && statistics->source == ColumnDynamic::Statistics::Source::MERGE)
+                size = statistics->variants_statistics.at(variant_info.variant_names[i]);
             /// Otherwise we can use only variant sizes from current column.
             else
                 size = variant_column.getVariantByGlobalDiscriminator(i).size();
             writeVarUInt(size, *stream);
         }
+
+        /// Second, write statistics for variants in shared variant.
+        /// Check if we have statistics calculated during merge of some data parts (Statistics::Source::MERGE).
+        if (statistics && statistics->source == ColumnDynamic::Statistics::Source::MERGE)
+        {
+            writeVarUInt(statistics->shared_variants_statistics.size(), *stream);
+            for (const auto & [variant_name, size] : statistics->shared_variants_statistics)
+            {
+                writeStringBinary(variant_name, *stream);
+                writeVarUInt(size, *stream);
+            }
+        }
+        /// If we don't have statistics for shared variants from merge, calculate it from the column.
+        else
+        {
+            std::unordered_map<String, size_t> shared_variants_statistics;
+            const auto & shared_variant = column_dynamic.getSharedVariant();
+            for (size_t i = 0; i != shared_variant.size(); ++i)
+            {
+                auto value = shared_variant.getDataAt(i);
+                ReadBufferFromMemory buf(value.data, value.size);
+                auto type = decodeDataType(buf);
+                auto type_name = type->getName();
+                if (auto it = shared_variants_statistics.find(type_name); it != shared_variants_statistics.end())
+                    ++it->second;
+                else if (shared_variants_statistics.size() < ColumnDynamic::Statistics::MAX_SHARED_VARIANT_STATISTICS_SIZE)
+                    shared_variants_statistics.emplace(type_name, 1);
+            }
+
+            writeVarUInt(shared_variants_statistics.size(), *stream);
+            for (const auto & [variant_name, size] : shared_variants_statistics)
+            {
+                writeStringBinary(variant_name, *stream);
+                writeVarUInt(size, *stream);
+            }
+        }
+    }
+    /// Otherwise statistics will be written in the suffix, in this case we will recalculate
+    /// statistics during serialization to make it more precise.
+    else
+    {
+        dynamic_state->recalculate_statistics = true;
     }
 
     dynamic_state->variant_serialization = dynamic_state->variant_type->getDefaultSerialization();
@@ -182,33 +251,58 @@ ISerialization::DeserializeBinaryBulkStatePtr SerializationDynamic::deserializeD
         UInt64 structure_version;
         readBinaryLittleEndian(structure_version, *structure_stream);
         auto structure_state = std::make_shared<DeserializeBinaryBulkStateDynamicStructure>(structure_version);
-        /// Read internal Variant type name.
+        /// Read max_dynamic_types parameter.
+        readBinaryLittleEndian(structure_state->max_dynamic_types, *structure_stream);
+        /// Read information about variants.
+        DataTypes variants;
+        size_t num_variants;
+        readBinaryLittleEndian(num_variants, *structure_stream);
+        variants.reserve(num_variants + 1); /// +1 for shared variant.
         if (settings.data_types_binary_encoding)
         {
-            structure_state->variant_type = decodeDataType(*structure_stream);
+            for (size_t i = 0; i != num_variants; ++i)
+                variants.push_back(decodeDataType(*structure_stream));
         }
         else
         {
             String data_type_name;
-            readStringBinary(data_type_name, *structure_stream);
-            structure_state->variant_type = DataTypeFactory::instance().get(data_type_name);
+            for (size_t i = 0; i != num_variants; ++i)
+            {
+                readStringBinary(data_type_name, *structure_stream);
+                variants.push_back(DataTypeFactory::instance().get(data_type_name));
+            }
         }
-        const auto * variant_type = typeid_cast<const DataTypeVariant *>(structure_state->variant_type.get());
-        if (!variant_type)
-            throw Exception(ErrorCodes::INCORRECT_DATA, "Incorrect type of Dynamic nested column, expected Variant, got {}", structure_state->variant_type->getName());
+        /// Add shared variant, Dynamic column should always have it.
+        variants.push_back(ColumnDynamic::getSharedVariantDataType());
+        auto variant_type = std::make_shared<DataTypeVariant>(variants);
 
         /// Read statistics.
         if (settings.dynamic_read_statistics)
         {
-            const auto & variants = variant_type->getVariants();
+            ColumnDynamic::Statistics statistics(ColumnDynamic::Statistics::Source::READ);
+            /// First, read statistics for usual variants.
             size_t variant_size;
-            for (const auto & variant : variants)
+            for (const auto & variant : variant_type->getVariants())
             {
                 readVarUInt(variant_size, *structure_stream);
-                structure_state->statistics.data[variant->getName()] = variant_size;
+                statistics.variants_statistics[variant->getName()] = variant_size;
             }
+
+            /// Second, read statistics for shared variants.
+            size_t statistics_size;
+            readVarUInt(statistics_size, *structure_stream);
+            String variant_name;
+            for (size_t i = 0; i != statistics_size; ++i)
+            {
+                readStringBinary(variant_name, *structure_stream);
+                readVarUInt(variant_size, *structure_stream);
+                statistics.shared_variants_statistics[variant_name] = variant_size;
+            }
+
+            structure_state->statistics = std::make_shared<const ColumnDynamic::Statistics>(std::move(statistics));
         }
 
