Merge branch 'master' of github.com:ClickHouse/ClickHouse into divanik/add_local_and_azure_iceberg_support

2024-11-22 23:52:03 +00:00 · 2024-08-12 09:14:39 +00:00 · 2024-08-12 09:14:39 +00:00 · 52de06514a
commit 52de06514a
parent bc5d793382 0aa30b10d5
167 changed files with 4568 additions and 1636 deletions
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -187,14 +187,6 @@ else ()
    set(NO_WHOLE_ARCHIVE --no-whole-archive)
 endif ()

-if (NOT CMAKE_BUILD_TYPE_UC STREQUAL "RELEASE")
-    # Can be lld or ld-lld or lld-13 or /path/to/lld.
-    if (LINKER_NAME MATCHES "lld")
-        set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--gdb-index")
-        message (STATUS "Adding .gdb-index via --gdb-index linker option.")
-    endif ()
-endif()
-
 if (NOT (SANITIZE_COVERAGE OR WITH_COVERAGE)
    AND (CMAKE_BUILD_TYPE_UC STREQUAL "RELEASE"
        OR CMAKE_BUILD_TYPE_UC STREQUAL "RELWITHDEBINFO"
@ -402,7 +394,7 @@ if ((NOT OS_LINUX AND NOT OS_ANDROID) OR (CMAKE_BUILD_TYPE_UC STREQUAL "DEBUG")
    set(ENABLE_GWP_ASAN OFF)
 endif ()

-option (ENABLE_FIU "Enable Fiu" ON)
+option (ENABLE_LIBFIU "Enable libfiu" ON)

 option(WERROR "Enable -Werror compiler option" ON)

--- a/contrib/CMakeLists.txt
+++ b/contrib/CMakeLists.txt
@ -179,7 +179,7 @@ else()
    message(STATUS "Not using QPL")
 endif ()

-if (OS_LINUX AND ARCH_AMD64)
+if (OS_LINUX AND ARCH_AMD64 AND NOT NO_SSE3_OR_HIGHER)
    option (ENABLE_QATLIB "Enable Intel® QuickAssist Technology Library (QATlib)" ${ENABLE_LIBRARIES})
 elseif(ENABLE_QATLIB)
    message (${RECONFIGURE_MESSAGE_LEVEL} "QATLib is only supported on x86_64")
--- a/contrib/QAT-ZSTD-Plugin-cmake/CMakeLists.txt
+++ b/contrib/QAT-ZSTD-Plugin-cmake/CMakeLists.txt
@ -27,7 +27,7 @@ if (ENABLE_QAT_OUT_OF_TREE_BUILD)
                ${QAT_AL_INCLUDE_DIR}
                ${QAT_USDM_INCLUDE_DIR}
                ${ZSTD_LIBRARY_DIR})
-    target_compile_definitions(_qatzstd_plugin PRIVATE -DDEBUGLEVEL=0 PUBLIC -DENABLE_ZSTD_QAT_CODEC)
+    target_compile_definitions(_qatzstd_plugin PRIVATE -DDEBUGLEVEL=0)
    add_library (ch_contrib::qatzstd_plugin ALIAS _qatzstd_plugin)
 else () # In-tree build
    message(STATUS "Intel QATZSTD in-tree build")
@ -78,7 +78,7 @@ else () # In-tree build
        ${QAT_USDM_INCLUDE_DIR}
        ${ZSTD_LIBRARY_DIR}
        ${LIBQAT_HEADER_DIR})
-    target_compile_definitions(_qatzstd_plugin PRIVATE -DDEBUGLEVEL=0 PUBLIC -DENABLE_ZSTD_QAT_CODEC -DINTREE)
+    target_compile_definitions(_qatzstd_plugin PRIVATE -DDEBUGLEVEL=0 PUBLIC -DINTREE)
    target_include_directories(_qatzstd_plugin SYSTEM PUBLIC $<BUILD_INTERFACE:${QATZSTD_SRC_DIR}> $<INSTALL_INTERFACE:include>)
    add_library (ch_contrib::qatzstd_plugin ALIAS _qatzstd_plugin)
 endif ()
--- a/contrib/libfiu-cmake/CMakeLists.txt
+++ b/contrib/libfiu-cmake/CMakeLists.txt
@ -1,20 +1,21 @@
-if (NOT ENABLE_FIU)
-  message (STATUS "Not using fiu")
+if (NOT ENABLE_LIBFIU)
+  message (STATUS "Not using libfiu")
  return ()
 endif ()

-set(FIU_DIR "${ClickHouse_SOURCE_DIR}/contrib/libfiu/")
+set(LIBFIU_DIR "${ClickHouse_SOURCE_DIR}/contrib/libfiu/")

-set(FIU_SOURCES
-  ${FIU_DIR}/libfiu/fiu.c
-  ${FIU_DIR}/libfiu/fiu-rc.c
-  ${FIU_DIR}/libfiu/backtrace.c
-  ${FIU_DIR}/libfiu/wtable.c
+set(LIBFIU_SOURCES
+    ${LIBFIU_DIR}/libfiu/fiu.c
+    ${LIBFIU_DIR}/libfiu/fiu-rc.c
+    ${LIBFIU_DIR}/libfiu/backtrace.c
+    ${LIBFIU_DIR}/libfiu/wtable.c
 )

-set(FIU_HEADERS "${FIU_DIR}/libfiu")
+set(LIBFIU_HEADERS "${LIBFIU_DIR}/libfiu")

-add_library(_fiu ${FIU_SOURCES})
-target_compile_definitions(_fiu PUBLIC DUMMY_BACKTRACE)
-target_include_directories(_fiu PUBLIC ${FIU_HEADERS})
-add_library(ch_contrib::fiu ALIAS _fiu)
+add_library(_libfiu ${LIBFIU_SOURCES})
+target_compile_definitions(_libfiu PUBLIC DUMMY_BACKTRACE)
+target_compile_definitions(_libfiu PUBLIC FIU_ENABLE)
+target_include_directories(_libfiu PUBLIC ${LIBFIU_HEADERS})
+add_library(ch_contrib::libfiu ALIAS _libfiu)
--- a/contrib/qpl-cmake/CMakeLists.txt
+++ b/contrib/qpl-cmake/CMakeLists.txt
@ -728,10 +728,6 @@ add_library(_qpl STATIC ${LIB_DEPS})
 target_include_directories(_qpl
        PUBLIC $<BUILD_INTERFACE:${QPL_PROJECT_DIR}/include/> $<INSTALL_INTERFACE:include>)

-
-target_compile_definitions(_qpl
-        PUBLIC -DENABLE_QPL_COMPRESSION)
-
 target_link_libraries(_qpl
        PRIVATE ch_contrib::accel-config)

--- a/docker/packager/cctools/Dockerfile
+++ b/docker/packager/cctools/Dockerfile
@ -1,3 +1,5 @@
+# docker build -t clickhouse/cctools .
+
 # This is a hack to significantly reduce the build time of the clickhouse/binary-builder
 # It's based on the assumption that we don't care of the cctools version so much
 # It event does not depend on the clickhouse/fasttest in the `docker/images.json`
@ -30,5 +32,29 @@ RUN git clone https://github.com/tpoechtrager/cctools-port.git \
    && cd ../.. \
    && rm -rf cctools-port

+#
+# GDB
+#
+# ld from binutils is 2.38, which has the following error:
+#
+#     DWARF error: invalid or unhandled FORM value: 0x23
+#
+ENV LD=ld.lld-${LLVM_VERSION}
+ARG GDB_VERSION=15.1
+RUN apt-get update \
+    && apt-get install --yes \
+        libgmp-dev \
+        libmpfr-dev \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*
+RUN wget https://sourceware.org/pub/gdb/releases/gdb-$GDB_VERSION.tar.gz \
+    && tar -xvf gdb-$GDB_VERSION.tar.gz \
+    && cd gdb-$GDB_VERSION \
+    && ./configure --prefix=/opt/gdb \
+    && make -j $(nproc) \
+    && make install \
+    && rm -fr gdb-$GDB_VERSION gdb-$GDB_VERSION.tar.gz
+
 FROM scratch
 COPY --from=builder /cctools /cctools
+COPY --from=builder /opt/gdb /opt/gdb
--- a/docker/test/fasttest/Dockerfile
+++ b/docker/test/fasttest/Dockerfile
@ -83,7 +83,7 @@ RUN arch=${TARGETARCH:-amd64} \

 # Give suid to gdb to grant it attach permissions
 # chmod 777 to make the container user independent
-RUN chmod u+s /usr/bin/gdb \
+RUN chmod u+s /opt/gdb/bin/gdb \
  && mkdir -p /var/lib/clickhouse \
  && chmod 777 /var/lib/clickhouse

--- a/docker/test/integration/base/Dockerfile
+++ b/docker/test/integration/base/Dockerfile
@ -11,7 +11,6 @@ RUN apt-get update \
        curl \
        default-jre \
        g++ \
-        gdb \
        iproute2 \
        krb5-user \
        libicu-dev \
@ -73,3 +72,6 @@ maxClientCnxns=80' > /opt/zookeeper/conf/zoo.cfg && \

 ENV TZ=Etc/UTC
 RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone
+
+COPY --from=clickhouse/cctools:0d6b90a7a490 /opt/gdb /opt/gdb
+ENV PATH="/opt/gdb/bin:${PATH}"
--- a/docker/test/integration/runner/Dockerfile
+++ b/docker/test/integration/runner/Dockerfile
@ -30,7 +30,6 @@ RUN apt-get update \
    luajit \
    libssl-dev \
    libcurl4-openssl-dev \
-    gdb \
    default-jdk \
    software-properties-common \
    libkrb5-dev \
@ -87,6 +86,8 @@ COPY modprobe.sh /usr/local/bin/modprobe
 COPY dockerd-entrypoint.sh /usr/local/bin/
 COPY misc/ /misc/

+COPY --from=clickhouse/cctools:0d6b90a7a490 /opt/gdb /opt/gdb
+ENV PATH="/opt/gdb/bin:${PATH}"

 # Same options as in test/base/Dockerfile
 # (in case you need to override them in tests)
--- a/docker/test/performance-comparison/Dockerfile
+++ b/docker/test/performance-comparison/Dockerfile
@ -9,7 +9,6 @@ RUN apt-get update \
            curl \
            dmidecode \
            g++ \
-            gdb \
            git \
            gnuplot \
            imagemagick \
@ -42,6 +41,9 @@ RUN pip3 --no-cache-dir install -r requirements.txt

 COPY run.sh /

+COPY --from=clickhouse/cctools:0d6b90a7a490 /opt/gdb /opt/gdb
+ENV PATH="/opt/gdb/bin:${PATH}"
+
 CMD ["bash", "/run.sh"]

 # docker run --network=host --volume <workspace>:/workspace --volume=<output>:/output -e PR_TO_TEST=<> -e SHA_TO_TEST=<> clickhouse/performance-comparison
--- a/docker/test/stateless/Dockerfile
+++ b/docker/test/stateless/Dockerfile
@ -69,8 +69,8 @@ ENV MAX_RUN_TIME=0

 # Unrelated to vars in setup_minio.sh, but should be the same there
 # to have the same binaries for local running scenario
-ARG MINIO_SERVER_VERSION=2022-01-03T18-22-58Z
-ARG MINIO_CLIENT_VERSION=2022-01-05T23-52-51Z
+ARG MINIO_SERVER_VERSION=2024-08-03T04-33-23Z
+ARG MINIO_CLIENT_VERSION=2024-07-31T15-58-33Z
 ARG TARGETARCH

 # Download Minio-related binaries
--- a/docker/test/stateless/setup_minio.sh
+++ b/docker/test/stateless/setup_minio.sh
@ -59,8 +59,8 @@ find_os() {
 download_minio() {
  local os
  local arch
-  local minio_server_version=${MINIO_SERVER_VERSION:-2022-09-07T22-25-02Z}
-  local minio_client_version=${MINIO_CLIENT_VERSION:-2022-08-28T20-08-11Z}
+  local minio_server_version=${MINIO_SERVER_VERSION:-2024-08-03T04-33-23Z}
+  local minio_client_version=${MINIO_CLIENT_VERSION:-2024-07-31T15-58-33Z}

  os=$(find_os)
  arch=$(find_arch)
@ -82,10 +82,10 @@ setup_minio() {
  local test_type=$1
  ./mc alias set clickminio http://localhost:11111 clickhouse clickhouse
  ./mc admin user add clickminio test testtest
-  ./mc admin policy set clickminio readwrite user=test
+  ./mc admin policy attach clickminio readwrite --user=test
  ./mc mb --ignore-existing clickminio/test
  if [ "$test_type" = "stateless" ]; then
-    ./mc policy set public clickminio/test
+    ./mc anonymous set public clickminio/test
  fi
 }

@ -148,4 +148,4 @@ main() {
  setup_aws_credentials
 }

-main "$@"
+main "$@"
--- a/docker/test/util/Dockerfile
+++ b/docker/test/util/Dockerfile
@ -44,7 +44,6 @@ RUN apt-get update \
        bash \
        bsdmainutils \
        build-essential \
-        gdb \
        git \
        gperf \
        moreutils \
@ -58,3 +57,6 @@ RUN apt-get update \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

 COPY process_functional_tests_result.py /
+
+COPY --from=clickhouse/cctools:0d6b90a7a490 /opt/gdb /opt/gdb
+ENV PATH="/opt/gdb/bin:${PATH}"
--- a/docs/en/interfaces/prometheus.md
+++ b/docs/en/interfaces/prometheus.md
@ -75,7 +75,7 @@ Data are received by this protocol and written to a [TimeSeries](/en/engines/tab
        <my_rule_1>
            <url>/write</url>
            <handler>
-                <type>remote_write</type
+                <type>remote_write</type>
                <database>db_name</database>
                <table>time_series_table</table>
            </handler>
@ -105,7 +105,7 @@ Data are read from a [TimeSeries](/en/engines/table-engines/special/time_series)
        <my_rule_1>
            <url>/read</url>
            <handler>
-                <type>remote_read</type
+                <type>remote_read</type>
                <database>db_name</database>
                <table>time_series_table</table>
            </handler>
@ -144,14 +144,14 @@ Multiple protocols can be specified together in one place:
        <my_rule_2>
            <url>/write</url>
            <handler>
-                <type>remote_write</type
+                <type>remote_write</type>
                <table>db_name.time_series_table</table>
            </handler>
        </my_rule_2>
        <my_rule_3>
            <url>/read</url>
            <handler>
-                <type>remote_read</type
+                <type>remote_read</type>
                <table>db_name.time_series_table</table>
            </handler>
        </my_rule_3>
--- a/docs/en/sql-reference/data-types/geo.md
+++ b/docs/en/sql-reference/data-types/geo.md
@ -52,6 +52,48 @@ Result:
 └───────────────────────────────┴───────────────┘
 ```

+## LineString
+
+`LineString` is a line stored as an array of points: [Array](array.md)([Point](#point)).
+
+**Example**
+
+Query:
+
+```sql
+CREATE TABLE geo_linestring (l LineString) ENGINE = Memory();
+INSERT INTO geo_linestring VALUES([(0, 0), (10, 0), (10, 10), (0, 10)]);
+SELECT l, toTypeName(l) FROM geo_linestring;
+```
+Result:
+
+``` text
+┌─r─────────────────────────────┬─toTypeName(r)─┐
+│ [(0,0),(10,0),(10,10),(0,10)] │ LineString    │
+└───────────────────────────────┴───────────────┘
+```
+
+## MultiLineString
+
+`MultiLineString` is multiple lines stored as an array of `LineString`: [Array](array.md)([LineString](#linestring)).
+
+**Example**
+
+Query:
+
+```sql
+CREATE TABLE geo_multilinestring (l MultiLineString) ENGINE = Memory();
+INSERT INTO geo_multilinestring VALUES([[(0, 0), (10, 0), (10, 10), (0, 10)], [(1, 1), (2, 2), (3, 3)]]);
+SELECT l, toTypeName(l) FROM geo_multilinestring;
+```
+Result:
+
+``` text
+┌─l───────────────────────────────────────────────────┬─toTypeName(l)───┐
+│ [[(0,0),(10,0),(10,10),(0,10)],[(1,1),(2,2),(3,3)]] │ MultiLineString │
+└─────────────────────────────────────────────────────┴─────────────────┘
+```
+
 ## Polygon

 `Polygon` is a polygon with holes stored as an array of rings: [Array](array.md)([Ring](#ring)). First element of outer array is the outer shape of polygon and all the following elements are holes.
--- a/docs/en/sql-reference/functions/geo/polygon.md
+++ b/docs/en/sql-reference/functions/geo/polygon.md
@ -6,11 +6,13 @@ title: "Functions for Working with Polygons"

 ## WKT

-Returns a WKT (Well Known Text) geometric object from various [Geo Data Types](../../data-types/geo.md). Supported WKT objects are: 
+Returns a WKT (Well Known Text) geometric object from various [Geo Data Types](../../data-types/geo.md). Supported WKT objects are:

 - POINT
 - POLYGON
 - MULTIPOLYGON
+- LINESTRING
+- MULTILINESTRING

 **Syntax**

@ -26,12 +28,16 @@ WKT(geo_data)
 - [Ring](../../data-types/geo.md#ring)
 - [Polygon](../../data-types/geo.md#polygon)
 - [MultiPolygon](../../data-types/geo.md#multipolygon)
+- [LineString](../../data-types/geo.md#linestring)
+- [MultiLineString](../../data-types/geo.md#multilinestring)

 **Returned value**

 - WKT geometric object `POINT` is returned for a Point.
 - WKT geometric object `POLYGON` is returned for a Polygon
- WKT geometric object `MULTIPOLYGON` is returned for a MultiPolygon. 
+- WKT geometric object `MULTIPOLYGON` is returned for a MultiPolygon.
+- WKT geometric object `LINESTRING` is returned for a LineString.
+- WKT geometric object `MULTILINESTRING` is returned for a MultiLineString.

 **Examples**

@ -84,7 +90,7 @@ SELECT

 ### Input parameters

-String starting with `MULTIPOLYGON` 
+String starting with `MULTIPOLYGON`

 ### Returned value

@ -170,6 +176,34 @@ SELECT readWKTLineString('LINESTRING (1 1, 2 2, 3 3, 1 1)');
 [(1,1),(2,2),(3,3),(1,1)]
 ```

+## readWKTMultiLineString
+
+Parses a Well-Known Text (WKT) representation of a MultiLineString geometry and returns it in the internal ClickHouse format.
+
+### Syntax
+
+```sql
+readWKTMultiLineString(wkt_string)
+```
+
+### Arguments
+
+- `wkt_string`: The input WKT string representing a MultiLineString geometry.
+
+### Returned value
+
+The function returns a ClickHouse internal representation of the multilinestring geometry.
+
+### Example
+
+```sql
+SELECT readWKTMultiLineString('MULTILINESTRING ((1 1, 2 2, 3 3), (4 4, 5 5, 6 6))');
+```
+
+```response
+[[(1,1),(2,2),(3,3)],[(4,4),(5,5),(6,6)]]
+```
+
 ## readWKTRing

 Parses a Well-Known Text (WKT) representation of a Polygon geometry and returns a ring (closed linestring) in the internal ClickHouse format.
@ -219,7 +253,7 @@ UInt8, 0 for false, 1 for true

 ## polygonsDistanceSpherical

-Calculates the minimal distance between two points where one point belongs to the first polygon and the second to another polygon. Spherical means that coordinates are interpreted as coordinates on a pure and ideal sphere, which is not true for the Earth. Using this type of coordinate system speeds up execution, but of course is not precise. 
+Calculates the minimal distance between two points where one point belongs to the first polygon and the second to another polygon. Spherical means that coordinates are interpreted as coordinates on a pure and ideal sphere, which is not true for the Earth. Using this type of coordinate system speeds up execution, but of course is not precise.

 ### Example

--- a/docs/en/sql-reference/functions/type-conversion-functions.md
+++ b/docs/en/sql-reference/functions/type-conversion-functions.md
--- a/docs/en/sql-reference/statements/select/join.md
+++ b/docs/en/sql-reference/statements/select/join.md
@ -186,7 +186,7 @@ Otherwise, you'll get `INVALID_JOIN_ON_EXPRESSION`.

 :::

-Clickhouse currently supports `ALL INNER/LEFT/RIGHT/FULL JOIN` with inequality conditions in addition to equality conditions. The inequality conditions are supported only for `hash` and `grace_hash` join algorithms. The inequality conditions are not supported with `join_use_nulls`.
+Clickhouse currently supports `ALL/ANY/SEMI/ANTI INNER/LEFT/RIGHT/FULL JOIN` with inequality conditions in addition to equality conditions. The inequality conditions are supported only for `hash` and `grace_hash` join algorithms. The inequality conditions are not supported with `join_use_nulls`.

 **Example**

--- a/programs/local/LocalServer.cpp
+++ b/programs/local/LocalServer.cpp
@ -143,7 +143,7 @@ void LocalServer::initialize(Poco::Util::Application & self)

    if (fs::exists(config_path))
    {
-        ConfigProcessor config_processor(config_path, false, true);
+        ConfigProcessor config_processor(config_path);
        ConfigProcessor::setConfigPath(fs::path(config_path).parent_path());
        auto loaded_config = config_processor.loadConfig();
        getClientConfiguration().add(loaded_config.configuration.duplicate(), PRIO_DEFAULT, false);
--- a/src/Backups/WithRetries.cpp
+++ b/src/Backups/WithRetries.cpp
@ -68,13 +68,19 @@ const WithRetries::KeeperSettings & WithRetries::getKeeperSettings() const

 WithRetries::FaultyKeeper WithRetries::getFaultyZooKeeper() const
 {
-    /// We need to create new instance of ZooKeeperWithFaultInjection each time a copy a pointer to ZooKeeper client there
+    zkutil::ZooKeeperPtr current_zookeeper;
+    {
+        std::lock_guard lock(zookeeper_mutex);
+        current_zookeeper = zookeeper;
+    }
+
+    /// We need to create new instance of ZooKeeperWithFaultInjection each time and copy a pointer to ZooKeeper client there
    /// The reason is that ZooKeeperWithFaultInjection may reset the underlying pointer and there could be a race condition
    /// when the same object is used from multiple threads.
    auto faulty_zookeeper = ZooKeeperWithFaultInjection::createInstance(
        settings.keeper_fault_injection_probability,
        settings.keeper_fault_injection_seed,
-        zookeeper,
+        current_zookeeper,
        log->name(),
        log);

--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@ -354,8 +354,8 @@ target_link_libraries(clickhouse_common_io
            Poco::Foundation
 )

-if (TARGET ch_contrib::fiu)
-    target_link_libraries(clickhouse_common_io PUBLIC ch_contrib::fiu)
+if (TARGET ch_contrib::libfiu)
+    target_link_libraries(clickhouse_common_io PUBLIC ch_contrib::libfiu)
 endif()

 if (TARGET ch_contrib::cpuid)
@ -557,14 +557,13 @@ target_link_libraries (clickhouse_common_io PRIVATE ch_contrib::lz4)

 if (TARGET ch_contrib::qpl)
    dbms_target_link_libraries(PUBLIC ch_contrib::qpl)
+    target_link_libraries (clickhouse_compression PUBLIC ch_contrib::qpl)
+    target_link_libraries (clickhouse_compression PUBLIC ch_contrib::accel-config)
 endif ()

-if (TARGET ch_contrib::accel-config)
-    dbms_target_link_libraries(PUBLIC ch_contrib::accel-config)
-endif ()
-
-if (TARGET ch_contrib::qatzstd_plugin)
+if (TARGET ch_contrib::accel-config AND TARGET ch_contrib::qatzstd_plugin)
    dbms_target_link_libraries(PUBLIC ch_contrib::qatzstd_plugin)
+    dbms_target_link_libraries(PUBLIC ch_contrib::accel-config)
    target_link_libraries(clickhouse_common_io PUBLIC ch_contrib::qatzstd_plugin)
 endif ()

--- a/src/Client/ClientApplicationBase.cpp
+++ b/src/Client/ClientApplicationBase.cpp
@ -200,8 +200,6 @@ void ClientApplicationBase::init(int argc, char ** argv)
        ("pager", po::value<std::string>(), "Pipe all output into this command (less or similar)")
        ("max_memory_usage_in_client", po::value<std::string>(), "Set memory limit in client/local server")

-        ("fuzzer-args", po::value<std::string>(), "Command line arguments for the LLVM's libFuzzer driver. Only relevant if the application is compiled with libFuzzer.")
-
        ("client_logs_file", po::value<std::string>(), "Path to a file for writing client logs. Currently we only have fatal logs (when the client crashes)")
    ;

--- a/src/Client/ClientBase.cpp
+++ b/src/Client/ClientBase.cpp
@ -477,6 +477,8 @@ void ClientBase::onProfileInfo(const ProfileInfo & profile_info)
 {
    if (profile_info.hasAppliedLimit() && output_format)
        output_format->setRowsBeforeLimit(profile_info.getRowsBeforeLimit());
+    if (profile_info.hasAppliedAggregation() && output_format)
+        output_format->setRowsBeforeAggregation(profile_info.getRowsBeforeAggregation());
 }


--- a/src/Client/Connection.cpp
+++ b/src/Client/Connection.cpp
@ -1319,7 +1319,7 @@ Progress Connection::receiveProgress() const
 ProfileInfo Connection::receiveProfileInfo() const
 {
    ProfileInfo profile_info;
-    profile_info.read(*in);
+    profile_info.read(*in, server_revision);
    return profile_info;
 }

--- a/src/Common/FailPoint.cpp
+++ b/src/Common/FailPoint.cpp
@ -7,6 +7,8 @@
 #include <condition_variable>
 #include <mutex>

+#include "config.h"
+
 namespace DB
 {

@ -15,7 +17,7 @@ namespace ErrorCodes
 extern const int LOGICAL_ERROR;
 };

-#if FIU_ENABLE
+#if USE_LIBFIU
 static struct InitFiu
 {
    InitFiu()
@ -135,7 +137,7 @@ void FailPointInjection::pauseFailPoint(const String & fail_point_name)

 void FailPointInjection::enableFailPoint(const String & fail_point_name)
 {
-#if FIU_ENABLE
+#if USE_LIBFIU
 #define SUB_M(NAME, flags, pause)                                                                               \
    if (fail_point_name == FailPoints::NAME)                                                                    \
    {                                                                                                           \
--- a/src/Common/FailPoint.h
+++ b/src/Common/FailPoint.h
@ -1,17 +1,16 @@
 #pragma once
-#include "config.h"

 #include <Common/Exception.h>
 #include <Core/Types.h>
 #include <Poco/Util/AbstractConfiguration.h>

+#include "config.h"
+
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wdocumentation"
 #pragma clang diagnostic ignored "-Wreserved-macro-identifier"
-
-#include <fiu.h>
-#include <fiu-control.h>
-
+#  include <fiu.h>
+#  include <fiu-control.h>
 #pragma clang diagnostic pop

 #include <unordered_map>
--- a/src/Common/ShellCommand.cpp
+++ b/src/Common/ShellCommand.cpp
@ -237,7 +237,14 @@ std::unique_ptr<ShellCommand> ShellCommand::executeImpl(
        res->write_fds.emplace(fd, fds.fds_rw[1]);
    }

-    LOG_TRACE(getLogger(), "Started shell command '{}' with pid {}", filename, pid);
+    LOG_TRACE(
+        getLogger(),
+        "Started shell command '{}' with pid {} and file descriptors: out {}, err {}",
+        filename,
+        pid,
+        res->out.getFD(),
+        res->err.getFD());
+
    return res;
 }

--- a/src/Common/config.h.in
+++ b/src/Common/config.h.in
@ -32,6 +32,8 @@
 #cmakedefine01 USE_IDNA
 #cmakedefine01 USE_NLP
 #cmakedefine01 USE_VECTORSCAN
+#cmakedefine01 USE_QPL
+#cmakedefine01 USE_QATLIB
 #cmakedefine01 USE_LIBURING
 #cmakedefine01 USE_AVRO
 #cmakedefine01 USE_CAPNP
@ -59,7 +61,7 @@
 #cmakedefine01 USE_SKIM
 #cmakedefine01 USE_PRQL
 #cmakedefine01 USE_ULID
-#cmakedefine01 FIU_ENABLE
+#cmakedefine01 USE_LIBFIU
 #cmakedefine01 USE_BCRYPT
 #cmakedefine01 USE_LIBARCHIVE
 #cmakedefine01 USE_POCKETFFT
--- a/src/Compression/CompressionCodecDeflateQpl.cpp
+++ b/src/Compression/CompressionCodecDeflateQpl.cpp
@ -1,7 +1,3 @@
-#ifdef ENABLE_QPL_COMPRESSION
-
-#include <cstdio>
-#include <thread>
 #include <Compression/CompressionCodecDeflateQpl.h>
 #include <Compression/CompressionFactory.h>
 #include <Compression/CompressionInfo.h>
@ -11,6 +7,10 @@
 #include <Common/randomSeed.h>
 #include <base/scope_guard.h>
 #include <base/getPageSize.h>
+#include <cstdio>
+#include <thread>
+
+#if USE_QPL

 #include "libaccel_config.h"

--- a/src/Compression/CompressionCodecDeflateQpl.h
+++ b/src/Compression/CompressionCodecDeflateQpl.h
@ -4,6 +4,11 @@
 #include <map>
 #include <random>
 #include <pcg_random.hpp>
+
+#include "config.h"
+
+#if USE_QPL
+
 #include <qpl/qpl.h>

 namespace Poco
@ -117,3 +122,4 @@ private:
 };

 }
+#endif
--- a/src/Compression/CompressionCodecZSTDQAT.cpp
+++ b/src/Compression/CompressionCodecZSTDQAT.cpp
@ -1,4 +1,6 @@
-#ifdef ENABLE_ZSTD_QAT_CODEC
+#include "config.h"
+
+#if USE_QATLIB

 #include <Common/logger_useful.h>
 #include <Compression/CompressionCodecZSTD.h>
@ -6,6 +8,7 @@
 #include <Parsers/ASTLiteral.h>
 #include <Parsers/IAST.h>
 #include <Poco/Logger.h>
+
 #include <qatseqprod.h>
 #include <zstd.h>

--- a/src/Compression/CompressionFactory.cpp
+++ b/src/Compression/CompressionFactory.cpp
@ -1,20 +1,20 @@
-#include "config.h"
-
 #include <Compression/CompressionFactory.h>
+#include <Compression/CompressionCodecMultiple.h>
+#include <Compression/CompressionCodecNone.h>
+#include <IO/ReadBuffer.h>
+#include <IO/WriteHelpers.h>
 #include <Parsers/ASTFunction.h>
 #include <Parsers/ASTIdentifier.h>
 #include <Parsers/ASTLiteral.h>
-#include <Poco/String.h>
-#include <IO/ReadBuffer.h>
-#include <Parsers/queryToString.h>
-#include <Parsers/parseQuery.h>
 #include <Parsers/ExpressionElementParsers.h>
-#include <Compression/CompressionCodecMultiple.h>
-#include <Compression/CompressionCodecNone.h>
-#include <IO/WriteHelpers.h>
+#include <Parsers/parseQuery.h>
+#include <Parsers/queryToString.h>
+#include <Poco/String.h>

 #include <boost/algorithm/string/join.hpp>

+#include "config.h"
+
 namespace DB
 {

@ -175,11 +175,11 @@ void registerCodecNone(CompressionCodecFactory & factory);
 void registerCodecLZ4(CompressionCodecFactory & factory);
 void registerCodecLZ4HC(CompressionCodecFactory & factory);
 void registerCodecZSTD(CompressionCodecFactory & factory);
-#ifdef ENABLE_ZSTD_QAT_CODEC
+#if USE_QATLIB
 void registerCodecZSTDQAT(CompressionCodecFactory & factory);
 #endif
 void registerCodecMultiple(CompressionCodecFactory & factory);
-#ifdef ENABLE_QPL_COMPRESSION
+#if USE_QPL
 void registerCodecDeflateQpl(CompressionCodecFactory & factory);
 #endif

@ -198,7 +198,7 @@ CompressionCodecFactory::CompressionCodecFactory()
    registerCodecNone(*this);
    registerCodecLZ4(*this);
    registerCodecZSTD(*this);
-#ifdef ENABLE_ZSTD_QAT_CODEC
+#if USE_QATLIB
    registerCodecZSTDQAT(*this);
 #endif
    registerCodecLZ4HC(*this);
@ -209,7 +209,7 @@ CompressionCodecFactory::CompressionCodecFactory()
    registerCodecGorilla(*this);
    registerCodecEncrypted(*this);
    registerCodecFPC(*this);
-#ifdef ENABLE_QPL_COMPRESSION
+#if USE_QPL
    registerCodecDeflateQpl(*this);
 #endif
    registerCodecGCD(*this);
--- a/src/Core/ProtocolDefines.h
+++ b/src/Core/ProtocolDefines.h
@ -81,6 +81,8 @@ static constexpr auto DBMS_MIN_REVISION_WITH_TABLE_READ_ONLY_CHECK = 54467;

 static constexpr auto DBMS_MIN_REVISION_WITH_SYSTEM_KEYWORDS_TABLE = 54468;

+static constexpr auto DBMS_MIN_REVISION_WITH_ROWS_BEFORE_AGGREGATION = 54469;
+
 /// Version of ClickHouse TCP protocol.
 ///
 /// Should be incremented manually on protocol changes.
@ -88,6 +90,6 @@ static constexpr auto DBMS_MIN_REVISION_WITH_SYSTEM_KEYWORDS_TABLE = 54468;
 /// 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 = 54468;
+static constexpr auto DBMS_TCP_PROTOCOL_VERSION = 54469;

 }
--- a/src/Core/Settings.h
+++ b/src/Core/Settings.h
@ -606,7 +606,7 @@ class IColumn;
    M(Bool, optimize_if_chain_to_multiif, false, "Replace if(cond1, then1, if(cond2, ...)) chains to multiIf. Currently it's not beneficial for numeric types.", 0) \
    M(Bool, optimize_multiif_to_if, true, "Replace 'multiIf' with only one condition to 'if'.", 0) \
    M(Bool, optimize_if_transform_strings_to_enum, false, "Replaces string-type arguments in If and Transform to enum. Disabled by default cause it could make inconsistent change in distributed query that would lead to its fail.", 0) \
-    M(Bool, optimize_functions_to_subcolumns, false, "Transform functions to subcolumns, if possible, to reduce amount of read data. E.g. 'length(arr)' -> 'arr.size0', 'col IS NULL' -> 'col.null' ", 0) \
+    M(Bool, optimize_functions_to_subcolumns, true, "Transform functions to subcolumns, if possible, to reduce amount of read data. E.g. 'length(arr)' -> 'arr.size0', 'col IS NULL' -> 'col.null' ", 0) \
    M(Bool, optimize_using_constraints, false, "Use constraints for query optimization", 0)                                                                                                                                           \
    M(Bool, optimize_substitute_columns, false, "Use constraints for column substitution", 0)                                                                                                                                         \
    M(Bool, optimize_append_index, false, "Use constraints in order to append index condition (indexHint)", 0) \
@ -1244,6 +1244,7 @@ class IColumn;
    M(Bool, insert_distributed_one_random_shard, false, "If setting is enabled, inserting into distributed table will choose a random shard to write when there is no sharding key", 0) \
    \
    M(Bool, exact_rows_before_limit, false, "When enabled, ClickHouse will provide exact value for rows_before_limit_at_least statistic, but with the cost that the data before limit will have to be read completely", 0) \
+    M(Bool, rows_before_aggregation, false, "When enabled, ClickHouse will provide exact value for rows_before_aggregation statistic, represents the number of rows read before aggregation", 0) \
    M(UInt64, cross_to_inner_join_rewrite, 1, "Use inner join instead of comma/cross join if there are joining expressions in the WHERE section. Values: 0 - no rewrite, 1 - apply if possible for comma/cross, 2 - force rewrite all comma joins, cross - if possible", 0) \
    \
    M(Bool, output_format_arrow_low_cardinality_as_dictionary, false, "Enable output LowCardinality type as Dictionary Arrow type", 0) \
--- a/src/Core/SettingsChangesHistory.cpp
+++ b/src/Core/SettingsChangesHistory.cpp
@ -335,6 +335,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
    },
    {"24.8",
        {
+            {"rows_before_aggregation", true, true, "Provide exact value for rows_before_aggregation statistic, represents the number of rows read before aggregation"},
            {"restore_replace_external_table_functions_to_null", false, false, "New setting."},
            {"restore_replace_external_engines_to_null", false, false, "New setting."},
            {"input_format_json_max_depth", 1000000, 1000, "It was unlimited in previous versions, but that was unsafe."},
@ -343,6 +344,7 @@ static std::initializer_list<std::pair<ClickHouseVersion, SettingsChangesHistory
            {"allow_archive_path_syntax", true, true, "Added new setting to allow disabling archive path syntax."},
            {"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"},
        }
    },
    {"24.7",
--- a/src/DataTypes/DataTypeCustomGeo.cpp
+++ b/src/DataTypes/DataTypeCustomGeo.cpp
@ -24,6 +24,13 @@ void registerDataTypeDomainGeo(DataTypeFactory & factory)
            std::make_unique<DataTypeCustomDesc>(std::make_unique<DataTypeLineStringName>()));
    });

+    // Custom type for multiple lines stored as Array(LineString)
+    factory.registerSimpleDataTypeCustom("MultiLineString", []
+    {
+        return std::make_pair(DataTypeFactory::instance().get("Array(LineString)"),
+            std::make_unique<DataTypeCustomDesc>(std::make_unique<DataTypeMultiLineStringName>()));
+    });
+
    // Custom type for simple polygon without holes stored as Array(Point)
    factory.registerSimpleDataTypeCustom("Ring", []
    {
--- a/src/DataTypes/DataTypeCustomGeo.h
+++ b/src/DataTypes/DataTypeCustomGeo.h
@ -17,6 +17,12 @@ public:
    DataTypeLineStringName() : DataTypeCustomFixedName("LineString") {}
 };

+class DataTypeMultiLineStringName : public DataTypeCustomFixedName
+{
+public:
+    DataTypeMultiLineStringName() : DataTypeCustomFixedName("MultiLineString") {}
+};
+
 class DataTypeRingName : public DataTypeCustomFixedName
 {
 public:
--- a/src/Disks/DiskFomAST.cpp
+++ b/src/Disks/DiskFomAST.cpp
@ -0,0 +1,150 @@
+#include <Disks/DiskFomAST.h>
+#include <Common/assert_cast.h>
+#include <Common/filesystemHelpers.h>
+#include <Disks/getDiskConfigurationFromAST.h>
+#include <Disks/DiskSelector.h>
+#include <Parsers/formatAST.h>
+#include <Parsers/ASTExpressionList.h>
+#include <Parsers/ASTLiteral.h>
+#include <Parsers/ASTIdentifier.h>
+#include <Parsers/ASTFunction.h>
+#include <Parsers/isDiskFunction.h>
+#include <Interpreters/Context.h>
+#include <Parsers/IAST.h>
+#include <Interpreters/InDepthNodeVisitor.h>
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int BAD_ARGUMENTS;
+}
+
+std::string getOrCreateCustomDisk(DiskConfigurationPtr config, const std::string & serialization, ContextPtr context, bool attach)
+{
+    Poco::Util::AbstractConfiguration::Keys disk_settings_keys;
+    config->keys(disk_settings_keys);
+    /// Check that no settings are defined when disk from the config is referred.
+    if (disk_settings_keys.empty())
+        throw Exception(
+            ErrorCodes::BAD_ARGUMENTS,
+            "Disk function must have arguments. Invalid disk description.");
+
+    if (disk_settings_keys.size() == 1 && disk_settings_keys.front() == "name" && !attach)
+        throw Exception(
+            ErrorCodes::BAD_ARGUMENTS,
+            "Disk function `{}` must have other arguments apart from `name`, which describe disk configuration. Invalid disk description.",
+            serialization);
+
+    auto disk_settings_hash = sipHash128(serialization.data(), serialization.size());
+
+    std::string disk_name;
+    if (config->has("name"))
+    {
+        disk_name = config->getString("name");
+    }
+    else
+    {
+        /// We need a unique name for a created custom disk, but it needs to be the same
+        /// after table is reattached or server is restarted, so take a hash of the disk
+        /// configuration serialized ast as a disk name suffix.
+        disk_name = DiskSelector::TMP_INTERNAL_DISK_PREFIX + toString(disk_settings_hash);
+    }
+
+
+    auto disk = context->getOrCreateDisk(disk_name, [&](const DisksMap & disks_map) -> DiskPtr {
+        auto result = DiskFactory::instance().create(
+            disk_name, *config, /* config_path */"", context, disks_map, /* attach */attach, /* custom_disk */true);
+        /// Mark that disk can be used without storage policy.
+        result->markDiskAsCustom(disk_settings_hash);
+        return result;
+    });
+
+    if (!disk->isCustomDisk())
+        throw Exception(ErrorCodes::BAD_ARGUMENTS,
+            "Disk `{}` already exists and is described by the config."
+            " It is impossible to redefine it.",
+            disk_name);
+
+    if (disk->getCustomDiskSettings() != disk_settings_hash && !attach)
+        throw Exception(
+                ErrorCodes::BAD_ARGUMENTS,
+                "The disk `{}` is already configured as a custom disk in another table. It can't be redefined with different settings.",
+                disk_name);
+
+    if (!attach && !disk->isRemote())
+    {
+        static constexpr auto custom_local_disks_base_dir_in_config = "custom_local_disks_base_directory";
+        auto disk_path_expected_prefix = context->getConfigRef().getString(custom_local_disks_base_dir_in_config, "");
+
+        if (disk_path_expected_prefix.empty())
+            throw Exception(
+                ErrorCodes::BAD_ARGUMENTS,
+                "Base path for custom local disks must be defined in config file by `{}`",
+                custom_local_disks_base_dir_in_config);
+
+        if (!pathStartsWith(disk->getPath(), disk_path_expected_prefix))
+            throw Exception(
+                ErrorCodes::BAD_ARGUMENTS,
+                "Path of the custom local disk must be inside `{}` directory",
+                disk_path_expected_prefix);
+    }
+
+    return disk_name;
+}
+
+class DiskConfigurationFlattener
+{
+public:
+    struct Data
+    {
+        ContextPtr context;
+        bool attach;
+    };
+
+    static bool needChildVisit(const ASTPtr &, const ASTPtr &) { return true; }
+
+    static void visit(ASTPtr & ast, Data & data)
+    {
+        if (isDiskFunction(ast))
+        {
+            const auto * function = ast->as<ASTFunction>();
+            const auto * function_args_expr = assert_cast<const ASTExpressionList *>(function->arguments.get());
+            const auto & function_args = function_args_expr->children;
+            auto config = getDiskConfigurationFromAST(function_args, data.context);
+            auto disk_setting_string = serializeAST(*function);
+            auto disk_name = getOrCreateCustomDisk(config, disk_setting_string, data.context, data.attach);
+            ast = std::make_shared<ASTLiteral>(disk_name);
+        }
+    }
+};
+
+
+std::string DiskFomAST::createCustomDisk(const ASTPtr & disk_function_ast, ContextPtr context, bool attach)
+{
+    if (!isDiskFunction(disk_function_ast))
+        throw Exception(ErrorCodes::BAD_ARGUMENTS, "Expected a disk function");
+
+    auto ast = disk_function_ast->clone();
+
+    using FlattenDiskConfigurationVisitor = InDepthNodeVisitor<DiskConfigurationFlattener, false>;
+    FlattenDiskConfigurationVisitor::Data data{context, attach};
+    FlattenDiskConfigurationVisitor{data}.visit(ast);
+
+    return assert_cast<const ASTLiteral &>(*ast).value.get<String>();
+}
+
+void DiskFomAST::ensureDiskIsNotCustom(const std::string & disk_name, ContextPtr context)
+{
+    auto disk = context->getDisk(disk_name);
+
+    if (disk->isCustomDisk())
+        throw Exception(
+            ErrorCodes::BAD_ARGUMENTS,
+            "Disk name `{}` is a custom disk that is used in other table. "
+            "That disk could not be used by a reference by other tables. The custom disk should be fully specified with a disk function.",
+            disk_name);
+}
+
+}
--- a/src/Disks/DiskFomAST.h
+++ b/src/Disks/DiskFomAST.h
@ -0,0 +1,15 @@
+#pragma once
+#include <string>
+#include <Interpreters/Context_fwd.h>
+#include <Parsers/IAST_fwd.h>
+
+namespace DB
+{
+
+namespace DiskFomAST
+{
+    void ensureDiskIsNotCustom(const std::string & name, ContextPtr context);
+    std::string createCustomDisk(const ASTPtr & disk_function, ContextPtr context, bool attach);
+}
+
+}
--- a/src/Disks/DiskSelector.h
+++ b/src/Disks/DiskSelector.h
@ -6,6 +6,8 @@
 #include <Poco/Util/AbstractConfiguration.h>