+        structure_state->variant_type = std::move(variant_type);
         state = structure_state;
         addToSubstreamsDeserializeStatesCache(cache, settings.path, state);
     }
@@ -231,8 +325,16 @@ void SerializationDynamic::serializeBinaryBulkStateSuffix(
     /// Write statistics in suffix if needed.
     if (settings.dynamic_write_statistics == SerializeBinaryBulkSettings::DynamicStatisticsMode::SUFFIX)
     {
+        /// First, write statistics for usual variants.
         for (const auto & variant_name : dynamic_state->variant_names)
-            writeVarUInt(dynamic_state->statistics.data[variant_name], *stream);
+            writeVarUInt(dynamic_state->statistics.variants_statistics[variant_name], *stream);
+        /// Second, write statistics for shared variants.
+        writeVarUInt(dynamic_state->statistics.shared_variants_statistics.size(), *stream);
+        for (const auto & [variant_name, size] : dynamic_state->statistics.shared_variants_statistics)
+        {
+            writeStringBinary(variant_name, *stream);
+            writeVarUInt(size, *stream);
+        }
     }
 
     settings.path.push_back(Substream::DynamicData);
@@ -255,9 +357,42 @@ void SerializationDynamic::serializeBinaryBulkWithMultipleStreams(
     if (!variant_info.variant_type->equals(*dynamic_state->variant_type))
         throw Exception(ErrorCodes::LOGICAL_ERROR, "Mismatch of internal columns of Dynamic. Expected: {}, Got: {}", dynamic_state->variant_type->getName(), variant_info.variant_type->getName());
 
+    if (column_dynamic.getMaxDynamicTypes() != dynamic_state->max_dynamic_types)
+        throw Exception(ErrorCodes::LOGICAL_ERROR, "Mismatch of max_dynamic_types parameter of Dynamic. Expected: {}, Got: {}", dynamic_state->max_dynamic_types, column_dynamic.getMaxDynamicTypes());
+
     settings.path.push_back(Substream::DynamicData);
-    assert_cast<const SerializationVariant &>(*dynamic_state->variant_serialization)
-        .serializeBinaryBulkWithMultipleStreamsAndUpdateVariantStatistics(*variant_column, offset, limit, settings, dynamic_state->variant_state, dynamic_state->statistics.data);
+    if (dynamic_state->recalculate_statistics)
+    {
+        assert_cast<const SerializationVariant &>(*dynamic_state->variant_serialization)
+            .serializeBinaryBulkWithMultipleStreamsAndUpdateVariantStatistics(*variant_column, offset, limit, settings, dynamic_state->variant_state, dynamic_state->statistics.variants_statistics);
+        /// Calculate statistics for shared variants.
+        const auto & shared_variant = column_dynamic.getSharedVariant();
+        if (!shared_variant.empty())
+        {
+            const auto & local_discriminators = variant_column->getLocalDiscriminators();
+            const auto & offsets = variant_column->getOffsets();
+            const auto shared_variant_discr = variant_column->localDiscriminatorByGlobal(column_dynamic.getSharedVariantDiscriminator());
+            size_t end = limit == 0 || offset + limit > local_discriminators.size() ? local_discriminators.size() : offset + limit;
+            for (size_t i = offset; i != end; ++i)
+            {
+                if (local_discriminators[i] == shared_variant_discr)
+                {
+                    auto value = shared_variant.getDataAt(offsets[i]);
+                    ReadBufferFromMemory buf(value.data, value.size);
+                    auto type = decodeDataType(buf);
+                    auto type_name = type->getName();
+                    if (auto it = dynamic_state->statistics.shared_variants_statistics.find(type_name); it != dynamic_state->statistics.shared_variants_statistics.end())
+                        ++it->second;
+                    else if (dynamic_state->statistics.shared_variants_statistics.size() < ColumnDynamic::Statistics::MAX_SHARED_VARIANT_STATISTICS_SIZE)
+                        dynamic_state->statistics.shared_variants_statistics.emplace(type_name, 1);
+                }
+            }
+        }
+    }
+    else
+    {
+        assert_cast<const SerializationVariant &>(*dynamic_state->variant_serialization).serializeBinaryBulkWithMultipleStreams(*variant_column, offset, limit, settings, dynamic_state->variant_state);
+    }
     settings.path.pop_back();
 }
 
@@ -272,13 +407,17 @@ void SerializationDynamic::deserializeBinaryBulkWithMultipleStreams(
         return;
 
     auto mutable_column = column->assumeMutable();
+    auto & column_dynamic = assert_cast<ColumnDynamic &>(*mutable_column);
     auto * dynamic_state = checkAndGetState<DeserializeBinaryBulkStateDynamic>(state);
     auto * structure_state = checkAndGetState<DeserializeBinaryBulkStateDynamicStructure>(dynamic_state->structure_state);
 
     if (mutable_column->empty())
-        mutable_column = ColumnDynamic::create(structure_state->variant_type->createColumn(), structure_state->variant_type, max_dynamic_types, structure_state->statistics);
+    {
+        column_dynamic.setMaxDynamicPaths(structure_state->max_dynamic_types);
+        column_dynamic.setVariantType(structure_state->variant_type);
+        column_dynamic.setStatistics(structure_state->statistics);
+    }
 