 #include <map>
+#include <sstream>
+#include <string_view>

 namespace DB
 {
--- a/src/Disks/IDisk.h
+++ b/src/Disks/IDisk.h
@ -464,9 +464,9 @@ public:
    virtual void chmod(const String & /*path*/, mode_t /*mode*/) { throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Disk does not support chmod"); }

    /// Was disk created to be used without storage configuration?
-    bool isCustomDisk() const { return is_custom_disk; }
-
-    void markDiskAsCustom() { is_custom_disk = true; }
+    bool isCustomDisk() const { return custom_disk_settings_hash != 0; }
+    UInt128 getCustomDiskSettings() const { return custom_disk_settings_hash; }
+    void markDiskAsCustom(UInt128 settings_hash) { custom_disk_settings_hash = settings_hash; }

    virtual DiskPtr getDelegateDiskIfExists() const { return nullptr; }

@ -504,7 +504,8 @@ protected:

 private:
    ThreadPool copying_thread_pool;
-    bool is_custom_disk = false;
+    // 0 means the disk is not custom, the disk is predefined in the config
+    UInt128 custom_disk_settings_hash = 0;

    /// Check access to the disk.
    void checkAccess();
--- a/src/Disks/ObjectStorages/S3/S3ObjectStorage.cpp
+++ b/src/Disks/ObjectStorages/S3/S3ObjectStorage.cpp
@ -305,7 +305,8 @@ void S3ObjectStorage::listObjects(const std::string & path, RelativePathsWithMet

    S3::ListObjectsV2Request request;
    request.SetBucket(uri.bucket);
-    request.SetPrefix(path);
+    if (path != "/")
+        request.SetPrefix(path);
    if (max_keys)
        request.SetMaxKeys(static_cast<int>(max_keys));
    else
--- a/src/Disks/StoragePolicy.h
+++ b/src/Disks/StoragePolicy.h
@ -12,7 +12,6 @@
 #include <Common/formatReadable.h>

 #include <memory>
-#include <mutex>
 #include <unordered_map>
 #include <unistd.h>
 #include <boost/noncopyable.hpp>
--- a/src/Disks/getOrCreateDiskFromAST.cpp
+++ b/src/Disks/getOrCreateDiskFromAST.cpp
@ -1,121 +0,0 @@
-#include <Disks/getOrCreateDiskFromAST.h>
-#include <Common/logger_useful.h>
-#include <Common/assert_cast.h>
-#include <Common/filesystemHelpers.h>
-#include <Disks/getDiskConfigurationFromAST.h>
-#include <Disks/DiskSelector.h>
-#include <Parsers/formatAST.h>
-#include <Parsers/ASTExpressionList.h>
-#include <Parsers/ASTLiteral.h>
-#include <Parsers/ASTIdentifier.h>
-#include <Parsers/ASTFunction.h>
-#include <Parsers/isDiskFunction.h>
-#include <Interpreters/Context.h>
-#include <Parsers/IAST.h>
-#include <Interpreters/InDepthNodeVisitor.h>
-
-namespace DB
-{
-
-namespace ErrorCodes
-{
-    extern const int BAD_ARGUMENTS;
-}
-
-namespace
-{
-    std::string getOrCreateDiskFromDiskAST(const ASTFunction & function, ContextPtr context, bool attach)
-    {
-        const auto * function_args_expr = assert_cast<const ASTExpressionList *>(function.arguments.get());
-        const auto & function_args = function_args_expr->children;
-        auto config = getDiskConfigurationFromAST(function_args, context);
-
-        std::string disk_name;
-        if (config->has("name"))
-        {
-            disk_name = config->getString("name");
-        }
-        else
-        {
-            /// We need a unique name for a created custom disk, but it needs to be the same
-            /// after table is reattached or server is restarted, so take a hash of the disk
-            /// configuration serialized ast as a disk name suffix.
-            auto disk_setting_string = serializeAST(function);
-            disk_name = DiskSelector::TMP_INTERNAL_DISK_PREFIX
-                + toString(sipHash128(disk_setting_string.data(), disk_setting_string.size()));
-        }
-
-        auto result_disk = context->getOrCreateDisk(disk_name, [&](const DisksMap & disks_map) -> DiskPtr {
-            auto disk = DiskFactory::instance().create(
-                disk_name, *config, /* config_path */"", context, disks_map, /* attach */attach, /* custom_disk */true);
-            /// Mark that disk can be used without storage policy.
-            disk->markDiskAsCustom();
-            return disk;
-        });
-
-        if (!result_disk->isCustomDisk())
-            throw Exception(ErrorCodes::BAD_ARGUMENTS, "Disk with name `{}` already exist", disk_name);
-
-        if (!attach && !result_disk->isRemote())
-        {
-            static constexpr auto custom_local_disks_base_dir_in_config = "custom_local_disks_base_directory";
-            auto disk_path_expected_prefix = context->getConfigRef().getString(custom_local_disks_base_dir_in_config, "");
-
-            if (disk_path_expected_prefix.empty())
-                throw Exception(
-                    ErrorCodes::BAD_ARGUMENTS,
-                    "Base path for custom local disks must be defined in config file by `{}`",
-                    custom_local_disks_base_dir_in_config);
-
-            if (!pathStartsWith(result_disk->getPath(), disk_path_expected_prefix))
-                throw Exception(
-                    ErrorCodes::BAD_ARGUMENTS,
-                    "Path of the custom local disk must be inside `{}` directory",
-                    disk_path_expected_prefix);
-        }
-
-        return disk_name;
-    }
-
-    class DiskConfigurationFlattener
-    {
-    public:
-        struct Data
-        {
-            ContextPtr context;
-            bool attach;
-        };
-
-        static bool needChildVisit(const ASTPtr &, const ASTPtr &) { return true; }
-
-        static void visit(ASTPtr & ast, Data & data)
-        {
-            if (isDiskFunction(ast))
-            {
-                auto disk_name = getOrCreateDiskFromDiskAST(*ast->as<ASTFunction>(), data.context, data.attach);
-                ast = std::make_shared<ASTLiteral>(disk_name);
-            }
-        }
-    };
-
-    /// Visits children first.
-    using FlattenDiskConfigurationVisitor = InDepthNodeVisitor<DiskConfigurationFlattener, false>;
-}
-
-
-std::string getOrCreateDiskFromDiskAST(const ASTPtr & disk_function, ContextPtr context, bool attach)
-{
-    if (!isDiskFunction(disk_function))
-        throw Exception(ErrorCodes::BAD_ARGUMENTS, "Expected a disk function");
-
-    auto ast = disk_function->clone();
-
-    FlattenDiskConfigurationVisitor::Data data{context, attach};
-    FlattenDiskConfigurationVisitor{data}.visit(ast);
-
-    auto disk_name = assert_cast<const ASTLiteral &>(*ast).value.get<String>();
-    LOG_TRACE(getLogger("getOrCreateDiskFromDiskAST"), "Result disk name: {}", disk_name);
-    return disk_name;
-}
-
-}
--- a/src/Disks/getOrCreateDiskFromAST.h
+++ b/src/Disks/getOrCreateDiskFromAST.h
@ -1,18 +0,0 @@
-#pragma once
-#include <string>
-#include <Interpreters/Context_fwd.h>
-#include <Parsers/IAST_fwd.h>
-
-namespace DB
-{
-
-class ASTFunction;
-
-/**
- * Create a DiskPtr from disk AST function like disk(<disk_configuration>),
- * add it to DiskSelector by a unique (but always the same for given configuration) disk name
- * and return this name.
- */
-std::string getOrCreateDiskFromDiskAST(const ASTPtr & disk_function, ContextPtr context, bool attach);
-
-}
--- a/src/Formats/JSONUtils.cpp
+++ b/src/Formats/JSONUtils.cpp
@ -487,6 +487,8 @@ namespace JSONUtils
        size_t rows,
        size_t rows_before_limit,
        bool applied_limit,
+        size_t rows_before_aggregation,
+        bool applied_aggregation,
        const Stopwatch & watch,
        const Progress & progress,
        bool write_statistics,
@ -502,7 +504,12 @@ namespace JSONUtils
            writeTitle("rows_before_limit_at_least", out, 1, " ");
            writeIntText(rows_before_limit, out);
        }
-
+        if (applied_aggregation)
+        {
+            writeFieldDelimiter(out, 2);
+            writeTitle("rows_before_aggregation", out, 1, " ");
+            writeIntText(rows_before_aggregation, out);
+        }
        if (write_statistics)
        {
            writeFieldDelimiter(out, 2);
--- a/src/Formats/JSONUtils.h
+++ b/src/Formats/JSONUtils.h
@ -104,6 +104,8 @@ namespace JSONUtils
        size_t rows,
        size_t rows_before_limit,
        bool applied_limit,
+        size_t rows_before_aggregation,
+        bool applied_aggregation,
        const Stopwatch & watch,
        const Progress & progress,
        bool write_statistics,
--- a/src/Functions/CastOverloadResolver.cpp
+++ b/src/Functions/CastOverloadResolver.cpp
@ -3,6 +3,7 @@
 #include <Functions/FunctionHelpers.h>
 #include <DataTypes/DataTypeFactory.h>
 #include <DataTypes/DataTypeNullable.h>
+#include <DataTypes/DataTypeString.h>
 #include <Columns/ColumnString.h>
 #include <Core/Settings.h>
 #include <Interpreters/parseColumnsListForTableFunction.h>
@ -35,7 +36,7 @@ FunctionBasePtr createFunctionBaseCast(
 class CastOverloadResolverImpl : public IFunctionOverloadResolver
 {
 public:
-    const char * getNameImpl() const
+    static const char * getNameImpl(CastType cast_type, bool internal)
    {
        if (cast_type == CastType::accurate)
            return "accurateCast";
@ -49,7 +50,7 @@ public:

    String getName() const override
    {
-        return getNameImpl();
+        return getNameImpl(cast_type, internal);
    }

    size_t getNumberOfArguments() const override { return 2; }
@ -79,10 +80,22 @@ public:
        }
    }

+    static FunctionBasePtr createInternalCast(ColumnWithTypeAndName from, DataTypePtr to, CastType cast_type, std::optional<CastDiagnostic> diagnostic)
+    {
+        if (cast_type == CastType::accurateOrNull && !isVariant(to))
+            to = makeNullable(to);
+
+        ColumnsWithTypeAndName arguments;
+        arguments.emplace_back(std::move(from));
+        arguments.emplace_back().type = std::make_unique<DataTypeString>();
+
+        return createFunctionBaseCast(nullptr, getNameImpl(cast_type, true), arguments, to, diagnostic, cast_type);
+    }
+
 protected:
    FunctionBasePtr buildImpl(const ColumnsWithTypeAndName & arguments, const DataTypePtr & return_type) const override
    {
-        return createFunctionBaseCast(context, getNameImpl(), arguments, return_type, diagnostic, cast_type);
+        return createFunctionBaseCast(context, getNameImpl(cast_type, internal), arguments, return_type, diagnostic, cast_type);
    }

    DataTypePtr getReturnTypeImpl(const ColumnsWithTypeAndName & arguments) const override
@ -130,9 +143,9 @@ private:
 };


-FunctionOverloadResolverPtr createInternalCastOverloadResolver(CastType type, std::optional<CastDiagnostic> diagnostic)
+FunctionBasePtr createInternalCast(ColumnWithTypeAndName from, DataTypePtr to, CastType cast_type, std::optional<CastDiagnostic> diagnostic)
 {
-    return CastOverloadResolverImpl::create(ContextPtr{}, type, true, diagnostic);
+    return CastOverloadResolverImpl::createInternalCast(std::move(from), std::move(to), cast_type, std::move(diagnostic));
 }

 REGISTER_FUNCTION(CastOverloadResolvers)
--- a/src/Functions/CastOverloadResolver.h
+++ b/src/Functions/CastOverloadResolver.h
@ -3,6 +3,7 @@
 #include <memory>
 #include <optional>
 #include <Interpreters/Context_fwd.h>
+#include <Core/ColumnWithTypeAndName.h>


 namespace DB
@ -11,6 +12,9 @@ namespace DB
 class IFunctionOverloadResolver;
 using FunctionOverloadResolverPtr = std::shared_ptr<IFunctionOverloadResolver>;

+class IFunctionBase;
+using FunctionBasePtr = std::shared_ptr<const IFunctionBase>;
+
 enum class CastType : uint8_t
 {
    nonAccurate,
@ -24,6 +28,6 @@ struct CastDiagnostic
    std::string column_to;
 };

-FunctionOverloadResolverPtr createInternalCastOverloadResolver(CastType type, std::optional<CastDiagnostic> diagnostic);
+FunctionBasePtr createInternalCast(ColumnWithTypeAndName from, DataTypePtr to, CastType cast_type, std::optional<CastDiagnostic> diagnostic);

 }
--- a/src/Functions/geometryConverters.h
+++ b/src/Functions/geometryConverters.h
@ -31,6 +31,9 @@ namespace ErrorCodes
 template <typename Point>
 using LineString = boost::geometry::model::linestring<Point>;

+template <typename Point>
+using MultiLineString = boost::geometry::model::multi_linestring<LineString<Point>>;
+
 template <typename Point>
 using Ring = boost::geometry::model::ring<Point>;

@ -42,12 +45,14 @@ using MultiPolygon = boost::geometry::model::multi_polygon<Polygon<Point>>;

 using CartesianPoint = boost::geometry::model::d2::point_xy<Float64>;
 using CartesianLineString = LineString<CartesianPoint>;
+using CartesianMultiLineString = MultiLineString<CartesianPoint>;
 using CartesianRing = Ring<CartesianPoint>;
 using CartesianPolygon = Polygon<CartesianPoint>;
 using CartesianMultiPolygon = MultiPolygon<CartesianPoint>;

 using SphericalPoint = boost::geometry::model::point<Float64, 2, boost::geometry::cs::spherical_equatorial<boost::geometry::degree>>;
 using SphericalLineString = LineString<SphericalPoint>;
+using SphericalMultiLineString = MultiLineString<SphericalPoint>;
 using SphericalRing = Ring<SphericalPoint>;
 using SphericalPolygon = Polygon<SphericalPoint>;
 using SphericalMultiPolygon = MultiPolygon<SphericalPoint>;
@ -113,6 +118,28 @@ struct ColumnToLineStringsConverter
    }
 };

+/**
+ * Class which converts Column with type Array(Array(Tuple(Float64, Float64))) to a vector of boost multi_linestring type.
+*/
+template <typename Point>
+struct ColumnToMultiLineStringsConverter
+{
+    static std::vector<MultiLineString<Point>> convert(ColumnPtr col)
+    {
+        const IColumn::Offsets & offsets = typeid_cast<const ColumnArray &>(*col).getOffsets();
+        size_t prev_offset = 0;
+        std::vector<MultiLineString<Point>> answer(offsets.size());
+        auto all_linestrings = ColumnToLineStringsConverter<Point>::convert(typeid_cast<const ColumnArray &>(*col).getDataPtr());
+        for (size_t iter = 0; iter < offsets.size() && iter < all_linestrings.size(); ++iter)
+        {
+            for (size_t linestring_iter = prev_offset; linestring_iter < offsets[iter]; ++linestring_iter)
+                answer[iter].emplace_back(std::move(all_linestrings[linestring_iter]));
+            prev_offset = offsets[iter];
+        }
+        return answer;
+    }
+};
+
 /**
 * Class which converts Column with type Array(Tuple(Float64, Float64)) to a vector of boost ring type.
 */
@ -268,6 +295,38 @@ private:
    ColumnUInt64::MutablePtr offsets;
 };

+/// Serialize Point, MultiLineString as MultiLineString
+template <typename Point>
+class MultiLineStringSerializer
+{
+public:
+    MultiLineStringSerializer()
+        : offsets(ColumnUInt64::create())
+    {}
+
+    explicit MultiLineStringSerializer(size_t n)
+        : offsets(ColumnUInt64::create(n))
+    {}
+
+    void add(const MultiLineString<Point> & multilinestring)
+    {
+        size += multilinestring.size();
+        offsets->insertValue(size);
+        for (const auto & linestring : multilinestring)
+            linestring_serializer.add(linestring);
+    }
+
+    ColumnPtr finalize()
+    {
+        return ColumnArray::create(linestring_serializer.finalize(), std::move(offsets));
+    }
+
+private:
+    size_t size = 0;
+    LineStringSerializer<Point> linestring_serializer;
+    ColumnUInt64::MutablePtr offsets;
+};
+
 /// Almost the same as LineStringSerializer
 /// Serialize Point, Ring as Ring
 template <typename Point>
@ -411,6 +470,11 @@ static void callOnGeometryDataType(DataTypePtr type, F && f)
    else if (factory.get("LineString")->equals(*type) && type->getCustomName() && type->getCustomName()->getName() == "LineString")
        return f(ConverterType<ColumnToLineStringsConverter<Point>>());

+    /// We should take the name into consideration to avoid ambiguity.
+    /// Because for example both MultiLineString and Polygon are resolved to Array(Array(Point)).
+    else if (factory.get("MultiLineString")->equals(*type) && type->getCustomName() && type->getCustomName()->getName() == "MultiLineString")
+        return f(ConverterType<ColumnToMultiLineStringsConverter<Point>>());
+
    /// For backward compatibility if we call this function not on a custom type, we will consider Array(Tuple(Point)) as type Ring.
    else if (factory.get("Ring")->equals(*type))
        return f(ConverterType<ColumnToRingsConverter<Point>>());
--- a/src/Functions/polygonsIntersection.cpp
+++ b/src/Functions/polygonsIntersection.cpp
@ -75,6 +75,8 @@ public:
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be Point", getName());
            else if constexpr (std::is_same_v<ColumnToLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToLineStringsConverter<Point>, RightConverter>)
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be LineString", getName());
+            else if constexpr (std::is_same_v<ColumnToMultiLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToMultiLineStringsConverter<Point>, RightConverter>)
+                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be MultiLineString", getName());
            else
            {
                auto first = LeftConverter::convert(arguments[0].column->convertToFullColumnIfConst());
--- a/src/Functions/polygonsSymDifference.cpp
+++ b/src/Functions/polygonsSymDifference.cpp
@ -73,6 +73,8 @@ public:
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be Point", getName());
            else if constexpr (std::is_same_v<ColumnToLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToLineStringsConverter<Point>, RightConverter>)
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be LineString", getName());
+            else if constexpr (std::is_same_v<ColumnToMultiLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToMultiLineStringsConverter<Point>, RightConverter>)
+                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be MultiLineString", getName());
            else
            {
                auto first = LeftConverter::convert(arguments[0].column->convertToFullColumnIfConst());
--- a/src/Functions/polygonsUnion.cpp
+++ b/src/Functions/polygonsUnion.cpp
@ -73,6 +73,8 @@ public:
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be Point", getName());
            else if constexpr (std::is_same_v<ColumnToLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToLineStringsConverter<Point>, RightConverter>)
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be LineString", getName());
+            else if constexpr (std::is_same_v<ColumnToMultiLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToMultiLineStringsConverter<Point>, RightConverter>)
+                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be MultiLineString", getName());
            else
            {
                auto first = LeftConverter::convert(arguments[0].column->convertToFullColumnIfConst());
--- a/src/Functions/polygonsWithin.cpp
+++ b/src/Functions/polygonsWithin.cpp
@ -77,6 +77,8 @@ public:
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be Point", getName());
            else if constexpr (std::is_same_v<ColumnToLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToLineStringsConverter<Point>, RightConverter>)
                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be LineString", getName());
+            else if constexpr (std::is_same_v<ColumnToMultiLineStringsConverter<Point>, LeftConverter> || std::is_same_v<ColumnToMultiLineStringsConverter<Point>, RightConverter>)
+                throw Exception(ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT, "Any argument of function {} must not be MultiLineString", getName());
            else
            {
                auto first = LeftConverter::convert(arguments[0].column->convertToFullColumnIfConst());
--- a/src/Functions/readWkt.cpp
+++ b/src/Functions/readWkt.cpp
@ -87,6 +87,11 @@ struct ReadWKTLineStringNameHolder
    static constexpr const char * name = "readWKTLineString";
 };