-    auto & column_dynamic = assert_cast<ColumnDynamic &>(*mutable_column);
     const auto & variant_info = column_dynamic.getVariantInfo();
     if (!variant_info.variant_type->equals(*structure_state->variant_type))
         throw Exception(ErrorCodes::LOGICAL_ERROR, "Mismatch of internal columns of Dynamic. Expected: {}, Got: {}", structure_state->variant_type->getName(), variant_info.variant_type->getName());
@@ -329,24 +468,42 @@ void SerializationDynamic::serializeBinary(const IColumn & column, size_t row_nu
         encodeDataType(std::make_shared<DataTypeNothing>(), ostr);
         return;
     }
+    /// Check if this value is in shared variant. In this case it's already
+    /// in desired binary format.
+    else if (global_discr == dynamic_column.getSharedVariantDiscriminator())
+    {
+        auto value = dynamic_column.getSharedVariant().getDataAt(variant_column.offsetAt(row_num));
+        ostr.write(value.data, value.size);
+        return;
+    }
 
     const auto & variant_type = assert_cast<const DataTypeVariant &>(*variant_info.variant_type).getVariant(global_discr);
+    const auto & variant_type_name = variant_info.variant_names[global_discr];
     encodeDataType(variant_type, ostr);
-    variant_type->getDefaultSerialization()->serializeBinary(variant_column.getVariantByGlobalDiscriminator(global_discr), variant_column.offsetAt(row_num), ostr, settings);
+    dynamic_column.getVariantSerialization(variant_type, variant_type_name)->serializeBinary(variant_column.getVariantByGlobalDiscriminator(global_discr), variant_column.offsetAt(row_num), ostr, settings);
 }
 
-template <typename DeserializeFunc>
-static void deserializeVariant(
+template <typename ReturnType = void, typename DeserializeFunc>
+static ReturnType deserializeVariant(
     ColumnVariant & variant_column,
-    const DataTypePtr & variant_type,
+    const SerializationPtr & variant_serialization,
     ColumnVariant::Discriminator global_discr,
     ReadBuffer & istr,
     DeserializeFunc deserialize)
 {
     auto & variant = variant_column.getVariantByGlobalDiscriminator(global_discr);
-    deserialize(*variant_type->getDefaultSerialization(), variant, istr);
+    if constexpr (std::is_same_v<ReturnType, bool>)
+    {
+        if (!deserialize(*variant_serialization, variant, istr))
+            return ReturnType(false);
+    }
+    else
+    {
+        deserialize(*variant_serialization, variant, istr);
+    }
     variant_column.getLocalDiscriminators().push_back(variant_column.localDiscriminatorByGlobal(global_discr));
     variant_column.getOffsets().push_back(variant.size() - 1);
+    return ReturnType(true);
 }
 
 void SerializationDynamic::deserializeBinary(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -360,11 +517,12 @@ void SerializationDynamic::deserializeBinary(IColumn & column, ReadBuffer & istr
     }
 
     auto variant_type_name = variant_type->getName();
+    const auto & variant_serialization = dynamic_column.getVariantSerialization(variant_type, variant_type_name);
     const auto & variant_info = dynamic_column.getVariantInfo();
     auto it = variant_info.variant_name_to_discriminator.find(variant_type_name);
     if (it != variant_info.variant_name_to_discriminator.end())
     {
-        deserializeVariant(dynamic_column.getVariantColumn(), variant_type, it->second, istr, [&settings](const ISerialization & serialization, IColumn & variant, ReadBuffer & buf){ serialization.deserializeBinary(variant, buf, settings); });
+        deserializeVariant(dynamic_column.getVariantColumn(), variant_serialization, it->second, istr, [&settings](const ISerialization & serialization, IColumn & variant, ReadBuffer & buf){ serialization.deserializeBinary(variant, buf, settings); });
         return;
     }
 
@@ -372,25 +530,15 @@ void SerializationDynamic::deserializeBinary(IColumn & column, ReadBuffer & istr
     if (dynamic_column.addNewVariant(variant_type))
     {
         auto discr = variant_info.variant_name_to_discriminator.at(variant_type_name);
-        deserializeVariant(dynamic_column.getVariantColumn(), variant_type, discr, istr, [&settings](const ISerialization & serialization, IColumn & variant, ReadBuffer & buf){ serialization.deserializeBinary(variant, buf, settings); });
+        deserializeVariant(dynamic_column.getVariantColumn(), variant_serialization, discr, istr, [&settings](const ISerialization & serialization, IColumn & variant, ReadBuffer & buf){ serialization.deserializeBinary(variant, buf, settings); });
         return;
     }
 