+struct ReadWKTMultiLineStringNameHolder
+{
+    static constexpr const char * name = "readWKTMultiLineString";
+};
+
 struct ReadWKTRingNameHolder
 {
    static constexpr const char * name = "readWKTRing";
@ -131,6 +136,31 @@ Parses a Well-Known Text (WKT) representation of a LineString geometry and retur
        },
        .categories{"Unique identifiers"}
    });
+    factory.registerFunction<FunctionReadWKT<DataTypeMultiLineStringName, CartesianMultiLineString, MultiLineStringSerializer<CartesianPoint>, ReadWKTMultiLineStringNameHolder>>(FunctionDocumentation
+    {
+        .description=R"(
+Parses a Well-Known Text (WKT) representation of a MultiLineString geometry and returns it in the internal ClickHouse format.
+)",
+        .syntax = "readWKTMultiLineString(wkt_string)",
+        .arguments{
+            {"wkt_string", "The input WKT string representing a MultiLineString geometry."}
+        },
+        .returned_value = "The function returns a ClickHouse internal representation of the multilinestring geometry.",
+        .examples{
+            {"first call", "SELECT readWKTMultiLineString('MULTILINESTRING ((1 1, 2 2, 3 3), (4 4, 5 5, 6 6))');", R"(
+┌─readWKTMultiLineString('MULTILINESTRING ((1 1, 2 2, 3 3), (4 4, 5 5, 6 6))')─┐
+│ [[(1,1),(2,2),(3,3)],[(4,4),(5,5),(6,6)]]                                    │
+└──────────────────────────────────────────────────────────────────────────────┘
+
+            )"},
+            {"second call", "SELECT toTypeName(readWKTLineString('MULTILINESTRING ((1 1, 2 2, 3 3, 1 1))'));", R"(
+┌─toTypeName(readWKTLineString('MULTILINESTRING ((1 1, 2 2, 3 3, 1 1))'))─┐
+│ MultiLineString                                                         │
+└─────────────────────────────────────────────────────────────────────────┘
+            )"},
+        },
+        .categories{"Unique identifiers"}
+    });
    factory.registerFunction<FunctionReadWKT<DataTypeRingName, CartesianRing, RingSerializer<CartesianPoint>, ReadWKTRingNameHolder>>();
    factory.registerFunction<FunctionReadWKT<DataTypePolygonName, CartesianPolygon, PolygonSerializer<CartesianPoint>, ReadWKTPolygonNameHolder>>();
    factory.registerFunction<FunctionReadWKT<DataTypeMultiPolygonName, CartesianMultiPolygon, MultiPolygonSerializer<CartesianPoint>, ReadWKTMultiPolygonNameHolder>>();
--- a/src/Functions/toBool.cpp
+++ b/src/Functions/toBool.cpp
@ -54,8 +54,7 @@ namespace
                }
            };

-            FunctionOverloadResolverPtr func_builder_cast = createInternalCastOverloadResolver(CastType::nonAccurate, {});
-            auto func_cast = func_builder_cast->build(cast_args);
+            auto func_cast = createInternalCast(arguments[0], result_type, CastType::nonAccurate, {});
            return func_cast->execute(cast_args, result_type, arguments[0].column->size());
        }
    };
--- a/src/Interpreters/ActionsDAG.cpp
+++ b/src/Interpreters/ActionsDAG.cpp
@ -301,11 +301,11 @@ const ActionsDAG::Node & ActionsDAG::addCast(const Node & node_to_cast, const Da
    column.column = DataTypeString().createColumnConst(0, cast_type_constant_value);
    column.type = std::make_shared<DataTypeString>();

-    const auto * cast_type_constant_node = &addColumn(std::move(column));
+    const auto * cast_type_constant_node = &addColumn(column);
    ActionsDAG::NodeRawConstPtrs children = {&node_to_cast, cast_type_constant_node};
-    FunctionOverloadResolverPtr func_builder_cast = createInternalCastOverloadResolver(CastType::nonAccurate, {});
+    auto func_base_cast = createInternalCast(ColumnWithTypeAndName{node_to_cast.result_type, node_to_cast.result_name}, cast_type, CastType::nonAccurate, {});

-    return addFunction(func_builder_cast, std::move(children), result_name);
+    return addFunction(func_base_cast, std::move(children), result_name);
 }

 const ActionsDAG::Node & ActionsDAG::addFunctionImpl(
@ -1547,11 +1547,11 @@ ActionsDAG ActionsDAG::makeConvertingActions(
            const auto * left_arg = dst_node;

            CastDiagnostic diagnostic = {dst_node->result_name, res_elem.name};
-            FunctionOverloadResolverPtr func_builder_cast
-                = createInternalCastOverloadResolver(CastType::nonAccurate, std::move(diagnostic));
+            ColumnWithTypeAndName left_column{nullptr, dst_node->result_type, {}};
+            auto func_base_cast = createInternalCast(std::move(left_column), res_elem.type, CastType::nonAccurate, std::move(diagnostic));

            NodeRawConstPtrs children = { left_arg, right_arg };
-            dst_node = &actions_dag.addFunction(func_builder_cast, std::move(children), {});
+            dst_node = &actions_dag.addFunction(func_base_cast, std::move(children), {});
        }

        if (dst_node->column && isColumnConst(*dst_node->column) && !(res_elem.column && isColumnConst(*res_elem.column)))
--- a/src/Interpreters/HashJoin/FullHashJoin.cpp
+++ b/src/Interpreters/HashJoin/FullHashJoin.cpp
@ -1,4 +1,4 @@
-#include <Interpreters/HashJoin/HashJoinMethods.h>
+#include <Interpreters/HashJoin/HashJoinMethodsImpl.h>

 namespace DB
 {
--- a/src/Interpreters/HashJoin/HashJoin.cpp
+++ b/src/Interpreters/HashJoin/HashJoin.cpp
@ -291,12 +291,13 @@ void HashJoin::dataMapInit(MapsVariant & map)
 {
    if (kind == JoinKind::Cross)
        return;
-    joinDispatchInit(kind, strictness, map);
-    joinDispatch(kind, strictness, map, [&](auto, auto, auto & map_) { map_.create(data->type); });
+    auto prefer_use_maps_all = table_join->getMixedJoinExpression() != nullptr;
+    joinDispatchInit(kind, strictness, map, prefer_use_maps_all);
+    joinDispatch(kind, strictness, map, prefer_use_maps_all, [&](auto, auto, auto & map_) { map_.create(data->type); });

    if (reserve_num)
    {
-        joinDispatch(kind, strictness, map, [&](auto, auto, auto & map_) { map_.reserve(data->type, reserve_num); });
+        joinDispatch(kind, strictness, map, prefer_use_maps_all, [&](auto, auto, auto & map_) { map_.reserve(data->type, reserve_num); });
    }

    if (!data)
@ -327,9 +328,10 @@ size_t HashJoin::getTotalRowCount() const
    }
    else
    {
+        auto prefer_use_maps_all = table_join->getMixedJoinExpression() != nullptr;
        for (const auto & map : data->maps)
        {
-            joinDispatch(kind, strictness, map, [&](auto, auto, auto & map_) { res += map_.getTotalRowCount(data->type); });
+            joinDispatch(kind, strictness, map, prefer_use_maps_all, [&](auto, auto, auto & map_) { res += map_.getTotalRowCount(data->type); });
        }
    }

@ -367,9 +369,10 @@ size_t HashJoin::getTotalByteCount() const

    if (data->type != Type::CROSS)
    {
+        auto prefer_use_maps_all = table_join->getMixedJoinExpression() != nullptr;
        for (const auto & map : data->maps)
        {
-            joinDispatch(kind, strictness, map, [&](auto, auto, auto & map_) { res += map_.getTotalByteCountImpl(data->type); });
+            joinDispatch(kind, strictness, map, prefer_use_maps_all, [&](auto, auto, auto & map_) { res += map_.getTotalByteCountImpl(data->type); });
        }
    }
    return res;
@ -520,6 +523,8 @@ bool HashJoin::addBlockToJoin(const Block & source_block_, bool check_limits)
        return true;
    }

+    bool prefer_use_maps_all = table_join->getMixedJoinExpression() != nullptr;
+
    size_t total_rows = 0;
    size_t total_bytes = 0;
    {
@ -592,7 +597,7 @@ bool HashJoin::addBlockToJoin(const Block & source_block_, bool check_limits)
            bool is_inserted = false;
            if (kind != JoinKind::Cross)
            {
-                joinDispatch(kind, strictness, data->maps[onexpr_idx], [&](auto kind_, auto strictness_, auto & map)
+                joinDispatch(kind, strictness, data->maps[onexpr_idx], prefer_use_maps_all, [&](auto kind_, auto strictness_, auto & map)
                {
                    size_t size = HashJoinMethods<kind_, strictness_, std::decay_t<decltype(map)>>::insertFromBlockImpl(
                            *this,
@ -608,10 +613,10 @@ bool HashJoin::addBlockToJoin(const Block & source_block_, bool check_limits)
                            is_inserted);

                    if (flag_per_row)
-                        used_flags->reinit<kind_, strictness_>(stored_block);
+                        used_flags->reinit<kind_, strictness_, std::is_same_v<std::decay_t<decltype(map)>, MapsAll>>(stored_block);
                    else if (is_inserted)
                        /// Number of buckets + 1 value from zero storage
-                        used_flags->reinit<kind_, strictness_>(size + 1);
+                        used_flags->reinit<kind_, strictness_, std::is_same_v<std::decay_t<decltype(map)>, MapsAll>>(size + 1);
                });
            }

@ -873,7 +878,7 @@ ColumnWithTypeAndName HashJoin::joinGet(const Block & block, const Block & block
        keys.insert(std::move(key));
    }

-    static_assert(!MapGetter<JoinKind::Left, JoinStrictness::Any>::flagged,
+    static_assert(!MapGetter<JoinKind::Left, JoinStrictness::Any, false>::flagged,
                  "joinGet are not protected from hash table changes between block processing");

    std::vector<const MapsOne *> maps_vector;
@ -914,16 +919,34 @@ void HashJoin::joinBlock(Block & block, ExtraBlockPtr & not_processed)
        materializeBlockInplace(block);
    }

+    bool prefer_use_maps_all = table_join->getMixedJoinExpression() != nullptr;
    {
        std::vector<const std::decay_t<decltype(data->maps[0])> * > maps_vector;
        for (size_t i = 0; i < table_join->getClauses().size(); ++i)
            maps_vector.push_back(&data->maps[i]);

-        if (joinDispatch(kind, strictness, maps_vector, [&](auto kind_, auto strictness_, auto & maps_vector_)
+        if (joinDispatch(kind, strictness, maps_vector, prefer_use_maps_all, [&](auto kind_, auto strictness_, auto & maps_vector_)
        {
-            using MapType = typename MapGetter<kind_, strictness_>::Map;
-            Block remaining_block = HashJoinMethods<kind_, strictness_, MapType>::joinBlockImpl(
-                *this, block, sample_block_with_columns_to_add, maps_vector_);
+            Block remaining_block;
+            if constexpr (std::is_same_v<std::decay_t<decltype(maps_vector_)>, std::vector<const MapsAll *>>)
+            {
+                remaining_block = HashJoinMethods<kind_, strictness_, MapsAll>::joinBlockImpl(
+                    *this, block, sample_block_with_columns_to_add, maps_vector_);
+            }
+            else if constexpr (std::is_same_v<std::decay_t<decltype(maps_vector_)>, std::vector<const MapsOne *>>)
+            {
+                remaining_block = HashJoinMethods<kind_, strictness_, MapsOne>::joinBlockImpl(
+                    *this, block, sample_block_with_columns_to_add, maps_vector_);
+            }
+            else if constexpr (std::is_same_v<std::decay_t<decltype(maps_vector_)>, std::vector<const MapsAsof *>>)
+            {
+                remaining_block = HashJoinMethods<kind_, strictness_, MapsAsof>::joinBlockImpl(
+                    *this, block, sample_block_with_columns_to_add, maps_vector_);
+            }
+            else
+            {
+                throw Exception(ErrorCodes::LOGICAL_ERROR, "Unknown maps type");
+            }
            if (remaining_block.rows())
                not_processed = std::make_shared<ExtraBlock>(ExtraBlock{std::move(remaining_block)});
            else
@ -1023,7 +1046,8 @@ public:
                rows_added = fillColumnsFromMap(map, columns_right);
            };

-            if (!joinDispatch(parent.kind, parent.strictness, parent.data->maps.front(), fill_callback))
+            bool prefer_use_maps_all = parent.table_join->getMixedJoinExpression() != nullptr;
+            if (!joinDispatch(parent.kind, parent.strictness, parent.data->maps.front(), prefer_use_maps_all, fill_callback))
                throw Exception(ErrorCodes::LOGICAL_ERROR, "Unknown JOIN strictness '{}' (must be on of: ANY, ALL, ASOF)", parent.strictness);
        }

@ -1220,11 +1244,12 @@ void HashJoin::reuseJoinedData(const HashJoin & join)
    if (flag_per_row)
        throw Exception(ErrorCodes::NOT_IMPLEMENTED, "StorageJoin with ORs is not supported");

+    bool prefer_use_maps_all = join.table_join->getMixedJoinExpression() != nullptr;
    for (auto & map : data->maps)
    {
-        joinDispatch(kind, strictness, map, [this](auto kind_, auto strictness_, auto & map_)
+        joinDispatch(kind, strictness, map, prefer_use_maps_all, [this](auto kind_, auto strictness_, auto & map_)
        {
-            used_flags->reinit<kind_, strictness_>(map_.getBufferSizeInCells(data->type) + 1);
+            used_flags->reinit<kind_, strictness_, std::is_same_v<std::decay_t<decltype(map_)>, MapsAll>>(map_.getBufferSizeInCells(data->type) + 1);
        });
    }
 }
@ -1304,7 +1329,9 @@ void HashJoin::validateAdditionalFilterExpression(ExpressionActionsPtr additiona
            additional_filter_expression->dumpActions());
    }

-    bool is_supported = (strictness == JoinStrictness::All) && (isInnerOrLeft(kind) || isRightOrFull(kind));
+    bool is_supported = ((strictness == JoinStrictness::All) && (isInnerOrLeft(kind) || isRightOrFull(kind)))
+        || ((strictness == JoinStrictness::Semi || strictness == JoinStrictness::Any || strictness == JoinStrictness::Anti)
+                && (isLeft(kind) || isRight(kind))) || (strictness == JoinStrictness::Any && (isInner(kind)));
    if (!is_supported)
    {
        throw Exception(ErrorCodes::INVALID_JOIN_ON_EXPRESSION,
--- a/src/Interpreters/HashJoin/HashJoinMethods.h
+++ b/src/Interpreters/HashJoin/HashJoinMethods.h
@ -12,15 +12,8 @@
 #include <Poco/Logger.h>
 #include <Common/logger_useful.h>

-
 namespace DB
 {
-namespace ErrorCodes
-{
-    extern const int UNSUPPORTED_JOIN_KEYS;
-    extern const int LOGICAL_ERROR;
-}
-
 /// Inserting an element into a hash table of the form `key -> reference to a string`, which will then be used by JOIN.
 template <typename HashMap, typename KeyGetter>
 struct Inserter
@ -64,7 +57,6 @@ struct Inserter
    }
 };

-
 /// MapsTemplate is one of MapsOne, MapsAll and MapsAsof
 template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
 class HashJoinMethods
@ -81,27 +73,7 @@ public:
        ConstNullMapPtr null_map,
        UInt8ColumnDataPtr join_mask,
        Arena & pool,
-        bool & is_inserted)
-    {
-        switch (type)
-        {
-            case HashJoin::Type::EMPTY:
-                [[fallthrough]];
-            case HashJoin::Type::CROSS:
-                /// Do nothing. We will only save block, and it is enough
-                is_inserted = true;
-                return 0;
-
-        #define M(TYPE) \
-            case HashJoin::Type::TYPE: \
-                return insertFromBlockImplTypeCase<typename KeyGetterForType<HashJoin::Type::TYPE, std::remove_reference_t<decltype(*maps.TYPE)>>::Type>(\
-                    join, *maps.TYPE, rows, key_columns, key_sizes, stored_block, null_map, join_mask, pool, is_inserted); \
-                break;
-
-            APPLY_FOR_JOIN_VARIANTS(M)
-        #undef M
-        }
-    }
+        bool & is_inserted);

    using MapsTemplateVector = std::vector<const MapsTemplate *>;

@ -110,280 +82,36 @@ public:
        Block & block,
        const Block & block_with_columns_to_add,
        const MapsTemplateVector & maps_,
-        bool is_join_get = false)
-    {
-        constexpr JoinFeatures<KIND, STRICTNESS> join_features;
-
-        std::vector<JoinOnKeyColumns> join_on_keys;
-        const auto & onexprs = join.table_join->getClauses();
-        for (size_t i = 0; i < onexprs.size(); ++i)
-        {
-            const auto & key_names = !is_join_get ? onexprs[i].key_names_left : onexprs[i].key_names_right;
-            join_on_keys.emplace_back(block, key_names, onexprs[i].condColumnNames().first, join.key_sizes[i]);
-        }
-        size_t existing_columns = block.columns();
-
-        /** If you use FULL or RIGHT JOIN, then the columns from the "left" table must be materialized.
-          * Because if they are constants, then in the "not joined" rows, they may have different values
-          *  - default values, which can differ from the values of these constants.
-          */
-        if constexpr (join_features.right || join_features.full)
-        {
-            materializeBlockInplace(block);
-        }
-
-        /** For LEFT/INNER JOIN, the saved blocks do not contain keys.
-          * For FULL/RIGHT JOIN, the saved blocks contain keys;
-          *  but they will not be used at this stage of joining (and will be in `AdderNonJoined`), and they need to be skipped.
-          * For ASOF, the last column is used as the ASOF column
-          */
-        AddedColumns<!join_features.is_any_join> added_columns(
-            block,
-            block_with_columns_to_add,
-            join.savedBlockSample(),
-            join,
-            std::move(join_on_keys),
-            join.table_join->getMixedJoinExpression(),
-            join_features.is_asof_join,
-            is_join_get);
-
-        bool has_required_right_keys = (join.required_right_keys.columns() != 0);
-        added_columns.need_filter = join_features.need_filter || has_required_right_keys;
-        added_columns.max_joined_block_rows = join.max_joined_block_rows;
-        if (!added_columns.max_joined_block_rows)
-            added_columns.max_joined_block_rows = std::numeric_limits<size_t>::max();
-        else
-            added_columns.reserve(join_features.need_replication);
-
-        size_t num_joined = switchJoinRightColumns(maps_, added_columns, join.data->type, *join.used_flags);
-        /// Do not hold memory for join_on_keys anymore
-        added_columns.join_on_keys.clear();
-        Block remaining_block = sliceBlock(block, num_joined);
-
-        added_columns.buildOutput();
-        for (size_t i = 0; i < added_columns.size(); ++i)
-            block.insert(added_columns.moveColumn(i));
-
-        std::vector<size_t> right_keys_to_replicate [[maybe_unused]];
-
-        if constexpr (join_features.need_filter)
-        {
-            /// If ANY INNER | RIGHT JOIN - filter all the columns except the new ones.
-            for (size_t i = 0; i < existing_columns; ++i)
-                block.safeGetByPosition(i).column = block.safeGetByPosition(i).column->filter(added_columns.filter, -1);
-
-            /// Add join key columns from right block if needed using value from left table because of equality
-            for (size_t i = 0; i < join.required_right_keys.columns(); ++i)
-            {
-                const auto & right_key = join.required_right_keys.getByPosition(i);
-                /// asof column is already in block.
-                if (join_features.is_asof_join && right_key.name == join.table_join->getOnlyClause().key_names_right.back())
-                    continue;
-
-                const auto & left_column = block.getByName(join.required_right_keys_sources[i]);
-                const auto & right_col_name = join.getTableJoin().renamedRightColumnName(right_key.name);
-                auto right_col = copyLeftKeyColumnToRight(right_key.type, right_col_name, left_column);
-                block.insert(std::move(right_col));
-            }
-        }
-        else if (has_required_right_keys)
-        {
-            /// Add join key columns from right block if needed.
-            for (size_t i = 0; i < join.required_right_keys.columns(); ++i)
-            {
-                const auto & right_key = join.required_right_keys.getByPosition(i);
-                auto right_col_name = join.getTableJoin().renamedRightColumnName(right_key.name);
-                /// asof column is already in block.
-                if (join_features.is_asof_join && right_key.name == join.table_join->getOnlyClause().key_names_right.back())
-                    continue;
-
-                const auto & left_column = block.getByName(join.required_right_keys_sources[i]);
-                auto right_col = copyLeftKeyColumnToRight(right_key.type, right_col_name, left_column, &added_columns.filter);
-                block.insert(std::move(right_col));
-
-                if constexpr (join_features.need_replication)
-                    right_keys_to_replicate.push_back(block.getPositionByName(right_col_name));
-            }
-        }
-
-        if constexpr (join_features.need_replication)
-        {
-            std::unique_ptr<IColumn::Offsets> & offsets_to_replicate = added_columns.offsets_to_replicate;
-
-            /// If ALL ... JOIN - we replicate all the columns except the new ones.
-            for (size_t i = 0; i < existing_columns; ++i)
-            {
-                block.safeGetByPosition(i).column = block.safeGetByPosition(i).column->replicate(*offsets_to_replicate);
-            }
-
-            /// Replicate additional right keys
-            for (size_t pos : right_keys_to_replicate)
-            {
-               block.safeGetByPosition(pos).column = block.safeGetByPosition(pos).column->replicate(*offsets_to_replicate);
-            }
-        }
-
-        return remaining_block;
-    }
-
+        bool is_join_get = false);
 private:
    template <typename KeyGetter, bool is_asof_join>
-    static KeyGetter createKeyGetter(const ColumnRawPtrs & key_columns, const Sizes & key_sizes)
-    {
-        if constexpr (is_asof_join)
-        {
-            auto key_column_copy = key_columns;
-            auto key_size_copy = key_sizes;
-            key_column_copy.pop_back();
-            key_size_copy.pop_back();
-            return KeyGetter(key_column_copy, key_size_copy, nullptr);
-        }
-        else
-            return KeyGetter(key_columns, key_sizes, nullptr);
-    }
+    static KeyGetter createKeyGetter(const ColumnRawPtrs & key_columns, const Sizes & key_sizes);

    template <typename KeyGetter, typename HashMap>
-    static size_t NO_INLINE insertFromBlockImplTypeCase(
+    static size_t insertFromBlockImplTypeCase(
        HashJoin & join, HashMap & map, size_t rows, const ColumnRawPtrs & key_columns,
-        const Sizes & key_sizes, Block * stored_block, ConstNullMapPtr null_map, UInt8ColumnDataPtr join_mask, Arena & pool, bool & is_inserted)
-    {
-        [[maybe_unused]] constexpr bool mapped_one = std::is_same_v<typename HashMap::mapped_type, RowRef>;
-        constexpr bool is_asof_join = STRICTNESS == JoinStrictness::Asof;
-
-        const IColumn * asof_column [[maybe_unused]] = nullptr;
-        if constexpr (is_asof_join)
-            asof_column = key_columns.back();
-
-        auto key_getter = createKeyGetter<KeyGetter, is_asof_join>(key_columns, key_sizes);
-
-        /// For ALL and ASOF join always insert values
-        is_inserted = !mapped_one || is_asof_join;
-
-        for (size_t i = 0; i < rows; ++i)
-        {
-            if (null_map && (*null_map)[i])
-            {
-                /// nulls are not inserted into hash table,
-                /// keep them for RIGHT and FULL joins
-                is_inserted = true;
-                continue;
-            }
-
-            /// Check condition for right table from ON section
-            if (join_mask && !(*join_mask)[i])
-                continue;
-
-            if constexpr (is_asof_join)
-                Inserter<HashMap, KeyGetter>::insertAsof(join, map, key_getter, stored_block, i, pool, *asof_column);
-            else if constexpr (mapped_one)
-                is_inserted |= Inserter<HashMap, KeyGetter>::insertOne(join, map, key_getter, stored_block, i, pool);
-            else
-                Inserter<HashMap, KeyGetter>::insertAll(join, map, key_getter, stored_block, i, pool);
-        }
-        return map.getBufferSizeInCells();
-    }
+        const Sizes & key_sizes, Block * stored_block, ConstNullMapPtr null_map, UInt8ColumnDataPtr join_mask, Arena & pool, bool & is_inserted);

    template <typename AddedColumns>
    static size_t switchJoinRightColumns(
        const std::vector<const MapsTemplate *> & mapv,
        AddedColumns & added_columns,
        HashJoin::Type type,
-        JoinStuff::JoinUsedFlags & used_flags)
-    {
-        constexpr bool is_asof_join = STRICTNESS == JoinStrictness::Asof;
-        switch (type)
-        {
-            case HashJoin::Type::EMPTY: {
-                if constexpr (!is_asof_join)
-                {
-                    using KeyGetter = KeyGetterEmpty<typename MapsTemplate::MappedType>;
-                    std::vector<KeyGetter> key_getter_vector;
-                    key_getter_vector.emplace_back();
-
-                    using MapTypeVal = typename KeyGetter::MappedType;
-                    std::vector<const MapTypeVal *> a_map_type_vector;
-                    a_map_type_vector.emplace_back();
-                    return joinRightColumnsSwitchNullability<KeyGetter>(
-                        std::move(key_getter_vector), a_map_type_vector, added_columns, used_flags);
-                }
-                throw Exception(ErrorCodes::UNSUPPORTED_JOIN_KEYS, "Unsupported JOIN keys. Type: {}", type);
-            }
-        #define M(TYPE) \
-            case HashJoin::Type::TYPE: \
-            {                                                           \
-                using MapTypeVal = const typename std::remove_reference_t<decltype(MapsTemplate::TYPE)>::element_type; \
-                using KeyGetter = typename KeyGetterForType<HashJoin::Type::TYPE, MapTypeVal>::Type; \
-                std::vector<const MapTypeVal *> a_map_type_vector(mapv.size()); \
-                std::vector<KeyGetter> key_getter_vector;                     \
-                for (size_t d = 0; d < added_columns.join_on_keys.size(); ++d)                   \
-                {       \
-                    const auto & join_on_key = added_columns.join_on_keys[d];     \
-                    a_map_type_vector[d] = mapv[d]->TYPE.get();              \
-                    key_getter_vector.push_back(std::move(createKeyGetter<KeyGetter, is_asof_join>(join_on_key.key_columns, join_on_key.key_sizes))); \
-                }                                                           \
-                return joinRightColumnsSwitchNullability<KeyGetter>( \
-                                std::move(key_getter_vector), a_map_type_vector, added_columns, used_flags); \
-            }
-            APPLY_FOR_JOIN_VARIANTS(M)
-        #undef M
-
-        default:
-            throw Exception(ErrorCodes::UNSUPPORTED_JOIN_KEYS, "Unsupported JOIN keys (type: {})", type);
-        }
-    }
+        JoinStuff::JoinUsedFlags & used_flags);

    template <typename KeyGetter, typename Map, typename AddedColumns>
    static size_t joinRightColumnsSwitchNullability(
        std::vector<KeyGetter> && key_getter_vector,
        const std::vector<const Map *> & mapv,
        AddedColumns & added_columns,
-        JoinStuff::JoinUsedFlags & used_flags)
-    {
-        if (added_columns.need_filter)
-        {
-            return joinRightColumnsSwitchMultipleDisjuncts<KeyGetter, Map, true>(
-            std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags);
-        }
-        else
-        {
-            return joinRightColumnsSwitchMultipleDisjuncts<KeyGetter, Map, false>(
-            std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags);
-        }
-    }
+        JoinStuff::JoinUsedFlags & used_flags);

    template <typename KeyGetter, typename Map, bool need_filter, typename AddedColumns>
    static size_t joinRightColumnsSwitchMultipleDisjuncts(
        std::vector<KeyGetter> && key_getter_vector,
        const std::vector<const Map *> & mapv,
        AddedColumns & added_columns,
-        JoinStuff::JoinUsedFlags & used_flags)
-    {
-        constexpr JoinFeatures<KIND, STRICTNESS> join_features;
-        if constexpr (join_features.is_all_join)
-        {
-        if (added_columns.additional_filter_expression)
-        {
-            bool mark_per_row_used = join_features.right || join_features.full || mapv.size() > 1;
-            return joinRightColumnsWithAddtitionalFilter<KeyGetter, Map, join_features.need_replication>(
-                std::forward<std::vector<KeyGetter>>(key_getter_vector),
-                mapv,
-                added_columns,
-                used_flags,
-                need_filter,
-                join_features.need_flags,
-                join_features.add_missing,
-                mark_per_row_used);
-        }
-        }
-
-        if (added_columns.additional_filter_expression)
-            throw Exception(ErrorCodes::LOGICAL_ERROR, "Additional filter expression is not supported for this JOIN");
-
-        return mapv.size() > 1 ? joinRightColumns<KeyGetter, Map, need_filter, true>(
-                   std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags)
-                               : joinRightColumns<KeyGetter, Map, need_filter, false>(
-                                   std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags);
-    }
+        JoinStuff::JoinUsedFlags & used_flags);

    /// Joins right table columns which indexes are present in right_indexes using specified map.
    /// Makes filter (1 if row presented in right table) and returns offsets to replicate (for ALL JOINS).
@ -392,464 +120,30 @@ private:
        std::vector<KeyGetter> && key_getter_vector,
        const std::vector<const Map *> & mapv,
        AddedColumns & added_columns,
-        JoinStuff::JoinUsedFlags & used_flags)
-    {
-        constexpr JoinFeatures<KIND, STRICTNESS> join_features;
-
-        size_t rows = added_columns.rows_to_add;
-        if constexpr (need_filter)
-            added_columns.filter = IColumn::Filter(rows, 0);
-
-        Arena pool;
-
-        if constexpr (join_features.need_replication)
-            added_columns.offsets_to_replicate = std::make_unique<IColumn::Offsets>(rows);
-
-        IColumn::Offset current_offset = 0;
-        size_t max_joined_block_rows = added_columns.max_joined_block_rows;
-        size_t i = 0;
-        for (; i < rows; ++i)
-        {
-            if constexpr (join_features.need_replication)
-            {
-                if (unlikely(current_offset >= max_joined_block_rows))
-                {
-                    added_columns.offsets_to_replicate->resize_assume_reserved(i);
-                    added_columns.filter.resize_assume_reserved(i);
-                    break;
-                }
-            }
-
-            bool right_row_found = false;
-
-            KnownRowsHolder<flag_per_row> known_rows;
-            for (size_t onexpr_idx = 0; onexpr_idx < added_columns.join_on_keys.size(); ++onexpr_idx)
-            {
-                const auto & join_keys = added_columns.join_on_keys[onexpr_idx];
-                if (join_keys.null_map && (*join_keys.null_map)[i])
-                        continue;
-
-                bool row_acceptable = !join_keys.isRowFiltered(i);
-                using FindResult = typename KeyGetter::FindResult;
-                auto find_result = row_acceptable ? key_getter_vector[onexpr_idx].findKey(*(mapv[onexpr_idx]), i, pool) : FindResult();
-
-                if (find_result.isFound())
-                {
-                    right_row_found = true;
-                    auto & mapped = find_result.getMapped();
-                    if constexpr (join_features.is_asof_join)
-                    {
-                        const IColumn & left_asof_key = added_columns.leftAsofKey();
-
-                        auto row_ref = mapped->findAsof(left_asof_key, i);
-                        if (row_ref.block)
-                        {
-                            setUsed<need_filter>(added_columns.filter, i);
-                            if constexpr (flag_per_row)
-                                used_flags.template setUsed<join_features.need_flags, flag_per_row>(row_ref.block, row_ref.row_num, 0);
-                            else
-                                used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
-
-                            added_columns.appendFromBlock(*row_ref.block, row_ref.row_num, join_features.add_missing);
-                        }
-                        else
-                            addNotFoundRow<join_features.add_missing, join_features.need_replication>(added_columns, current_offset);
-                    }
-                    else if constexpr (join_features.is_all_join)
-                    {
-                        setUsed<need_filter>(added_columns.filter, i);
-                        used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
-                        auto used_flags_opt = join_features.need_flags ? &used_flags : nullptr;
-                        addFoundRowAll<Map, join_features.add_missing>(mapped, added_columns, current_offset, known_rows, used_flags_opt);
-                    }
-                    else if constexpr ((join_features.is_any_join || join_features.is_semi_join) && join_features.right)
-                    {
-                        /// Use first appeared left key + it needs left columns replication
-                        bool used_once = used_flags.template setUsedOnce<join_features.need_flags, flag_per_row>(find_result);
-                        if (used_once)
-                        {
-                            auto used_flags_opt = join_features.need_flags ? &used_flags : nullptr;
-                            setUsed<need_filter>(added_columns.filter, i);
-                            addFoundRowAll<Map, join_features.add_missing>(
-                                mapped, added_columns, current_offset, known_rows, used_flags_opt);
-                        }
-                    }
-                    else if constexpr (join_features.is_any_join && KIND == JoinKind::Inner)
-                    {
-                        bool used_once = used_flags.template setUsedOnce<join_features.need_flags, flag_per_row>(find_result);
-
-                        /// Use first appeared left key only
-                        if (used_once)
-                        {
-                            setUsed<need_filter>(added_columns.filter, i);
-                            added_columns.appendFromBlock(*mapped.block, mapped.row_num, join_features.add_missing);
-                        }
-
-                        break;
-                    }
-                    else if constexpr (join_features.is_any_join && join_features.full)
-                    {
-                        /// TODO
-                    }
-                    else if constexpr (join_features.is_anti_join)
-                    {
-                        if constexpr (join_features.right && join_features.need_flags)
-                            used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
-                    }
-                    else /// ANY LEFT, SEMI LEFT, old ANY (RightAny)
-                    {
-                        setUsed<need_filter>(added_columns.filter, i);
-                        used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
-                        added_columns.appendFromBlock(*mapped.block, mapped.row_num, join_features.add_missing);
-
-                        if (join_features.is_any_or_semi_join)
-                        {
-                            break;
-                        }
-                    }
-                }
-            }
-
-            if (!right_row_found)
-            {
-                if constexpr (join_features.is_anti_join && join_features.left)
-                    setUsed<need_filter>(added_columns.filter, i);
-                addNotFoundRow<join_features.add_missing, join_features.need_replication>(added_columns, current_offset);
-            }
-
-            if constexpr (join_features.need_replication)
-            {
-                (*added_columns.offsets_to_replicate)[i] = current_offset;
-            }
-        }
-
-        added_columns.applyLazyDefaults();
-        return i;
-    }
+        JoinStuff::JoinUsedFlags & used_flags);

    template <bool need_filter>
-    static void setUsed(IColumn::Filter & filter [[maybe_unused]], size_t pos [[maybe_unused]])
-    {
-        if constexpr (need_filter)
-            filter[pos] = 1;
-    }
+    static void setUsed(IColumn::Filter & filter [[maybe_unused]], size_t pos [[maybe_unused]]);

    template <typename AddedColumns>
    static ColumnPtr buildAdditionalFilter(
        size_t left_start_row,
        const std::vector<RowRef> & selected_rows,
        const std::vector<size_t> & row_replicate_offset,
-        AddedColumns & added_columns)
-    {
-        ColumnPtr result_column;
-        do
-        {
-            if (selected_rows.empty())
-            {
-                result_column = ColumnUInt8::create();
-                break;
-            }
-            const Block & sample_right_block = *selected_rows.begin()->block;
-            if (!sample_right_block || !added_columns.additional_filter_expression)
-            {
-                auto filter = ColumnUInt8::create();
-                filter->insertMany(1, selected_rows.size());
-                result_column = std::move(filter);
-                break;
-            }
-
-            auto required_cols = added_columns.additional_filter_expression->getRequiredColumnsWithTypes();
-            if (required_cols.empty())
-            {
-                Block block;
-                added_columns.additional_filter_expression->execute(block);
-                result_column = block.getByPosition(0).column->cloneResized(selected_rows.size());
-                break;
-            }
-            NameSet required_column_names;
-            for (auto & col : required_cols)
-                required_column_names.insert(col.name);
-
-            Block executed_block;
-            size_t right_col_pos = 0;
-            for (const auto & col : sample_right_block.getColumnsWithTypeAndName())
-            {
-                if (required_column_names.contains(col.name))
-                {
-                    auto new_col = col.column->cloneEmpty();
-                    for (const auto & selected_row : selected_rows)
-                    {
-                        const auto & src_col = selected_row.block->getByPosition(right_col_pos);
-                        new_col->insertFrom(*src_col.column, selected_row.row_num);
-                    }
-                    executed_block.insert({std::move(new_col), col.type, col.name});
-                }
-                right_col_pos += 1;
-            }
-            if (!executed_block)
-            {
-                result_column = ColumnUInt8::create();
-                break;
-            }
-
-            for (const auto & col_name : required_column_names)
-            {
-                const auto * src_col = added_columns.left_block.findByName(col_name);
-                if (!src_col)
-                    continue;
-                auto new_col = src_col->column->cloneEmpty();
-                size_t prev_left_offset = 0;
-                for (size_t i = 1; i < row_replicate_offset.size(); ++i)
-                {
-                    const size_t & left_offset = row_replicate_offset[i];
-                    size_t rows = left_offset - prev_left_offset;
-                    if (rows)
-                        new_col->insertManyFrom(*src_col->column, left_start_row + i - 1, rows);
-                    prev_left_offset = left_offset;
-                }
-                executed_block.insert({std::move(new_col), src_col->type, col_name});
-            }
-            if (!executed_block)
-            {
-                throw Exception(
-                    ErrorCodes::LOGICAL_ERROR,
-                    "required columns: [{}], but not found any in left/right table. right table: {}, left table: {}",
-                    required_cols.toString(),
-                    sample_right_block.dumpNames(),
-                    added_columns.left_block.dumpNames());
-            }
-
-            for (const auto & col : executed_block.getColumnsWithTypeAndName())
-                if (!col.column || !col.type)
-                    throw Exception(ErrorCodes::LOGICAL_ERROR, "Illegal nullptr column in input block: {}", executed_block.dumpStructure());
-
-            added_columns.additional_filter_expression->execute(executed_block);
-            result_column = executed_block.getByPosition(0).column->convertToFullColumnIfConst();
-            executed_block.clear();
-        } while (false);
-
-        result_column = result_column->convertToFullIfNeeded();
-        if (result_column->isNullable())
-        {
-            /// Convert Nullable(UInt8) to UInt8 ensuring that nulls are zeros
-            /// Trying to avoid copying data, since we are the only owner of the column.
-            ColumnPtr mask_column = assert_cast<const ColumnNullable &>(*result_column).getNullMapColumnPtr();
-
-            MutableColumnPtr mutable_column;
-            {
-                ColumnPtr nested_column = assert_cast<const ColumnNullable &>(*result_column).getNestedColumnPtr();
-                result_column.reset();
-                mutable_column = IColumn::mutate(std::move(nested_column));
-            }
-
-            auto & column_data = assert_cast<ColumnUInt8 &>(*mutable_column).getData();
-            const auto & mask_column_data = assert_cast<const ColumnUInt8 &>(*mask_column).getData();
-            for (size_t i = 0; i < column_data.size(); ++i)
-            {
-                if (mask_column_data[i])
-                    column_data[i] = 0;
-            }
-            return mutable_column;
-        }
-        return result_column;
-    }
+        AddedColumns & added_columns);

    /// First to collect all matched rows refs by join keys, then filter out rows which are not true in additional filter expression.
-    template <typename KeyGetter, typename Map, bool need_replication, typename AddedColumns>
+    template <typename KeyGetter, typename Map, typename AddedColumns>
    static size_t joinRightColumnsWithAddtitionalFilter(
        std::vector<KeyGetter> && key_getter_vector,
        const std::vector<const Map *> & mapv,
        AddedColumns & added_columns,
        JoinStuff::JoinUsedFlags & used_flags [[maybe_unused]],
        bool need_filter [[maybe_unused]],
-        bool need_flags [[maybe_unused]],
-        bool add_missing [[maybe_unused]],
-        bool flag_per_row [[maybe_unused]])
-    {
-        size_t left_block_rows = added_columns.rows_to_add;
-        if (need_filter)
-            added_columns.filter = IColumn::Filter(left_block_rows, 0);
-
-        std::unique_ptr<Arena> pool;
-
-        if constexpr (need_replication)
-            added_columns.offsets_to_replicate = std::make_unique<IColumn::Offsets>(left_block_rows);
-
-        std::vector<size_t> row_replicate_offset;
-        row_replicate_offset.reserve(left_block_rows);
-
-        using FindResult = typename KeyGetter::FindResult;
-        size_t max_joined_block_rows = added_columns.max_joined_block_rows;
-        size_t left_row_iter = 0;
-        PreSelectedRows selected_rows;
-        selected_rows.reserve(left_block_rows);
-        std::vector<FindResult> find_results;
-        find_results.reserve(left_block_rows);
-        bool exceeded_max_block_rows = false;
-        IColumn::Offset total_added_rows = 0;
-        IColumn::Offset current_added_rows = 0;
-
-        auto collect_keys_matched_rows_refs = [&]()
-        {
-            pool = std::make_unique<Arena>();
-            find_results.clear();
-            row_replicate_offset.clear();
-            row_replicate_offset.push_back(0);
-            current_added_rows = 0;
-            selected_rows.clear();
-            for (; left_row_iter < left_block_rows; ++left_row_iter)
-            {
-                if constexpr (need_replication)
-                {
-                    if (unlikely(total_added_rows + current_added_rows >= max_joined_block_rows))
-                    {
-                        break;
-                    }
-                }
-                KnownRowsHolder<true> all_flag_known_rows;
-                KnownRowsHolder<false> single_flag_know_rows;
-                for (size_t join_clause_idx = 0; join_clause_idx < added_columns.join_on_keys.size(); ++join_clause_idx)
-                {
-                    const auto & join_keys = added_columns.join_on_keys[join_clause_idx];
-                    if (join_keys.null_map && (*join_keys.null_map)[left_row_iter])
-                        continue;
-
-                    bool row_acceptable = !join_keys.isRowFiltered(left_row_iter);
-                    auto find_result = row_acceptable
-                        ? key_getter_vector[join_clause_idx].findKey(*(mapv[join_clause_idx]), left_row_iter, *pool)
-                        : FindResult();
-
-                    if (find_result.isFound())
-                    {
-                        auto & mapped = find_result.getMapped();
-                        find_results.push_back(find_result);
-                        if (flag_per_row)
-                            addFoundRowAll<Map, false, true>(mapped, selected_rows, current_added_rows, all_flag_known_rows, nullptr);
-                        else
-                            addFoundRowAll<Map, false, false>(mapped, selected_rows, current_added_rows, single_flag_know_rows, nullptr);
-                    }
-                }
-                row_replicate_offset.push_back(current_added_rows);
-            }
-        };
-
-        auto copy_final_matched_rows = [&](size_t left_start_row, ColumnPtr filter_col)
-        {
-            const PaddedPODArray<UInt8> & filter_flags = assert_cast<const ColumnUInt8 &>(*filter_col).getData();
-
-            size_t prev_replicated_row = 0;
-            auto selected_right_row_it = selected_rows.begin();
-            size_t find_result_index = 0;
-            for (size_t i = 1, n = row_replicate_offset.size(); i < n; ++i)
-            {
-                bool any_matched = false;
-                /// For all right join, flag_per_row is true, we need mark used flags for each row.
-                if (flag_per_row)
-                {
-                    for (size_t replicated_row = prev_replicated_row; replicated_row < row_replicate_offset[i]; ++replicated_row)
-                    {
-                        if (filter_flags[replicated_row])
-                        {
-                            any_matched = true;
-                            added_columns.appendFromBlock(*selected_right_row_it->block, selected_right_row_it->row_num, add_missing);
-                            total_added_rows += 1;
-                            if (need_flags)
-                                used_flags.template setUsed<true, true>(selected_right_row_it->block, selected_right_row_it->row_num, 0);
-                        }
-                        ++selected_right_row_it;
-                    }
-                }
-                else
-                {
-                    for (size_t replicated_row = prev_replicated_row; replicated_row < row_replicate_offset[i]; ++replicated_row)
-                    {
-                        if (filter_flags[replicated_row])
-                        {
-                            any_matched = true;
-                            added_columns.appendFromBlock(*selected_right_row_it->block, selected_right_row_it->row_num, add_missing);
-                            total_added_rows += 1;
-                        }
-                        ++selected_right_row_it;
-                    }
-                }
-                if (!any_matched)
-                {
-                    if (add_missing)
-                        addNotFoundRow<true, need_replication>(added_columns, total_added_rows);
-                    else
-                        addNotFoundRow<false, need_replication>(added_columns, total_added_rows);
-                }
-                else
-                {
-                    if (!flag_per_row && need_flags)
-                        used_flags.template setUsed<true, false>(find_results[find_result_index]);
-                    if (need_filter)
-                        setUsed<true>(added_columns.filter, left_start_row + i - 1);
-                    if (add_missing)
-                        added_columns.applyLazyDefaults();
-                }
-                find_result_index += (prev_replicated_row != row_replicate_offset[i]);
-
-                if constexpr (need_replication)
-                {
-                    (*added_columns.offsets_to_replicate)[left_start_row + i - 1] = total_added_rows;
-                }
-                prev_replicated_row = row_replicate_offset[i];
-            }
-        };
-
-        while (left_row_iter < left_block_rows && !exceeded_max_block_rows)
-        {
-            auto left_start_row = left_row_iter;
-            collect_keys_matched_rows_refs();
-            if (selected_rows.size() != current_added_rows || row_replicate_offset.size() != left_row_iter - left_start_row + 1)
-            {
-                throw Exception(
-                    ErrorCodes::LOGICAL_ERROR,
-                    "Sizes are mismatched. selected_rows.size:{}, current_added_rows:{}, row_replicate_offset.size:{}, left_row_iter: {}, "
-                    "left_start_row: {}",
-                    selected_rows.size(),
-                    current_added_rows,
-                    row_replicate_offset.size(),
-                    left_row_iter,
-                    left_start_row);
-            }
-            auto filter_col = buildAdditionalFilter(left_start_row, selected_rows, row_replicate_offset, added_columns);
-            copy_final_matched_rows(left_start_row, filter_col);
-
-            if constexpr (need_replication)
-            {
-                // Add a check for current_added_rows to avoid run the filter expression on too small size batch.
-                if (total_added_rows >= max_joined_block_rows || current_added_rows < 1024)
-                    exceeded_max_block_rows = true;
-            }
-        }
-
-        if constexpr (need_replication)
-        {
-            added_columns.offsets_to_replicate->resize_assume_reserved(left_row_iter);
-            added_columns.filter.resize_assume_reserved(left_row_iter);
-        }
-        added_columns.applyLazyDefaults();
-        return left_row_iter;
-    }
+        bool flag_per_row [[maybe_unused]]);