     /// We reached maximum number of variants and couldn't add new variant.
-    /// This case should be really rare in real use cases.
-    /// We should always be able to add String variant and insert value as String.
-    dynamic_column.addStringVariant();
+    /// In this case we insert this value into shared variant in binary form.
     auto tmp_variant_column = variant_type->createColumn();
-    variant_type->getDefaultSerialization()->deserializeBinary(*tmp_variant_column, istr, settings);
-    auto string_column = castColumn(ColumnWithTypeAndName(tmp_variant_column->getPtr(), variant_type, ""), std::make_shared<DataTypeString>());
-    auto & variant_column = dynamic_column.getVariantColumn();
-    variant_column.insertIntoVariantFrom(variant_info.variant_name_to_discriminator.at("String"), *string_column, 0);
-}
-
-void SerializationDynamic::serializeTextCSV(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
-{
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeTextCSV(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    variant_serialization->deserializeBinary(*tmp_variant_column, istr, settings);
+    dynamic_column.insertValueIntoSharedVariant(*tmp_variant_column, variant_type, variant_type_name, 0);
 }
 
 template <typename ReadFieldFunc, typename TryDeserializeVariantFunc, typename DeserializeVariant>
@@ -406,6 +554,7 @@ static void deserializeTextImpl(
     auto & dynamic_column = assert_cast<ColumnDynamic &>(column);
     auto & variant_column = dynamic_column.getVariantColumn();
     const auto & variant_info = dynamic_column.getVariantInfo();
+    const auto & variant_types = assert_cast<const DataTypeVariant &>(*variant_info.variant_type).getVariants();
     String field = read_field(istr);
     auto field_buf = std::make_unique<ReadBufferFromString>(field);
     JSONInferenceInfo json_info;
@@ -413,27 +562,81 @@ static void deserializeTextImpl(
     if (escaping_rule == FormatSettings::EscapingRule::JSON)
         transformFinalInferredJSONTypeIfNeeded(variant_type, settings, &json_info);
 
-    if (checkIfTypeIsComplete(variant_type) && dynamic_column.addNewVariant(variant_type))
+    /// If inferred type is not complete, we cannot add it as a new variant.
+    /// Let's try to deserialize this field into existing variants.
+    /// If failed, insert this value as String.
+    if (!checkIfTypeIsComplete(variant_type))
+    {
+        size_t shared_variant_discr = dynamic_column.getSharedVariantDiscriminator();
+        for (size_t i = 0; i != variant_types.size(); ++i)
+        {
+            field_buf = std::make_unique<ReadBufferFromString>(field);
+            if (i != shared_variant_discr
+                && deserializeVariant<bool>(
+                    variant_column,
+                    dynamic_column.getVariantSerialization(variant_types[i], variant_info.variant_names[i]),
+                    i,
+                    *field_buf,
+                    try_deserialize_variant))
+                return;
+        }
+
+        variant_type = std::make_shared<DataTypeString>();
+        /// To be able to deserialize field as String with Quoted escaping rule, it should be quoted.
+        if (escaping_rule == FormatSettings::EscapingRule::Quoted && (field.size() < 2 || field.front() != '\'' || field.back() != '\''))
+            field = "'" + field + "'";
+    }
+    else if (dynamic_column.addNewVariant(variant_type, variant_type->getName()))
     {
         auto discr = variant_info.variant_name_to_discriminator.at(variant_type->getName());
-        deserializeVariant(dynamic_column.getVariantColumn(), variant_type, discr, *field_buf, deserialize_variant);
+        deserializeVariant(dynamic_column.getVariantColumn(), dynamic_column.getVariantSerialization(variant_type), discr, *field_buf, deserialize_variant);
         return;
     }
 
-    /// We couldn't infer type or add new variant. Try to insert field into current variants.
+    /// We couldn't infer type or add new variant. Insert it into shared variant.
+    auto tmp_variant_column = variant_type->createColumn();
     field_buf = std::make_unique<ReadBufferFromString>(field);
-    if (try_deserialize_variant(*variant_info.variant_type->getDefaultSerialization(), variant_column, *field_buf))
-        return;
+    auto variant_type_name = variant_type->getName();
+    deserialize_variant(*dynamic_column.getVariantSerialization(variant_type, variant_type_name), *tmp_variant_column, *field_buf);
+    dynamic_column.insertValueIntoSharedVariant(*tmp_variant_column, variant_type, variant_type_name, 0);
+}
 
-    /// We couldn't insert field into any existing variant, add String variant and read value as String.
-    dynamic_column.addStringVariant();
+template <typename NestedSerialize>
+static void serializeTextImpl(
+    const IColumn & column,
+    size_t row_num,
+    WriteBuffer & ostr,
+    const FormatSettings & settings,
+    NestedSerialize nested_serialize)
+{
+    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
+    const auto & variant_column = dynamic_column.getVariantColumn();
+    /// Check if this row has value in shared variant. In this case we should first deserialize it from binary format.
+    if (variant_column.globalDiscriminatorAt(row_num) == dynamic_column.getSharedVariantDiscriminator())
+    {
+        auto value = dynamic_column.getSharedVariant().getDataAt(variant_column.offsetAt(row_num));
+        ReadBufferFromMemory buf(value.data, value.size);
+        auto variant_type = decodeDataType(buf);
+        auto tmp_variant_column = variant_type->createColumn();
+        auto variant_serialization = dynamic_column.getVariantSerialization(variant_type);
+        variant_serialization->deserializeBinary(*tmp_variant_column, buf, settings);
+        nested_serialize(*variant_serialization, *tmp_variant_column, 0, ostr);
+    }
+    /// Otherwise just use serialization for Variant.
+    else
+    {
+        nested_serialize(*dynamic_column.getVariantInfo().variant_type->getDefaultSerialization(), variant_column, row_num, ostr);
+    }
+}
 
-    if (escaping_rule == FormatSettings::EscapingRule::Quoted && (field.size() < 2 || field.front() != '\'' || field.back() != '\''))
-        field = "'" + field + "'";
+void SerializationDynamic::serializeTextCSV(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
+{
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeTextCSV(col, row, buf, settings);
+    };
 
-    field_buf = std::make_unique<ReadBufferFromString>(field);
-    auto string_discr = variant_info.variant_name_to_discriminator.at("String");
-    deserializeVariant(dynamic_column.getVariantColumn(), std::make_shared<DataTypeString>(), string_discr, *field_buf, deserialize_variant);
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 void SerializationDynamic::deserializeTextCSV(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -466,8 +669,12 @@ bool SerializationDynamic::tryDeserializeTextCSV(DB::IColumn & column, DB::ReadB
 
 void SerializationDynamic::serializeTextEscaped(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
 {
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeTextEscaped(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeTextEscaped(col, row, buf, settings);
+    };
+
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 void SerializationDynamic::deserializeTextEscaped(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -500,8 +707,12 @@ bool SerializationDynamic::tryDeserializeTextEscaped(DB::IColumn & column, DB::R
 
 void SerializationDynamic::serializeTextQuoted(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
 {
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeTextQuoted(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeTextQuoted(col, row, buf, settings);
+    };
+
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 void SerializationDynamic::deserializeTextQuoted(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -534,8 +745,12 @@ bool SerializationDynamic::tryDeserializeTextQuoted(DB::IColumn & column, DB::Re
 
 void SerializationDynamic::serializeTextJSON(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
 {
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeTextJSON(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeTextJSON(col, row, buf, settings);
+    };
+
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 void SerializationDynamic::deserializeTextJSON(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -568,8 +783,12 @@ bool SerializationDynamic::tryDeserializeTextJSON(DB::IColumn & column, DB::Read
 
 void SerializationDynamic::serializeTextRaw(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
 {
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeTextRaw(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeTextRaw(col, row, buf, settings);
+    };
+
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 void SerializationDynamic::deserializeTextRaw(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -602,8 +821,12 @@ bool SerializationDynamic::tryDeserializeTextRaw(DB::IColumn & column, DB::ReadB
 
 void SerializationDynamic::serializeText(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
 {
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeText(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeText(col, row, buf, settings);
+    };
+
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 void SerializationDynamic::deserializeWholeText(IColumn & column, ReadBuffer & istr, const FormatSettings & settings) const
@@ -636,8 +859,12 @@ bool SerializationDynamic::tryDeserializeWholeText(DB::IColumn & column, DB::Rea
 
 void SerializationDynamic::serializeTextXML(const IColumn & column, size_t row_num, WriteBuffer & ostr, const FormatSettings & settings) const
 {
-    const auto & dynamic_column = assert_cast<const ColumnDynamic &>(column);
-    dynamic_column.getVariantInfo().variant_type->getDefaultSerialization()->serializeTextXML(dynamic_column.getVariantColumn(), row_num, ostr, settings);
+    auto nested_serialize = [&settings](const ISerialization & serialization, const IColumn & col, size_t row, WriteBuffer & buf)
+    {
+        serialization.serializeTextXML(col, row, buf, settings);
+    };
+
+    serializeTextImpl(column, row_num, ostr, settings, nested_serialize);
 }
 
 }

src/DataTypes/Serializations/SerializationDynamic.h

@@ -105,9 +105,13 @@ private:
     {
         DynamicStructureSerializationVersion structure_version;
         DataTypePtr variant_type;
-        ColumnDynamic::Statistics statistics = {.source = ColumnDynamic::Statistics::Source::READ, .data = {}};
+        size_t max_dynamic_types;
+        ColumnDynamic::StatisticsPtr statistics;
 
-        explicit DeserializeBinaryBulkStateDynamicStructure(UInt64 structure_version_) : structure_version(structure_version_) {}
+        explicit DeserializeBinaryBulkStateDynamicStructure(UInt64 structure_version_)
+            : structure_version(structure_version_)
+        {
+        }
     };
 
     size_t max_dynamic_types;

src/DataTypes/Serializations/SerializationDynamicElement.cpp

@@ -4,7 +4,10 @@
 #include <DataTypes/Serializations/SerializationDynamic.h>
 #include <DataTypes/DataTypeVariant.h>
 #include <DataTypes/DataTypeFactory.h>
+#include <DataTypes/DataTypesBinaryEncoding.h>
 #include <Columns/ColumnDynamic.h>
+#include <Columns/ColumnLowCardinality.h>
+#include <Formats/FormatSettings.h>
 #include <IO/ReadHelpers.h>
 
 namespace DB
@@ -21,6 +24,8 @@ struct DeserializeBinaryBulkStateDynamicElement : public ISerialization::Deseria
     ISerialization::DeserializeBinaryBulkStatePtr structure_state;
     SerializationPtr variant_serialization;
     ISerialization::DeserializeBinaryBulkStatePtr variant_element_state;
+    bool read_from_shared_variant;
+    ColumnPtr shared_variant;
 };
 
 void SerializationDynamicElement::enumerateStreams(
@@ -73,9 +78,10 @@ void SerializationDynamicElement::deserializeBinaryBulkStatePrefix(
 