    /// Cut first num_rows rows from block in place and returns block with remaining rows
-    static Block sliceBlock(Block & block, size_t num_rows)
-    {
-        size_t total_rows = block.rows();
-        if (num_rows >= total_rows)
-            return {};
-        size_t remaining_rows = total_rows - num_rows;
-        Block remaining_block = block.cloneEmpty();
-        for (size_t i = 0; i < block.columns(); ++i)
-        {
-            auto & col = block.getByPosition(i);
-            remaining_block.getByPosition(i).column = col.column->cut(num_rows, remaining_rows);
-            col.column = col.column->cut(0, num_rows);
-        }
-        return remaining_block;
-    }
+    static Block sliceBlock(Block & block, size_t num_rows);

    /** Since we do not store right key columns,
      * this function is used to copy left key columns to right key columns.
@ -864,70 +158,22 @@ private:
        const DataTypePtr & right_key_type,
        const String & renamed_right_column,
        const ColumnWithTypeAndName & left_column,
-        const IColumn::Filter * null_map_filter = nullptr)
-    {
-        ColumnWithTypeAndName right_column = left_column;
-        right_column.name = renamed_right_column;
+        const IColumn::Filter * null_map_filter = nullptr);

-        if (null_map_filter)
-            right_column.column = JoinCommon::filterWithBlanks(right_column.column, *null_map_filter);
+    static void correctNullabilityInplace(ColumnWithTypeAndName & column, bool nullable);

-        bool should_be_nullable = isNullableOrLowCardinalityNullable(right_key_type);
-        if (null_map_filter)
-            correctNullabilityInplace(right_column, should_be_nullable, *null_map_filter);
-        else
-            correctNullabilityInplace(right_column, should_be_nullable);
-
-        if (!right_column.type->equals(*right_key_type))
-        {
-            right_column.column = castColumnAccurate(right_column, right_key_type);
-            right_column.type = right_key_type;
-        }
-
-        right_column.column = right_column.column->convertToFullColumnIfConst();
-        return right_column;
-    }
-
-    static void correctNullabilityInplace(ColumnWithTypeAndName & column, bool nullable)
-    {
-        if (nullable)
-        {
-            JoinCommon::convertColumnToNullable(column);
-        }
-        else
-        {
-            /// We have to replace values masked by NULLs with defaults.
-            if (column.column)
-                if (const auto * nullable_column = checkAndGetColumn<ColumnNullable>(&*column.column))
-                    column.column = JoinCommon::filterWithBlanks(column.column, nullable_column->getNullMapColumn().getData(), true);
-
-            JoinCommon::removeColumnNullability(column);
-        }
-    }
-
-    static void correctNullabilityInplace(ColumnWithTypeAndName & column, bool nullable, const IColumn::Filter & negative_null_map)
-    {
-        if (nullable)
-        {
-            JoinCommon::convertColumnToNullable(column);
-            if (column.type->isNullable() && !negative_null_map.empty())
-            {
-                MutableColumnPtr mutable_column = IColumn::mutate(std::move(column.column));
-                assert_cast<ColumnNullable &>(*mutable_column).applyNegatedNullMap(negative_null_map);
-                column.column = std::move(mutable_column);
-            }
-        }
-        else
-            JoinCommon::removeColumnNullability(column);
-    }
+    static void correctNullabilityInplace(ColumnWithTypeAndName & column, bool nullable, const IColumn::Filter & negative_null_map);
 };

 /// Instantiate template class ahead in different .cpp files to avoid `too large translation unit`.
 extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::RightAny, HashJoin::MapsOne>;
 extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Any, HashJoin::MapsOne>;
+extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Any, HashJoin::MapsAll>;
 extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::All, HashJoin::MapsAll>;
 extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Semi, HashJoin::MapsOne>;
+extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Semi, HashJoin::MapsAll>;
 extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Anti, HashJoin::MapsOne>;
+extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Anti, HashJoin::MapsAll>;
 extern template class HashJoinMethods<JoinKind::Left, JoinStrictness::Asof, HashJoin::MapsAsof>;

 extern template class HashJoinMethods<JoinKind::Right, JoinStrictness::RightAny, HashJoin::MapsOne>;
@ -939,6 +185,7 @@ extern template class HashJoinMethods<JoinKind::Right, JoinStrictness::Asof, Has