     auto dynamic_element_state = std::make_shared<DeserializeBinaryBulkStateDynamicElement>();
     dynamic_element_state->structure_state = std::move(structure_state);
-    const auto & variant_type = checkAndGetState<SerializationDynamic::DeserializeBinaryBulkStateDynamicStructure>(dynamic_element_state->structure_state)->variant_type;
+    const auto & variant_type = assert_cast<const DataTypeVariant &>(
+        *checkAndGetState<SerializationDynamic::DeserializeBinaryBulkStateDynamicStructure>(dynamic_element_state->structure_state)->variant_type);
     /// Check if we actually have required element in the Variant.
-    if (auto global_discr = assert_cast<const DataTypeVariant &>(*variant_type).tryGetVariantDiscriminator(dynamic_element_name))
+    if (auto global_discr = variant_type.tryGetVariantDiscriminator(dynamic_element_name))
     {
         settings.path.push_back(Substream::DynamicData);
         if (is_null_map_subcolumn)
@@ -83,6 +89,21 @@ void SerializationDynamicElement::deserializeBinaryBulkStatePrefix(
         else
             dynamic_element_state->variant_serialization = std::make_shared<SerializationVariantElement>(nested_serialization, dynamic_element_name, *global_discr);
         dynamic_element_state->variant_serialization->deserializeBinaryBulkStatePrefix(settings, dynamic_element_state->variant_element_state, cache);
+        dynamic_element_state->read_from_shared_variant = false;
+        settings.path.pop_back();
+    }
+    /// If we don't have this element in the Variant, we will read shared variant and try to find it there.
+    else
+    {
+        auto shared_variant_global_discr = variant_type.tryGetVariantDiscriminator(ColumnDynamic::getSharedVariantTypeName());
+        chassert(shared_variant_global_discr.has_value());
+        settings.path.push_back(Substream::DynamicData);
+        dynamic_element_state->variant_serialization = std::make_shared<SerializationVariantElement>(
+            ColumnDynamic::getSharedVariantDataType()->getDefaultSerialization(),
+            ColumnDynamic::getSharedVariantTypeName(),
+            *shared_variant_global_discr);
+        dynamic_element_state->variant_serialization->deserializeBinaryBulkStatePrefix(settings, dynamic_element_state->variant_element_state, cache);
+        dynamic_element_state->read_from_shared_variant = true;
         settings.path.pop_back();
     }
 
@@ -115,23 +136,103 @@ void SerializationDynamicElement::deserializeBinaryBulkWithMultipleStreams(
 
     auto * dynamic_element_state = checkAndGetState<DeserializeBinaryBulkStateDynamicElement>(state);
 
-    if (dynamic_element_state->variant_serialization)
+    /// Check if this subcolumn should not be read from shared variant.
+    /// In this case just read data from the corresponding variant.
+    if (!dynamic_element_state->read_from_shared_variant)
     {
         settings.path.push_back(Substream::DynamicData);
-        dynamic_element_state->variant_serialization->deserializeBinaryBulkWithMultipleStreams(result_column, limit, settings, dynamic_element_state->variant_element_state, cache);
+        dynamic_element_state->variant_serialization->deserializeBinaryBulkWithMultipleStreams(
+            result_column, limit, settings, dynamic_element_state->variant_element_state, cache);
         settings.path.pop_back();
     }
-    else if (is_null_map_subcolumn)
-    {
-        auto mutable_column = result_column->assumeMutable();
-        auto & data = assert_cast<ColumnUInt8 &>(*mutable_column).getData();
-        data.resize_fill(data.size() + limit, 1);
-    }
+    /// Otherwise, read the shared variant column and extract requested type from it.
     else
     {
-        auto mutable_column = result_column->assumeMutable();
-        mutable_column->insertManyDefaults(limit);
-        result_column = std::move(mutable_column);
+        settings.path.push_back(Substream::DynamicData);
+        /// Initialize shared_variant column if needed.
+        if (result_column->empty())
+            dynamic_element_state->shared_variant = makeNullable(ColumnDynamic::getSharedVariantDataType()->createColumn());
+        size_t prev_size = result_column->size();
+        dynamic_element_state->variant_serialization->deserializeBinaryBulkWithMultipleStreams(
+            dynamic_element_state->shared_variant, limit, settings, dynamic_element_state->variant_element_state, cache);
+        settings.path.pop_back();
+
+        /// If we need to read a subcolumn from variant column, create an empty variant column, fill it and extract subcolumn.
+        auto variant_type = DataTypeFactory::instance().get(dynamic_element_name);
+        auto result_type = makeNullableOrLowCardinalityNullableSafe(variant_type);
+        MutableColumnPtr variant_column = nested_subcolumn.empty() || is_null_map_subcolumn ? result_column->assumeMutable() : result_type->createColumn();
+        variant_column->reserve(variant_column->size() + limit);
+        MutableColumnPtr non_nullable_variant_column = variant_column->assumeMutable();
+        NullMap * null_map = nullptr;
+        bool is_low_cardinality_nullable = isColumnLowCardinalityNullable(*variant_column);
+        /// Resulting subolumn can be Nullable, but value is serialized in shared variant as non-Nullable.
+        /// Extract non-nullable column and remember the null map to fill it during deserialization.
+        if (isColumnNullable(*variant_column))
+        {
+            auto & nullable_variant_column = assert_cast<ColumnNullable &>(*variant_column);
+            non_nullable_variant_column = nullable_variant_column.getNestedColumnPtr()->assumeMutable();
+            null_map = &nullable_variant_column.getNullMapData();
+        }
+        else if (is_null_map_subcolumn)
+        {
+            null_map = &assert_cast<ColumnUInt8 &>(*variant_column).getData();
+        }
+
+        auto variant_serialization = variant_type->getDefaultSerialization();
+
+        const auto & nullable_shared_variant = assert_cast<const ColumnNullable &>(*dynamic_element_state->shared_variant);
+        const auto & shared_null_map = nullable_shared_variant.getNullMapData();
+        const auto & shared_variant = assert_cast<const ColumnString &>(nullable_shared_variant.getNestedColumn());
+        const FormatSettings format_settings;
+        for (size_t i = prev_size; i != shared_variant.size(); ++i)
+        {
+            if (!shared_null_map[i])
+            {
+                auto value = shared_variant.getDataAt(i);
+                ReadBufferFromMemory buf(value.data, value.size);
+                auto type = decodeDataType(buf);
+                if (type->getName() == dynamic_element_name)
+                {
+                    /// When requested type is LowCardinality the subcolumn type name will be LowCardinality(Nullable).
+                    /// Value in shared variant is serialized as LowCardinality and we cannot simply deserialize it
+                    /// inside LowCardinality(Nullable) column (it will try to deserialize null bit). In this case we
+                    /// have to create temporary LowCardinality column, deserialize value into it and insert it into
+                    /// resulting LowCardinality(Nullable) (insertion from LowCardinality column to LowCardinality(Nullable)
+                    /// column is allowed).
+                    if (is_low_cardinality_nullable)
+                    {
+                        auto tmp_column = variant_type->createColumn();
+                        variant_serialization->deserializeBinary(*tmp_column, buf, format_settings);
+                        non_nullable_variant_column->insertFrom(*tmp_column, 0);
+                    }
+                    else if (is_null_map_subcolumn)
+                    {
+                        null_map->push_back(0);
+                    }
+                    else
+                    {
+                        variant_serialization->deserializeBinary(*non_nullable_variant_column, buf, format_settings);
+                        if (null_map)
+                            null_map->push_back(0);
+                    }
+                }
+                else
+                {
+                    variant_column->insertDefault();
+                }
+            }
+            else
+            {
+                variant_column->insertDefault();
+            }
+        }
+
+        /// Extract nested subcolumn if needed.
+        if (!nested_subcolumn.empty() && !is_null_map_subcolumn)
+        {
+            auto subcolumn = result_type->getSubcolumn(nested_subcolumn, variant_column->getPtr());
+            result_column->assumeMutable()->insertRangeFrom(*subcolumn, 0, subcolumn->size());
+        }
     }
 }
 