 extern template class HashJoinMethods<JoinKind::Inner, JoinStrictness::RightAny, HashJoin::MapsOne>;
 extern template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Any, HashJoin::MapsOne>;
+extern template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Any, HashJoin::MapsAll>;
 extern template class HashJoinMethods<JoinKind::Inner, JoinStrictness::All, HashJoin::MapsAll>;
 extern template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Semi, HashJoin::MapsOne>;
 extern template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Anti, HashJoin::MapsOne>;
--- a/src/Interpreters/HashJoin/HashJoinMethodsImpl.h
+++ b/src/Interpreters/HashJoin/HashJoinMethodsImpl.h
@ -0,0 +1,936 @@
+#pragma once
+#include <Interpreters/HashJoin/HashJoinMethods.h>
+
+namespace DB
+{
+namespace ErrorCodes
+{
+extern const int UNSUPPORTED_JOIN_KEYS;
+extern const int LOGICAL_ERROR;
+}
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::insertFromBlockImpl(
+    HashJoin & join,
+    HashJoin::Type type,
+    MapsTemplate & maps,
+    size_t rows,
+    const ColumnRawPtrs & key_columns,
+    const Sizes & key_sizes,
+    Block * stored_block,
+    ConstNullMapPtr null_map,
+    UInt8ColumnDataPtr join_mask,
+    Arena & pool,
+    bool & is_inserted)
+{
+    switch (type)
+    {
+        case HashJoin::Type::EMPTY:
+            [[fallthrough]];
+        case HashJoin::Type::CROSS:
+            /// Do nothing. We will only save block, and it is enough
+            is_inserted = true;
+            return 0;
+
+#define M(TYPE) \
+    case HashJoin::Type::TYPE: \
+        return insertFromBlockImplTypeCase< \
+            typename KeyGetterForType<HashJoin::Type::TYPE, std::remove_reference_t<decltype(*maps.TYPE)>>::Type>( \
+            join, *maps.TYPE, rows, key_columns, key_sizes, stored_block, null_map, join_mask, pool, is_inserted); \
+        break;
+
+            APPLY_FOR_JOIN_VARIANTS(M)
+#undef M
+    }
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+Block HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::joinBlockImpl(
+    const HashJoin & join, Block & block, const Block & block_with_columns_to_add, const MapsTemplateVector & maps_, bool is_join_get)
+{
+    constexpr JoinFeatures<KIND, STRICTNESS, MapsTemplate> join_features;
+
+    std::vector<JoinOnKeyColumns> join_on_keys;
+    const auto & onexprs = join.table_join->getClauses();
+    for (size_t i = 0; i < onexprs.size(); ++i)
+    {
+        const auto & key_names = !is_join_get ? onexprs[i].key_names_left : onexprs[i].key_names_right;
+        join_on_keys.emplace_back(block, key_names, onexprs[i].condColumnNames().first, join.key_sizes[i]);
+    }
+    size_t existing_columns = block.columns();
+
+    /** If you use FULL or RIGHT JOIN, then the columns from the "left" table must be materialized.
+      * Because if they are constants, then in the "not joined" rows, they may have different values
+      *  - default values, which can differ from the values of these constants.
+      */
+    if constexpr (join_features.right || join_features.full)
+    {
+        materializeBlockInplace(block);
+    }
+
+    /** For LEFT/INNER JOIN, the saved blocks do not contain keys.
+      * For FULL/RIGHT JOIN, the saved blocks contain keys;
+      *  but they will not be used at this stage of joining (and will be in `AdderNonJoined`), and they need to be skipped.
+      * For ASOF, the last column is used as the ASOF column
+      */
+    AddedColumns<!join_features.is_any_join> added_columns(
+        block,
+        block_with_columns_to_add,
+        join.savedBlockSample(),
+        join,
+        std::move(join_on_keys),
+        join.table_join->getMixedJoinExpression(),
+        join_features.is_asof_join,
+        is_join_get);
+
+    bool has_required_right_keys = (join.required_right_keys.columns() != 0);
+    added_columns.need_filter = join_features.need_filter || has_required_right_keys;
+    added_columns.max_joined_block_rows = join.max_joined_block_rows;
+    if (!added_columns.max_joined_block_rows)
+        added_columns.max_joined_block_rows = std::numeric_limits<size_t>::max();
+    else
+        added_columns.reserve(join_features.need_replication);
+
+    size_t num_joined = switchJoinRightColumns(maps_, added_columns, join.data->type, *join.used_flags);
+    /// Do not hold memory for join_on_keys anymore
+    added_columns.join_on_keys.clear();
+    Block remaining_block = sliceBlock(block, num_joined);
+
+    added_columns.buildOutput();
+    for (size_t i = 0; i < added_columns.size(); ++i)
+        block.insert(added_columns.moveColumn(i));
+
+    std::vector<size_t> right_keys_to_replicate [[maybe_unused]];
+
+    if constexpr (join_features.need_filter)
+    {
+        /// If ANY INNER | RIGHT JOIN - filter all the columns except the new ones.
+        for (size_t i = 0; i < existing_columns; ++i)
+            block.safeGetByPosition(i).column = block.safeGetByPosition(i).column->filter(added_columns.filter, -1);
+
+        /// Add join key columns from right block if needed using value from left table because of equality
+        for (size_t i = 0; i < join.required_right_keys.columns(); ++i)
+        {
+            const auto & right_key = join.required_right_keys.getByPosition(i);
+            /// asof column is already in block.
+            if (join_features.is_asof_join && right_key.name == join.table_join->getOnlyClause().key_names_right.back())
+                continue;
+
+            const auto & left_column = block.getByName(join.required_right_keys_sources[i]);
+            const auto & right_col_name = join.getTableJoin().renamedRightColumnName(right_key.name);
+            auto right_col = copyLeftKeyColumnToRight(right_key.type, right_col_name, left_column);
+            block.insert(std::move(right_col));
+        }
+    }
+    else if (has_required_right_keys)
+    {
+        /// Add join key columns from right block if needed.
+        for (size_t i = 0; i < join.required_right_keys.columns(); ++i)
+        {
+            const auto & right_key = join.required_right_keys.getByPosition(i);
+            auto right_col_name = join.getTableJoin().renamedRightColumnName(right_key.name);
+            /// asof column is already in block.
+            if (join_features.is_asof_join && right_key.name == join.table_join->getOnlyClause().key_names_right.back())
+                continue;
+
+            const auto & left_column = block.getByName(join.required_right_keys_sources[i]);
+            auto right_col = copyLeftKeyColumnToRight(right_key.type, right_col_name, left_column, &added_columns.filter);
+            block.insert(std::move(right_col));
+
+            if constexpr (join_features.need_replication)
+                right_keys_to_replicate.push_back(block.getPositionByName(right_col_name));
+        }
+    }
+
+    if constexpr (join_features.need_replication)
+    {
+        std::unique_ptr<IColumn::Offsets> & offsets_to_replicate = added_columns.offsets_to_replicate;
+
+        /// If ALL ... JOIN - we replicate all the columns except the new ones.
+        for (size_t i = 0; i < existing_columns; ++i)
+        {
+            block.safeGetByPosition(i).column = block.safeGetByPosition(i).column->replicate(*offsets_to_replicate);
+        }
+
+        /// Replicate additional right keys
+        for (size_t pos : right_keys_to_replicate)
+        {
+            block.safeGetByPosition(pos).column = block.safeGetByPosition(pos).column->replicate(*offsets_to_replicate);
+        }
+    }
+    return remaining_block;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename KeyGetter, bool is_asof_join>
+KeyGetter HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::createKeyGetter(const ColumnRawPtrs & key_columns, const Sizes & key_sizes)
+{
+    if constexpr (is_asof_join)
+    {
+        auto key_column_copy = key_columns;
+        auto key_size_copy = key_sizes;
+        key_column_copy.pop_back();
+        key_size_copy.pop_back();
+        return KeyGetter(key_column_copy, key_size_copy, nullptr);
+    }
+    else
+        return KeyGetter(key_columns, key_sizes, nullptr);
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename KeyGetter, typename HashMap>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::insertFromBlockImplTypeCase(
+    HashJoin & join,
+    HashMap & map,
+    size_t rows,
+    const ColumnRawPtrs & key_columns,
+    const Sizes & key_sizes,
+    Block * stored_block,
+    ConstNullMapPtr null_map,
+    UInt8ColumnDataPtr join_mask,
+    Arena & pool,
+    bool & is_inserted)
+{
+    [[maybe_unused]] constexpr bool mapped_one = std::is_same_v<typename HashMap::mapped_type, RowRef>;
+    constexpr bool is_asof_join = STRICTNESS == JoinStrictness::Asof;
+
+    const IColumn * asof_column [[maybe_unused]] = nullptr;
+    if constexpr (is_asof_join)
+        asof_column = key_columns.back();
+
+    auto key_getter = createKeyGetter<KeyGetter, is_asof_join>(key_columns, key_sizes);
+
+    /// For ALL and ASOF join always insert values
+    is_inserted = !mapped_one || is_asof_join;
+
+    for (size_t i = 0; i < rows; ++i)
+    {
+        if (null_map && (*null_map)[i])
+        {
+            /// nulls are not inserted into hash table,
+            /// keep them for RIGHT and FULL joins
+            is_inserted = true;
+            continue;
+        }
+
+        /// Check condition for right table from ON section
+        if (join_mask && !(*join_mask)[i])
+            continue;
+
+        if constexpr (is_asof_join)
+            Inserter<HashMap, KeyGetter>::insertAsof(join, map, key_getter, stored_block, i, pool, *asof_column);
+        else if constexpr (mapped_one)
+            is_inserted |= Inserter<HashMap, KeyGetter>::insertOne(join, map, key_getter, stored_block, i, pool);
+        else
+            Inserter<HashMap, KeyGetter>::insertAll(join, map, key_getter, stored_block, i, pool);
+    }
+    return map.getBufferSizeInCells();
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename AddedColumns>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::switchJoinRightColumns(
+    const std::vector<const MapsTemplate *> & mapv,
+    AddedColumns & added_columns,
+    HashJoin::Type type,
+    JoinStuff::JoinUsedFlags & used_flags)
+{
+    constexpr bool is_asof_join = STRICTNESS == JoinStrictness::Asof;
+    switch (type)
+    {
+        case HashJoin::Type::EMPTY: {
+            if constexpr (!is_asof_join)
+            {
+                using KeyGetter = KeyGetterEmpty<typename MapsTemplate::MappedType>;
+                std::vector<KeyGetter> key_getter_vector;
+                key_getter_vector.emplace_back();
+
+                using MapTypeVal = typename KeyGetter::MappedType;
+                std::vector<const MapTypeVal *> a_map_type_vector;
+                a_map_type_vector.emplace_back();
+                return joinRightColumnsSwitchNullability<KeyGetter>(
+                    std::move(key_getter_vector), a_map_type_vector, added_columns, used_flags);
+            }
+            throw Exception(ErrorCodes::UNSUPPORTED_JOIN_KEYS, "Unsupported JOIN keys. Type: {}", type);
+        }
+#define M(TYPE) \
+    case HashJoin::Type::TYPE: { \
+        using MapTypeVal = const typename std::remove_reference_t<decltype(MapsTemplate::TYPE)>::element_type; \
+        using KeyGetter = typename KeyGetterForType<HashJoin::Type::TYPE, MapTypeVal>::Type; \
+        std::vector<const MapTypeVal *> a_map_type_vector(mapv.size()); \
+        std::vector<KeyGetter> key_getter_vector; \
+        for (size_t d = 0; d < added_columns.join_on_keys.size(); ++d) \
+        { \
+            const auto & join_on_key = added_columns.join_on_keys[d]; \
+            a_map_type_vector[d] = mapv[d]->TYPE.get(); \
+            key_getter_vector.push_back( \
+                std::move(createKeyGetter<KeyGetter, is_asof_join>(join_on_key.key_columns, join_on_key.key_sizes))); \
+        } \
+        return joinRightColumnsSwitchNullability<KeyGetter>(std::move(key_getter_vector), a_map_type_vector, added_columns, used_flags); \
+    }
+            APPLY_FOR_JOIN_VARIANTS(M)
+#undef M
+
+        default:
+            throw Exception(ErrorCodes::UNSUPPORTED_JOIN_KEYS, "Unsupported JOIN keys (type: {})", type);
+    }
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename KeyGetter, typename Map, typename AddedColumns>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::joinRightColumnsSwitchNullability(
+    std::vector<KeyGetter> && key_getter_vector,
+    const std::vector<const Map *> & mapv,
+    AddedColumns & added_columns,
+    JoinStuff::JoinUsedFlags & used_flags)
+{
+    if (added_columns.need_filter)
+    {
+        return joinRightColumnsSwitchMultipleDisjuncts<KeyGetter, Map, true>(
+            std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags);
+    }
+    else
+    {
+        return joinRightColumnsSwitchMultipleDisjuncts<KeyGetter, Map, false>(
+            std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags);
+    }
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename KeyGetter, typename Map, bool need_filter, typename AddedColumns>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::joinRightColumnsSwitchMultipleDisjuncts(
+    std::vector<KeyGetter> && key_getter_vector,
+    const std::vector<const Map *> & mapv,
+    AddedColumns & added_columns,
+    JoinStuff::JoinUsedFlags & used_flags)
+{
+    constexpr JoinFeatures<KIND, STRICTNESS, MapsTemplate> join_features;
+    if constexpr (join_features.is_maps_all)
+    {
+        if (added_columns.additional_filter_expression)
+        {
+            bool mark_per_row_used = join_features.right || join_features.full || mapv.size() > 1;
+            return joinRightColumnsWithAddtitionalFilter<KeyGetter, Map>(
+                std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags, need_filter, mark_per_row_used);
+        }
+    }
+
+    if (added_columns.additional_filter_expression)
+        throw Exception(ErrorCodes::LOGICAL_ERROR, "Additional filter expression is not supported for this JOIN");
+
+    return mapv.size() > 1 ? joinRightColumns<KeyGetter, Map, need_filter, true>(
+               std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags)
+                           : joinRightColumns<KeyGetter, Map, need_filter, false>(
+                               std::forward<std::vector<KeyGetter>>(key_getter_vector), mapv, added_columns, used_flags);
+}
+
+
+/// Joins right table columns which indexes are present in right_indexes using specified map.
+/// Makes filter (1 if row presented in right table) and returns offsets to replicate (for ALL JOINS).
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename KeyGetter, typename Map, bool need_filter, bool flag_per_row, typename AddedColumns>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::joinRightColumns(
+    std::vector<KeyGetter> && key_getter_vector,
+    const std::vector<const Map *> & mapv,
+    AddedColumns & added_columns,
+    JoinStuff::JoinUsedFlags & used_flags)
+{
+    constexpr JoinFeatures<KIND, STRICTNESS, MapsTemplate> join_features;
+
+    size_t rows = added_columns.rows_to_add;
+    if constexpr (need_filter)
+        added_columns.filter = IColumn::Filter(rows, 0);
+
+    Arena pool;
+
+    if constexpr (join_features.need_replication)
+        added_columns.offsets_to_replicate = std::make_unique<IColumn::Offsets>(rows);
+
+    IColumn::Offset current_offset = 0;
+    size_t max_joined_block_rows = added_columns.max_joined_block_rows;
+    size_t i = 0;
+    for (; i < rows; ++i)
+    {
+        if constexpr (join_features.need_replication)
+        {
+            if (unlikely(current_offset >= max_joined_block_rows))
+            {
+                added_columns.offsets_to_replicate->resize_assume_reserved(i);
+                added_columns.filter.resize_assume_reserved(i);
+                break;
+            }
+        }
+
+        bool right_row_found = false;
+        KnownRowsHolder<flag_per_row> known_rows;
+        for (size_t onexpr_idx = 0; onexpr_idx < added_columns.join_on_keys.size(); ++onexpr_idx)
+        {
+            const auto & join_keys = added_columns.join_on_keys[onexpr_idx];
+            if (join_keys.null_map && (*join_keys.null_map)[i])
+                continue;
+
+            bool row_acceptable = !join_keys.isRowFiltered(i);
+            using FindResult = typename KeyGetter::FindResult;
+            auto find_result = row_acceptable ? key_getter_vector[onexpr_idx].findKey(*(mapv[onexpr_idx]), i, pool) : FindResult();
+
+            if (find_result.isFound())
+            {
+                right_row_found = true;
+                auto & mapped = find_result.getMapped();
+                if constexpr (join_features.is_asof_join)
+                {
+                    const IColumn & left_asof_key = added_columns.leftAsofKey();
+
+                    auto row_ref = mapped->findAsof(left_asof_key, i);
+                    if (row_ref.block)
+                    {
+                        setUsed<need_filter>(added_columns.filter, i);
+                        if constexpr (flag_per_row)
+                            used_flags.template setUsed<join_features.need_flags, flag_per_row>(row_ref.block, row_ref.row_num, 0);
+                        else
+                            used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
+
+                        added_columns.appendFromBlock(*row_ref.block, row_ref.row_num, join_features.add_missing);
+                    }
+                    else
+                        addNotFoundRow<join_features.add_missing, join_features.need_replication>(added_columns, current_offset);
+                }
+                else if constexpr (join_features.is_all_join)
+                {
+                    setUsed<need_filter>(added_columns.filter, i);
+                    used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
+                    auto used_flags_opt = join_features.need_flags ? &used_flags : nullptr;
+                    addFoundRowAll<Map, join_features.add_missing>(mapped, added_columns, current_offset, known_rows, used_flags_opt);
+                }
+                else if constexpr ((join_features.is_any_join || join_features.is_semi_join) && join_features.right)
+                {
+                    /// Use first appeared left key + it needs left columns replication
+                    bool used_once = used_flags.template setUsedOnce<join_features.need_flags, flag_per_row>(find_result);
+                    if (used_once)
+                    {
+                        auto used_flags_opt = join_features.need_flags ? &used_flags : nullptr;
+                        setUsed<need_filter>(added_columns.filter, i);
+                        addFoundRowAll<Map, join_features.add_missing>(mapped, added_columns, current_offset, known_rows, used_flags_opt);
+                    }
+                }
+                else if constexpr (join_features.is_any_join && join_features.inner)
+                {
+                    bool used_once = used_flags.template setUsedOnce<join_features.need_flags, flag_per_row>(find_result);
+
+                    /// Use first appeared left key only
+                    if (used_once)
+                    {
+                        setUsed<need_filter>(added_columns.filter, i);
+                        added_columns.appendFromBlock(*mapped.block, mapped.row_num, join_features.add_missing);
+                    }
+
+                    break;
+                }
+                else if constexpr (join_features.is_any_join && join_features.full)
+                {
+                    /// TODO
+                }
+                else if constexpr (join_features.is_anti_join)
+                {
+                    if constexpr (join_features.right && join_features.need_flags)
+                        used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
+                }
+                else /// ANY LEFT, SEMI LEFT, old ANY (RightAny)
+                {
+                    setUsed<need_filter>(added_columns.filter, i);
+                    used_flags.template setUsed<join_features.need_flags, flag_per_row>(find_result);
+                    added_columns.appendFromBlock(*mapped.block, mapped.row_num, join_features.add_missing);
+
+                    if (join_features.is_any_or_semi_join)
+                    {
+                        break;
+                    }
+                }
+            }
+        }
+
+        if (!right_row_found)
+        {
+            if constexpr (join_features.is_anti_join && join_features.left)
+                setUsed<need_filter>(added_columns.filter, i);
+            addNotFoundRow<join_features.add_missing, join_features.need_replication>(added_columns, current_offset);
+        }
+
+        if constexpr (join_features.need_replication)
+        {
+            (*added_columns.offsets_to_replicate)[i] = current_offset;
+        }
+    }
+
+    added_columns.applyLazyDefaults();
+    return i;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <bool need_filter>
+void HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::setUsed(IColumn::Filter & filter [[maybe_unused]], size_t pos [[maybe_unused]])
+{
+    if constexpr (need_filter)
+        filter[pos] = 1;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename AddedColumns>
+ColumnPtr HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::buildAdditionalFilter(
+    size_t left_start_row,
+    const std::vector<RowRef> & selected_rows,
+    const std::vector<size_t> & row_replicate_offset,
+    AddedColumns & added_columns)
+{
+    ColumnPtr result_column;
+    do
+    {
+        if (selected_rows.empty())
+        {
+            result_column = ColumnUInt8::create();
+            break;
+        }
+        const Block & sample_right_block = *selected_rows.begin()->block;
+        if (!sample_right_block || !added_columns.additional_filter_expression)
+        {
+            auto filter = ColumnUInt8::create();
+            filter->insertMany(1, selected_rows.size());
+            result_column = std::move(filter);
+            break;
+        }
+
+        auto required_cols = added_columns.additional_filter_expression->getRequiredColumnsWithTypes();
+        if (required_cols.empty())
+        {
+            Block block;
+            added_columns.additional_filter_expression->execute(block);
+            result_column = block.getByPosition(0).column->cloneResized(selected_rows.size());
+            break;
+        }
+        NameSet required_column_names;
+        for (auto & col : required_cols)
+            required_column_names.insert(col.name);
+
+        Block executed_block;
+        size_t right_col_pos = 0;
+        for (const auto & col : sample_right_block.getColumnsWithTypeAndName())
+        {
+            if (required_column_names.contains(col.name))
+            {
+                auto new_col = col.column->cloneEmpty();
+                for (const auto & selected_row : selected_rows)
+                {
+                    const auto & src_col = selected_row.block->getByPosition(right_col_pos);
+                    new_col->insertFrom(*src_col.column, selected_row.row_num);
+                }
+                executed_block.insert({std::move(new_col), col.type, col.name});
+            }
+            right_col_pos += 1;
+        }
+        if (!executed_block)
+        {
+            result_column = ColumnUInt8::create();
+            break;
+        }
+
+        for (const auto & col_name : required_column_names)
+        {
+            const auto * src_col = added_columns.left_block.findByName(col_name);
+            if (!src_col)
+                continue;
+            auto new_col = src_col->column->cloneEmpty();
+            size_t prev_left_offset = 0;
+            for (size_t i = 1; i < row_replicate_offset.size(); ++i)
+            {
+                const size_t & left_offset = row_replicate_offset[i];
+                size_t rows = left_offset - prev_left_offset;
+                if (rows)
+                    new_col->insertManyFrom(*src_col->column, left_start_row + i - 1, rows);
+                prev_left_offset = left_offset;
+            }
+            executed_block.insert({std::move(new_col), src_col->type, col_name});
+        }
+        if (!executed_block)
+        {
+            throw Exception(
+                ErrorCodes::LOGICAL_ERROR,
+                "required columns: [{}], but not found any in left/right table. right table: {}, left table: {}",
+                required_cols.toString(),
+                sample_right_block.dumpNames(),
+                added_columns.left_block.dumpNames());
+        }
+
+        for (const auto & col : executed_block.getColumnsWithTypeAndName())
+            if (!col.column || !col.type)
+                throw Exception(ErrorCodes::LOGICAL_ERROR, "Illegal nullptr column in input block: {}", executed_block.dumpStructure());
+
+        added_columns.additional_filter_expression->execute(executed_block);
+        result_column = executed_block.getByPosition(0).column->convertToFullColumnIfConst();
+        executed_block.clear();
+    } while (false);
+
+    result_column = result_column->convertToFullIfNeeded();
+    if (result_column->isNullable())
+    {
+        /// Convert Nullable(UInt8) to UInt8 ensuring that nulls are zeros
+        /// Trying to avoid copying data, since we are the only owner of the column.
+        ColumnPtr mask_column = assert_cast<const ColumnNullable &>(*result_column).getNullMapColumnPtr();
+
+        MutableColumnPtr mutable_column;
+        {
+            ColumnPtr nested_column = assert_cast<const ColumnNullable &>(*result_column).getNestedColumnPtr();
+            result_column.reset();
+            mutable_column = IColumn::mutate(std::move(nested_column));
+        }
+
+        auto & column_data = assert_cast<ColumnUInt8 &>(*mutable_column).getData();
+        const auto & mask_column_data = assert_cast<const ColumnUInt8 &>(*mask_column).getData();
+        for (size_t i = 0; i < column_data.size(); ++i)
+        {
+            if (mask_column_data[i])
+                column_data[i] = 0;
+        }
+        return mutable_column;
+    }
+    return result_column;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+template <typename KeyGetter, typename Map, typename AddedColumns>
+size_t HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::joinRightColumnsWithAddtitionalFilter(
+    std::vector<KeyGetter> && key_getter_vector,
+    const std::vector<const Map *> & mapv,
+    AddedColumns & added_columns,
+    JoinStuff::JoinUsedFlags & used_flags [[maybe_unused]],
+    bool need_filter [[maybe_unused]],
+    bool flag_per_row [[maybe_unused]])
+{
+    constexpr JoinFeatures<KIND, STRICTNESS, MapsTemplate> join_features;
+    size_t left_block_rows = added_columns.rows_to_add;
+    if (need_filter)
+        added_columns.filter = IColumn::Filter(left_block_rows, 0);
+
+    std::unique_ptr<Arena> pool;
+
+    if constexpr (join_features.need_replication)
+        added_columns.offsets_to_replicate = std::make_unique<IColumn::Offsets>(left_block_rows);
+
+    std::vector<size_t> row_replicate_offset;
+    row_replicate_offset.reserve(left_block_rows);
+
+    using FindResult = typename KeyGetter::FindResult;
+    size_t max_joined_block_rows = added_columns.max_joined_block_rows;
+    size_t left_row_iter = 0;
+    PreSelectedRows selected_rows;
+    selected_rows.reserve(left_block_rows);
+    std::vector<FindResult> find_results;
+    find_results.reserve(left_block_rows);
+    bool exceeded_max_block_rows = false;
+    IColumn::Offset total_added_rows = 0;
+    IColumn::Offset current_added_rows = 0;
+
+    auto collect_keys_matched_rows_refs = [&]()
+    {
+        pool = std::make_unique<Arena>();
+        find_results.clear();
+        row_replicate_offset.clear();
+        row_replicate_offset.push_back(0);
+        current_added_rows = 0;
+        selected_rows.clear();
+        for (; left_row_iter < left_block_rows; ++left_row_iter)
+        {
+            if constexpr (join_features.need_replication)
+            {
+                if (unlikely(total_added_rows + current_added_rows >= max_joined_block_rows))
+                {
+                    break;
+                }
+            }
+            KnownRowsHolder<true> all_flag_known_rows;
+            KnownRowsHolder<false> single_flag_know_rows;
+            for (size_t join_clause_idx = 0; join_clause_idx < added_columns.join_on_keys.size(); ++join_clause_idx)
+            {
+                const auto & join_keys = added_columns.join_on_keys[join_clause_idx];
+                if (join_keys.null_map && (*join_keys.null_map)[left_row_iter])
+                    continue;
+
+                bool row_acceptable = !join_keys.isRowFiltered(left_row_iter);
+                auto find_result = row_acceptable
+                    ? key_getter_vector[join_clause_idx].findKey(*(mapv[join_clause_idx]), left_row_iter, *pool)
+                    : FindResult();
+
+                if (find_result.isFound())
+                {
+                    auto & mapped = find_result.getMapped();
+                    find_results.push_back(find_result);
+                    /// We don't add missing in addFoundRowAll here. we will add it after filter is applied.
+                    /// it's different from `joinRightColumns`.
+                    if (flag_per_row)
+                        addFoundRowAll<Map, false, true>(mapped, selected_rows, current_added_rows, all_flag_known_rows, nullptr);
+                    else
+                        addFoundRowAll<Map, false, false>(mapped, selected_rows, current_added_rows, single_flag_know_rows, nullptr);
+                }
+            }
+            row_replicate_offset.push_back(current_added_rows);
+        }
+    };
+
+    auto copy_final_matched_rows = [&](size_t left_start_row, ColumnPtr filter_col)
+    {
+        const PaddedPODArray<UInt8> & filter_flags = assert_cast<const ColumnUInt8 &>(*filter_col).getData();
+
+        size_t prev_replicated_row = 0;
+        auto selected_right_row_it = selected_rows.begin();
+        size_t find_result_index = 0;
+        for (size_t i = 1, n = row_replicate_offset.size(); i < n; ++i)
+        {
+            bool any_matched = false;
+            /// right/full join or multiple disjuncts, we need to mark used flags for each row.
+            if (flag_per_row)
+            {
+                for (size_t replicated_row = prev_replicated_row; replicated_row < row_replicate_offset[i]; ++replicated_row)
+                {
+                    if (filter_flags[replicated_row])
+                    {
+                        if constexpr (join_features.is_semi_join || join_features.is_any_join)
+                        {
+                            /// For LEFT/INNER SEMI/ANY JOIN, we need to add only first appeared row from left,
+                            if constexpr (join_features.left || join_features.inner)
+                            {
+                                if (!any_matched)
+                                {
+                                    // For inner join, we need mark each right row'flag, because we only use each right row once.
+                                    auto used_once = used_flags.template setUsedOnce<join_features.need_flags, true>(
+                                        selected_right_row_it->block, selected_right_row_it->row_num, 0);
+                                    if (used_once)
+                                    {
+                                        any_matched = true;
+                                        total_added_rows += 1;
+                                        added_columns.appendFromBlock(
+                                            *selected_right_row_it->block, selected_right_row_it->row_num, join_features.add_missing);
+                                    }
+                                }
+                            }
+                            else
+                            {
+                                auto used_once = used_flags.template setUsedOnce<join_features.need_flags, true>(
+                                    selected_right_row_it->block, selected_right_row_it->row_num, 0);
+                                if (used_once)
+                                {
+                                    any_matched = true;
+                                    total_added_rows += 1;
+                                    added_columns.appendFromBlock(
+                                        *selected_right_row_it->block, selected_right_row_it->row_num, join_features.add_missing);
+                                }
+                            }
+                        }
+                        else if constexpr (join_features.is_anti_join)
+                        {
+                            any_matched = true;
+                            if constexpr (join_features.right && join_features.need_flags)
+                                used_flags.template setUsed<true, true>(selected_right_row_it->block, selected_right_row_it->row_num, 0);
+                        }
+                        else
+                        {
+                            any_matched = true;
+                            total_added_rows += 1;
+                            added_columns.appendFromBlock(
+                                *selected_right_row_it->block, selected_right_row_it->row_num, join_features.add_missing);
+                            used_flags.template setUsed<join_features.need_flags, true>(
+                                selected_right_row_it->block, selected_right_row_it->row_num, 0);
+                        }
+                    }
+
+                    ++selected_right_row_it;
+                }
+            }
+            else
+            {
+                for (size_t replicated_row = prev_replicated_row; replicated_row < row_replicate_offset[i]; ++replicated_row)
+                {
+                    if constexpr (join_features.is_anti_join)
+                    {
+                        any_matched |= filter_flags[replicated_row];
+                    }
+                    else if constexpr (join_features.need_replication)
+                    {
+                        if (filter_flags[replicated_row])
+                        {
+                            any_matched = true;
+                            added_columns.appendFromBlock(
+                                *selected_right_row_it->block, selected_right_row_it->row_num, join_features.add_missing);
+                            total_added_rows += 1;
+                        }
+                        ++selected_right_row_it;
+                    }
+                    else
+                    {
+                        if (filter_flags[replicated_row])
+                        {
+                            any_matched = true;
+                            added_columns.appendFromBlock(
+                                *selected_right_row_it->block, selected_right_row_it->row_num, join_features.add_missing);
+                            total_added_rows += 1;
+                            selected_right_row_it = selected_right_row_it + row_replicate_offset[i] - replicated_row;
+                            break;
+                        }
+                        else
+                            ++selected_right_row_it;
+                    }
+                }
+            }
+
+
+            if constexpr (join_features.is_anti_join)
+            {
+                if (!any_matched)
+                {
+                    if constexpr (join_features.left)
+                        if (need_filter)
+                            setUsed<true>(added_columns.filter, left_start_row + i - 1);
+                    addNotFoundRow<join_features.add_missing, join_features.need_replication>(added_columns, total_added_rows);
+                }
+            }
+            else
+            {
+                if (!any_matched)
+                {
+                    addNotFoundRow<join_features.add_missing, join_features.need_replication>(added_columns, total_added_rows);
+                }
+                else
+                {
+                    if (!flag_per_row)
+                        used_flags.template setUsed<join_features.need_flags, false>(find_results[find_result_index]);
+                    if (need_filter)
+                        setUsed<true>(added_columns.filter, left_start_row + i - 1);
+                    if constexpr (join_features.add_missing)
+                        added_columns.applyLazyDefaults();
+                }
+            }
+            find_result_index += (prev_replicated_row != row_replicate_offset[i]);
+
+            if constexpr (join_features.need_replication)
+            {
+                (*added_columns.offsets_to_replicate)[left_start_row + i - 1] = total_added_rows;
+            }
+            prev_replicated_row = row_replicate_offset[i];
+        }
+    };
+
+    while (left_row_iter < left_block_rows && !exceeded_max_block_rows)
+    {
+        auto left_start_row = left_row_iter;
+        collect_keys_matched_rows_refs();
+        if (selected_rows.size() != current_added_rows || row_replicate_offset.size() != left_row_iter - left_start_row + 1)
+        {
+            throw Exception(
+                ErrorCodes::LOGICAL_ERROR,
+                "Sizes are mismatched. selected_rows.size:{}, current_added_rows:{}, row_replicate_offset.size:{}, left_row_iter: {}, "
+                "left_start_row: {}",
+                selected_rows.size(),
+                current_added_rows,
+                row_replicate_offset.size(),
+                left_row_iter,
+                left_start_row);
+        }
+        auto filter_col = buildAdditionalFilter(left_start_row, selected_rows, row_replicate_offset, added_columns);
+        copy_final_matched_rows(left_start_row, filter_col);
+
+        if constexpr (join_features.need_replication)
+        {
+            // Add a check for current_added_rows to avoid run the filter expression on too small size batch.
+            if (total_added_rows >= max_joined_block_rows || current_added_rows < 1024)
+                exceeded_max_block_rows = true;
+        }
+    }
+
+    if constexpr (join_features.need_replication)
+    {
+        added_columns.offsets_to_replicate->resize_assume_reserved(left_row_iter);
+        added_columns.filter.resize_assume_reserved(left_row_iter);
+    }
+    added_columns.applyLazyDefaults();
+    return left_row_iter;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+Block HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::sliceBlock(Block & block, size_t num_rows)
+{
+    size_t total_rows = block.rows();
+    if (num_rows >= total_rows)
+        return {};
+    size_t remaining_rows = total_rows - num_rows;
+    Block remaining_block = block.cloneEmpty();
+    for (size_t i = 0; i < block.columns(); ++i)
+    {
+        auto & col = block.getByPosition(i);
+        remaining_block.getByPosition(i).column = col.column->cut(num_rows, remaining_rows);
+        col.column = col.column->cut(0, num_rows);
+    }
+    return remaining_block;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+ColumnWithTypeAndName HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::copyLeftKeyColumnToRight(
+    const DataTypePtr & right_key_type,
+    const String & renamed_right_column,
+    const ColumnWithTypeAndName & left_column,
+    const IColumn::Filter * null_map_filter)
+{
+    ColumnWithTypeAndName right_column = left_column;
+    right_column.name = renamed_right_column;
+
+    if (null_map_filter)
+        right_column.column = JoinCommon::filterWithBlanks(right_column.column, *null_map_filter);
+
+    bool should_be_nullable = isNullableOrLowCardinalityNullable(right_key_type);
+    if (null_map_filter)
+        correctNullabilityInplace(right_column, should_be_nullable, *null_map_filter);
+    else
+        correctNullabilityInplace(right_column, should_be_nullable);
+
+    if (!right_column.type->equals(*right_key_type))
+    {
+        right_column.column = castColumnAccurate(right_column, right_key_type);
+        right_column.type = right_key_type;
+    }
+
+    right_column.column = right_column.column->convertToFullColumnIfConst();
+    return right_column;
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+void HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::correctNullabilityInplace(ColumnWithTypeAndName & column, bool nullable)
+{
+    if (nullable)
+    {
+        JoinCommon::convertColumnToNullable(column);
+    }
+    else
+    {
+        /// We have to replace values masked by NULLs with defaults.
+        if (column.column)
+            if (const auto * nullable_column = checkAndGetColumn<ColumnNullable>(&*column.column))
+                column.column = JoinCommon::filterWithBlanks(column.column, nullable_column->getNullMapColumn().getData(), true);
+
+        JoinCommon::removeColumnNullability(column);
+    }
+}
+
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename MapsTemplate>
+void HashJoinMethods<KIND, STRICTNESS, MapsTemplate>::correctNullabilityInplace(
+    ColumnWithTypeAndName & column, bool nullable, const IColumn::Filter & negative_null_map)
+{
+    if (nullable)
+    {
+        JoinCommon::convertColumnToNullable(column);
+        if (column.type->isNullable() && !negative_null_map.empty())
+        {
+            MutableColumnPtr mutable_column = IColumn::mutate(std::move(column.column));
+            assert_cast<ColumnNullable &>(*mutable_column).applyNegatedNullMap(negative_null_map);
+            column.column = std::move(mutable_column);
+        }
+    }
+    else
+        JoinCommon::removeColumnNullability(column);
+}
+}
--- a/src/Interpreters/HashJoin/InnerHashJoin.cpp
+++ b/src/Interpreters/HashJoin/InnerHashJoin.cpp
@ -1,10 +1,11 @@