src/DataTypes/Serializations/SerializationDynamicElement.h

@@ -13,11 +13,15 @@ private:
     /// To be able to deserialize Dynamic element as a subcolumn
     /// we need its type name and global discriminator.
     String dynamic_element_name;
+    /// Nested subcolumn of a type dynamic type. For example, for `Tuple(a UInt32)`.a
+    /// subcolumn dynamic_element_name = 'Tuple(a UInt32)' and nested_subcolumn = 'a'.
+    /// Needed to extract nested subcolumn from values in shared variant.
+    String nested_subcolumn;
     bool is_null_map_subcolumn;
 
 public:
-    SerializationDynamicElement(const SerializationPtr & nested_, const String & dynamic_element_name_, bool is_null_map_subcolumn_ = false)
-        : SerializationWrapper(nested_), dynamic_element_name(dynamic_element_name_), is_null_map_subcolumn(is_null_map_subcolumn_)
+    SerializationDynamicElement(const SerializationPtr & nested_, const String & dynamic_element_name_, const String & nested_subcolumn_, bool is_null_map_subcolumn_ = false)
+        : SerializationWrapper(nested_), dynamic_element_name(dynamic_element_name_), nested_subcolumn(nested_subcolumn_), is_null_map_subcolumn(is_null_map_subcolumn_)
     {
     }
 

src/DataTypes/Serializations/SerializationVariantElement.cpp

@@ -305,8 +305,10 @@ SerializationVariantElement::VariantSubcolumnCreator::VariantSubcolumnCreator(
     const String & variant_element_name_,
     ColumnVariant::Discriminator global_variant_discriminator_,
     ColumnVariant::Discriminator local_variant_discriminator_,
-    bool make_nullable_)
+    bool make_nullable_,
+    const ColumnPtr & null_map_)
     : local_discriminators(local_discriminators_)
+    , null_map(null_map_)
     , variant_element_name(variant_element_name_)
     , global_variant_discriminator(global_variant_discriminator_)
     , local_variant_discriminator(local_variant_discriminator_)
@@ -314,12 +316,13 @@ SerializationVariantElement::VariantSubcolumnCreator::VariantSubcolumnCreator(
 {
 }
 
-DataTypePtr SerializationVariantElement::VariantSubcolumnCreator::create(const DB::DataTypePtr & prev) const
+
+DataTypePtr SerializationVariantElement::VariantSubcolumnCreator::create(const DataTypePtr & prev) const
 {
     return make_nullable ? makeNullableOrLowCardinalityNullableSafe(prev) : prev;
 }
 
-SerializationPtr SerializationVariantElement::VariantSubcolumnCreator::create(const DB::SerializationPtr & prev) const
+SerializationPtr SerializationVariantElement::VariantSubcolumnCreator::create(const SerializationPtr & prev) const
 {
     return std::make_shared<SerializationVariantElement>(prev, variant_element_name, global_variant_discriminator);
 }
@@ -339,12 +342,16 @@ ColumnPtr SerializationVariantElement::VariantSubcolumnCreator::create(const DB:
         return res;
     }
 
-    /// In general case we should iterate through discriminators and create null-map for our variant.
-    NullMap null_map;
-    null_map.reserve(local_discriminators->size());
-    const auto & local_discriminators_data = assert_cast<const ColumnVariant::ColumnDiscriminators &>(*local_discriminators).getData();
-    for (auto local_discr : local_discriminators_data)
-        null_map.push_back(local_discr != local_variant_discriminator);
+    /// In general case we should iterate through discriminators and create null-map for our variant if we don't already have it.
+    std::optional<NullMap> null_map_from_discriminators;
+    if (!null_map)
+    {
+        null_map_from_discriminators = NullMap();
+        null_map_from_discriminators->reserve(local_discriminators->size());
+        const auto & local_discriminators_data = assert_cast<const ColumnVariant::ColumnDiscriminators &>(*local_discriminators).getData();
+        for (auto local_discr : local_discriminators_data)
+            null_map_from_discriminators->push_back(local_discr != local_variant_discriminator);
+    }
 
     /// Now we can create new column from null-map and variant column using IColumn::expand.
     auto res_column = IColumn::mutate(prev);
@@ -356,13 +363,21 @@ ColumnPtr SerializationVariantElement::VariantSubcolumnCreator::create(const DB:
     if (make_nullable && prev->lowCardinality())
         res_column = assert_cast<ColumnLowCardinality &>(*res_column).cloneNullable();
 
-    res_column->expand(null_map, /*inverted = */ true);
+    if (null_map_from_discriminators)
+        res_column->expand(*null_map_from_discriminators, /*inverted = */ true);
+    else
+        res_column->expand(assert_cast<const ColumnUInt8 &>(*null_map).getData(), /*inverted = */ true);
 
     if (make_nullable && prev->canBeInsideNullable())
     {
-        auto null_map_col = ColumnUInt8::create();
-        null_map_col->getData() = std::move(null_map);
-        return ColumnNullable::create(std::move(res_column), std::move(null_map_col));
+        if (null_map_from_discriminators)
+        {
+            auto null_map_col = ColumnUInt8::create();
+            null_map_col->getData() = std::move(*null_map_from_discriminators);
+            return ColumnNullable::create(std::move(res_column), std::move(null_map_col));
+        }
+
+        return ColumnNullable::create(std::move(res_column), null_map->assumeMutable());
     }
 
     return res_column;

src/DataTypes/Serializations/SerializationVariantElement.h

@@ -63,18 +63,22 @@ public:
 
     struct VariantSubcolumnCreator : public ISubcolumnCreator
     {
+    private:
         const ColumnPtr local_discriminators;
+        const ColumnPtr null_map; /// optional
         const String variant_element_name;
         const ColumnVariant::Discriminator global_variant_discriminator;
         const ColumnVariant::Discriminator local_variant_discriminator;
         bool make_nullable;
 
+    public:
         VariantSubcolumnCreator(
             const ColumnPtr & local_discriminators_,
             const String & variant_element_name_,
             ColumnVariant::Discriminator global_variant_discriminator_,
             ColumnVariant::Discriminator local_variant_discriminator_,
-            bool make_nullable_);
+            bool make_nullable_,
+            const ColumnPtr & null_map_ = nullptr);
 
         DataTypePtr create(const DataTypePtr & prev) const override;
         ColumnPtr create(const ColumnPtr & prev) const override;

src/DataTypes/fuzzers/CMakeLists.txt

@@ -1,2 +1,2 @@
 clickhouse_add_executable(data_type_deserialization_fuzzer data_type_deserialization_fuzzer.cpp ${SRCS})
-target_link_libraries(data_type_deserialization_fuzzer PRIVATE dbms clickhouse_aggregate_functions clickhouse_functions)
+target_link_libraries(data_type_deserialization_fuzzer PRIVATE clickhouse_functions clickhouse_aggregate_functions)

src/Databases/DatabaseReplicated.cpp

@@ -1153,8 +1153,7 @@ void DatabaseReplicated::recoverLostReplica(const ZooKeeperPtr & current_zookeep
         query_context->setSetting("allow_experimental_object_type", 1);
         query_context->setSetting("allow_experimental_variant_type", 1);
         query_context->setSetting("allow_experimental_dynamic_type", 1);
-        query_context->setSetting("allow_experimental_annoy_index", 1);
-        query_context->setSetting("allow_experimental_usearch_index", 1);
+        query_context->setSetting("allow_experimental_vector_similarity_index", 1);
         query_context->setSetting("allow_experimental_bigint_types", 1);
         query_context->setSetting("allow_experimental_window_functions", 1);
         query_context->setSetting("allow_experimental_geo_types", 1);
@@ -1584,6 +1583,8 @@ void DatabaseReplicated::dropTable(ContextPtr local_context, const String & tabl
     }
 
     auto table = tryGetTable(table_name, getContext());
+    if (!table)
+        throw Exception(ErrorCodes::LOGICAL_ERROR, "Table {} doesn't exist", table_name);
     if (table->getName() == "MaterializedView" || table->getName() == "WindowView")
     {
         /// Avoid recursive locking of metadata_mutex