-#include <Interpreters/HashJoin/HashJoinMethods.h>
+#include <Interpreters/HashJoin/HashJoinMethodsImpl.h>

 namespace DB
 {
 template class HashJoinMethods<JoinKind::Inner, JoinStrictness::RightAny, HashJoin::MapsOne>;
 template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Any, HashJoin::MapsOne>;
+template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Any, HashJoin::MapsAll>;
 template class HashJoinMethods<JoinKind::Inner, JoinStrictness::All, HashJoin::MapsAll>;
 template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Semi, HashJoin::MapsOne>;
 template class HashJoinMethods<JoinKind::Inner, JoinStrictness::Anti, HashJoin::MapsOne>;
--- a/src/Interpreters/HashJoin/JoinFeatures.h
+++ b/src/Interpreters/HashJoin/JoinFeatures.h
@ -3,15 +3,15 @@
 #include <Interpreters/joinDispatch.h>
 namespace DB
 {
-template <JoinKind KIND, JoinStrictness STRICTNESS>
+template <JoinKind KIND, JoinStrictness STRICTNESS, typename Map>
 struct JoinFeatures
 {
    static constexpr bool is_any_join = STRICTNESS == JoinStrictness::Any;
-    static constexpr bool is_any_or_semi_join = STRICTNESS == JoinStrictness::Any || STRICTNESS == JoinStrictness::RightAny || (STRICTNESS == JoinStrictness::Semi && KIND == JoinKind::Left);
    static constexpr bool is_all_join = STRICTNESS == JoinStrictness::All;
    static constexpr bool is_asof_join = STRICTNESS == JoinStrictness::Asof;
    static constexpr bool is_semi_join = STRICTNESS == JoinStrictness::Semi;
    static constexpr bool is_anti_join = STRICTNESS == JoinStrictness::Anti;
+    static constexpr bool is_any_or_semi_join = is_any_join || STRICTNESS == JoinStrictness::RightAny || (is_semi_join && KIND == JoinKind::Left);

    static constexpr bool left = KIND == JoinKind::Left;
    static constexpr bool right = KIND == JoinKind::Right;
@ -22,7 +22,8 @@ struct JoinFeatures
    static constexpr bool need_filter = !need_replication && (inner || right || (is_semi_join && left) || (is_anti_join && left));
    static constexpr bool add_missing = (left || full) && !is_semi_join;

-    static constexpr bool need_flags = MapGetter<KIND, STRICTNESS>::flagged;
+    static constexpr bool need_flags = MapGetter<KIND, STRICTNESS, std::is_same_v<std::decay_t<Map>, HashJoin::MapsAll>>::flagged;
+    static constexpr bool is_maps_all = std::is_same_v<std::decay_t<Map>, HashJoin::MapsAll>;
 };

 }
--- a/src/Interpreters/HashJoin/JoinUsedFlags.h
+++ b/src/Interpreters/HashJoin/JoinUsedFlags.h
@ -26,10 +26,10 @@ public:
    /// Update size for vector with flags.
    /// Calling this method invalidates existing flags.
    /// It can be called several times, but all of them should happen before using this structure.
-    template <JoinKind KIND, JoinStrictness STRICTNESS>
+    template <JoinKind KIND, JoinStrictness STRICTNESS, bool prefer_use_maps_all>
    void reinit(size_t size)
    {
-        if constexpr (MapGetter<KIND, STRICTNESS>::flagged)
+        if constexpr (MapGetter<KIND, STRICTNESS, prefer_use_maps_all>::flagged)
        {
            assert(flags[nullptr].size() <= size);
            need_flags = true;
@ -43,10 +43,10 @@ public:
        }
    }

-    template <JoinKind KIND, JoinStrictness STRICTNESS>
+    template <JoinKind KIND, JoinStrictness STRICTNESS, bool prefer_use_maps_all>
    void reinit(const Block * block_ptr)
    {
-        if constexpr (MapGetter<KIND, STRICTNESS>::flagged)
+        if constexpr (MapGetter<KIND, STRICTNESS, prefer_use_maps_all>::flagged)
        {
            assert(flags[block_ptr].size() <= block_ptr->rows());
            need_flags = true;
@ -148,6 +148,31 @@ public:
        }

    }
+    template <bool use_flags, bool flag_per_row>
+    bool setUsedOnce(const Block * block, size_t row_num, size_t offset)
+    {
+        if constexpr (!use_flags)
+            return true;
+
+        if constexpr (flag_per_row)
+        {
+            /// fast check to prevent heavy CAS with seq_cst order
+            if (flags[block][row_num].load(std::memory_order_relaxed))
+                return false;
+
+            bool expected = false;
+            return flags[block][row_num].compare_exchange_strong(expected, true);
+        }
+        else
+        {
+            /// fast check to prevent heavy CAS with seq_cst order
+            if (flags[nullptr][offset].load(std::memory_order_relaxed))
+                return false;
+
+            bool expected = false;
+            return flags[nullptr][offset].compare_exchange_strong(expected, true);
+        }
+    }
 };

 }
--- a/src/Interpreters/HashJoin/LeftHashJoin.cpp
+++ b/src/Interpreters/HashJoin/LeftHashJoin.cpp
@ -1,11 +1,14 @@
-#include <Interpreters/HashJoin/HashJoinMethods.h>
+#include <Interpreters/HashJoin/HashJoinMethodsImpl.h>

 namespace DB
 {
 template class HashJoinMethods<JoinKind::Left, JoinStrictness::RightAny, HashJoin::MapsOne>;
 template class HashJoinMethods<JoinKind::Left, JoinStrictness::Any, HashJoin::MapsOne>;
+template class HashJoinMethods<JoinKind::Left, JoinStrictness::Any, HashJoin::MapsAll>;
 template class HashJoinMethods<JoinKind::Left, JoinStrictness::All, HashJoin::MapsAll>;
 template class HashJoinMethods<JoinKind::Left, JoinStrictness::Semi, HashJoin::MapsOne>;
+template class HashJoinMethods<JoinKind::Left, JoinStrictness::Semi, HashJoin::MapsAll>;
 template class HashJoinMethods<JoinKind::Left, JoinStrictness::Anti, HashJoin::MapsOne>;
+template class HashJoinMethods<JoinKind::Left, JoinStrictness::Anti, HashJoin::MapsAll>;
 template class HashJoinMethods<JoinKind::Left, JoinStrictness::Asof, HashJoin::MapsAsof>;
 }
--- a/src/Interpreters/HashJoin/RightHashJoin.cpp
+++ b/src/Interpreters/HashJoin/RightHashJoin.cpp
@ -1,4 +1,4 @@
-#include <Interpreters/HashJoin/HashJoinMethods.h>
+#include <Interpreters/HashJoin/HashJoinMethodsImpl.h>

 namespace DB
 {
--- a/src/Interpreters/MutationsInterpreter.cpp
+++ b/src/Interpreters/MutationsInterpreter.cpp
@ -146,7 +146,6 @@ ColumnDependencies getAllColumnDependencies(


 bool isStorageTouchedByMutations(
-    MergeTreeData & storage,
    MergeTreeData::DataPartPtr source_part,
    const StorageMetadataPtr & metadata_snapshot,
    const std::vector<MutationCommand> & commands,
@ -155,7 +154,9 @@ bool isStorageTouchedByMutations(
    if (commands.empty())
        return false;

+    auto storage_from_part = std::make_shared<StorageFromMergeTreeDataPart>(source_part);
    bool all_commands_can_be_skipped = true;
+
    for (const auto & command : commands)
    {
        if (command.type == MutationCommand::APPLY_DELETED_MASK)
@ -170,7 +171,7 @@ bool isStorageTouchedByMutations(

            if (command.partition)
            {
-                const String partition_id = storage.getPartitionIDFromQuery(command.partition, context);
+                const String partition_id = storage_from_part->getPartitionIDFromQuery(command.partition, context);
                if (partition_id == source_part->info.partition_id)
                    all_commands_can_be_skipped = false;
            }
@ -184,20 +185,18 @@ bool isStorageTouchedByMutations(
    if (all_commands_can_be_skipped)
        return false;

-    auto storage_from_part = std::make_shared<StorageFromMergeTreeDataPart>(source_part);
-
    std::optional<InterpreterSelectQuery> interpreter_select_query;
    BlockIO io;

    if (context->getSettingsRef().allow_experimental_analyzer)
    {
-        auto select_query_tree = prepareQueryAffectedQueryTree(commands, storage.shared_from_this(), context);
+        auto select_query_tree = prepareQueryAffectedQueryTree(commands, storage_from_part, context);
        InterpreterSelectQueryAnalyzer interpreter(select_query_tree, context, SelectQueryOptions().ignoreLimits());
        io = interpreter.execute();
    }
    else
    {
-        ASTPtr select_query = prepareQueryAffectedAST(commands, storage.shared_from_this(), context);
+        ASTPtr select_query = prepareQueryAffectedAST(commands, storage_from_part, context);
        /// Interpreter must be alive, when we use result of execute() method.
        /// For some reason it may copy context and give it into ExpressionTransform
        /// after that we will use context from destroyed stack frame in our stream.
--- a/src/Interpreters/MutationsInterpreter.h
+++ b/src/Interpreters/MutationsInterpreter.h
@ -19,7 +19,6 @@ using QueryPipelineBuilderPtr = std::unique_ptr<QueryPipelineBuilder>;

 /// Return false if the data isn't going to be changed by mutations.
 bool isStorageTouchedByMutations(
-    MergeTreeData & storage,
    MergeTreeData::DataPartPtr source_part,
    const StorageMetadataPtr & metadata_snapshot,
    const std::vector<MutationCommand> & commands,
--- a/src/Interpreters/castColumn.cpp
+++ b/src/Interpreters/castColumn.cpp
@ -26,11 +26,9 @@ static ColumnPtr castColumn(CastType cast_type, const ColumnWithTypeAndName & ar
            ""
        }
    };
-    auto get_cast_func = [cast_type, &arguments]
+    auto get_cast_func = [from = arg, to = type, cast_type]
    {
-
-        FunctionOverloadResolverPtr func_builder_cast = createInternalCastOverloadResolver(cast_type, {});
-        return func_builder_cast->build(arguments);
+        return createInternalCast(from, to, cast_type, {});
    };

    FunctionBasePtr func_cast = cache ? cache->getOrSet(cast_type, from_name, to_name, std::move(get_cast_func)) : get_cast_func();
--- a/src/Interpreters/joinDispatch.h
+++ b/src/Interpreters/joinDispatch.h
@ -12,38 +12,53 @@
 namespace DB
 {

-template <JoinKind kind, JoinStrictness join_strictness>
+/// HashJoin::MapsOne is more efficient, it only store one row for each key in the map. It is recommended to use it whenever possible.
+/// When only need to match only one row from right table, use HashJoin::MapsOne. For example, LEFT ANY/SEMI/ANTI.
+///
+/// HashJoin::MapsAll will store all rows for each key in the map. It is used when need to match multiple rows from right table.
+/// For example, LEFT ALL, INNER ALL, RIGHT ALL/ANY.
+///
+/// prefer_use_maps_all is true when there is mixed inequal condition in the join condition. For example, `t1.a = t2.a AND t1.b > t2.b`.
+/// In this case, we need to use HashJoin::MapsAll to store all rows for each key in the map. We will select all matched rows from the map
+/// and filter them by `t1.b > t2.b`.
+///
+/// flagged indicates whether we need to store flags for each row whether it has been used in the join. See JoinUsedFlags.h.
+template <JoinKind kind, JoinStrictness join_strictness, bool prefer_use_maps_all>
 struct MapGetter;

-template <> struct MapGetter<JoinKind::Left, JoinStrictness::RightAny>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Inner, JoinStrictness::RightAny> { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Right, JoinStrictness::RightAny> { using Map = HashJoin::MapsOne; static constexpr bool flagged = true; };
-template <> struct MapGetter<JoinKind::Full, JoinStrictness::RightAny>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Left, JoinStrictness::RightAny, prefer_use_maps_all>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Inner, JoinStrictness::RightAny, prefer_use_maps_all> { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Right, JoinStrictness::RightAny, prefer_use_maps_all> { using Map = HashJoin::MapsOne; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Full, JoinStrictness::RightAny, prefer_use_maps_all>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = true; };

-template <> struct MapGetter<JoinKind::Left, JoinStrictness::Any>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Inner, JoinStrictness::Any> { using Map = HashJoin::MapsOne; static constexpr bool flagged = true; };
-template <> struct MapGetter<JoinKind::Right, JoinStrictness::Any> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
-template <> struct MapGetter<JoinKind::Full, JoinStrictness::Any>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <> struct MapGetter<JoinKind::Left, JoinStrictness::Any, false>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <> struct MapGetter<JoinKind::Left, JoinStrictness::Any, true>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = false; };
+template <> struct MapGetter<JoinKind::Inner, JoinStrictness::Any, true> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <> struct MapGetter<JoinKind::Inner, JoinStrictness::Any, false> { using Map = HashJoin::MapsOne; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Right, JoinStrictness::Any, prefer_use_maps_all> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Full, JoinStrictness::Any, prefer_use_maps_all>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };

-template <> struct MapGetter<JoinKind::Left, JoinStrictness::All>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Inner, JoinStrictness::All> { using Map = HashJoin::MapsAll; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Right, JoinStrictness::All> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
-template <> struct MapGetter<JoinKind::Full, JoinStrictness::All>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Left, JoinStrictness::All, prefer_use_maps_all>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Inner, JoinStrictness::All, prefer_use_maps_all> { using Map = HashJoin::MapsAll; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Right, JoinStrictness::All, prefer_use_maps_all> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Full, JoinStrictness::All, prefer_use_maps_all>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };

 /// Only SEMI LEFT and SEMI RIGHT are valid. INNER and FULL are here for templates instantiation.
-template <> struct MapGetter<JoinKind::Left, JoinStrictness::Semi>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Inner, JoinStrictness::Semi> { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Right, JoinStrictness::Semi> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
-template <> struct MapGetter<JoinKind::Full, JoinStrictness::Semi>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <> struct MapGetter<JoinKind::Left, JoinStrictness::Semi, false>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <> struct MapGetter<JoinKind::Left, JoinStrictness::Semi, true>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Inner, JoinStrictness::Semi, prefer_use_maps_all> { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Right, JoinStrictness::Semi, prefer_use_maps_all> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Full, JoinStrictness::Semi, prefer_use_maps_all>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };

-/// Only SEMI LEFT and SEMI RIGHT are valid. INNER and FULL are here for templates instantiation.
-template <> struct MapGetter<JoinKind::Left, JoinStrictness::Anti>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Inner, JoinStrictness::Anti> { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
-template <> struct MapGetter<JoinKind::Right, JoinStrictness::Anti> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
-template <> struct MapGetter<JoinKind::Full, JoinStrictness::Anti>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+/// Only ANTI LEFT and ANTI RIGHT are valid. INNER and FULL are here for templates instantiation.
+template <> struct MapGetter<JoinKind::Left, JoinStrictness::Anti, false>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <> struct MapGetter<JoinKind::Left, JoinStrictness::Anti, true>  { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Inner, JoinStrictness::Anti, prefer_use_maps_all> { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Right, JoinStrictness::Anti, prefer_use_maps_all> { using Map = HashJoin::MapsAll; static constexpr bool flagged = true; };
+template <bool prefer_use_maps_all> struct MapGetter<JoinKind::Full, JoinStrictness::Anti, prefer_use_maps_all>  { using Map = HashJoin::MapsOne; static constexpr bool flagged = false; };

-template <JoinKind kind>
-struct MapGetter<kind, JoinStrictness::Asof> { using Map = HashJoin::MapsAsof; static constexpr bool flagged = false; };
+template <JoinKind kind, bool prefer_use_maps_all>
+struct MapGetter<kind, JoinStrictness::Asof, prefer_use_maps_all> { using Map = HashJoin::MapsAsof; static constexpr bool flagged = false; };

 static constexpr std::array<JoinStrictness, 6> STRICTNESSES = {
    JoinStrictness::RightAny,
@ -62,7 +77,7 @@ static constexpr std::array<JoinKind, 4> KINDS = {
 };

 /// Init specified join map
-inline bool joinDispatchInit(JoinKind kind, JoinStrictness strictness, HashJoin::MapsVariant & maps)
+inline bool joinDispatchInit(JoinKind kind, JoinStrictness strictness, HashJoin::MapsVariant & maps, bool prefer_use_maps_all = false)
 {
    return static_for<0, KINDS.size() * STRICTNESSES.size()>([&](auto ij)
    {
@ -70,7 +85,10 @@ inline bool joinDispatchInit(JoinKind kind, JoinStrictness strictness, HashJoin:
        constexpr auto j = ij % STRICTNESSES.size();
        if (kind == KINDS[i] && strictness == STRICTNESSES[j])
        {
-            maps = typename MapGetter<KINDS[i], STRICTNESSES[j]>::Map();
+            if (prefer_use_maps_all)
+                maps = typename MapGetter<KINDS[i], STRICTNESSES[j], true>::Map();
+            else
+                maps = typename MapGetter<KINDS[i], STRICTNESSES[j], false>::Map();
            return true;
        }
        return false;
@ -79,7 +97,7 @@ inline bool joinDispatchInit(JoinKind kind, JoinStrictness strictness, HashJoin:

 /// Call function on specified join map
 template <typename MapsVariant, typename Func>
-inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, MapsVariant & maps, Func && func)
+inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, MapsVariant & maps, bool prefer_use_maps_all, Func && func)
 {
    return static_for<0, KINDS.size() * STRICTNESSES.size()>([&](auto ij)
    {
@ -89,10 +107,16 @@ inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, MapsVariant &
        constexpr auto j = ij % STRICTNESSES.size();
        if (kind == KINDS[i] && strictness == STRICTNESSES[j])
        {
-            func(
-                std::integral_constant<JoinKind, KINDS[i]>(),
-                std::integral_constant<JoinStrictness, STRICTNESSES[j]>(),
-                std::get<typename MapGetter<KINDS[i], STRICTNESSES[j]>::Map>(maps));
+            if (prefer_use_maps_all)
+                func(
+                    std::integral_constant<JoinKind, KINDS[i]>(),
+                    std::integral_constant<JoinStrictness, STRICTNESSES[j]>(),
+                    std::get<typename MapGetter<KINDS[i], STRICTNESSES[j], true>::Map>(maps));
+            else
+                func(
+                    std::integral_constant<JoinKind, KINDS[i]>(),
+                    std::integral_constant<JoinStrictness, STRICTNESSES[j]>(),
+                    std::get<typename MapGetter<KINDS[i], STRICTNESSES[j], false>::Map>(maps));
            return true;
        }
        return false;
@ -101,7 +125,7 @@ inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, MapsVariant &

 /// Call function on specified join map
 template <typename MapsVariant, typename Func>
-inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, std::vector<const MapsVariant *> & mapsv, Func && func)
+inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, std::vector<const MapsVariant *> & mapsv, bool prefer_use_maps_all, Func && func)
 {
    return static_for<0, KINDS.size() * STRICTNESSES.size()>([&](auto ij)
    {
@ -111,17 +135,31 @@ inline bool joinDispatch(JoinKind kind, JoinStrictness strictness, std::vector<c
        constexpr auto j = ij % STRICTNESSES.size();
        if (kind == KINDS[i] && strictness == STRICTNESSES[j])
        {
-            using MapType = typename MapGetter<KINDS[i], STRICTNESSES[j]>::Map;
-            std::vector<const MapType *> v;
-            v.reserve(mapsv.size());
-            for (const auto & el : mapsv)
-                v.push_back(&std::get<MapType>(*el));
+            if (prefer_use_maps_all)
+            {
+                using MapType = typename MapGetter<KINDS[i], STRICTNESSES[j], true>::Map;
+                std::vector<const MapType *> v;
+                v.reserve(mapsv.size());
+                for (const auto & el : mapsv)
+                    v.push_back(&std::get<MapType>(*el));

-            func(
-                std::integral_constant<JoinKind, KINDS[i]>(),
-                std::integral_constant<JoinStrictness, STRICTNESSES[j]>(),
-                v
-                /*std::get<typename MapGetter<KINDS[i], STRICTNESSES[j]>::Map>(maps)*/);
+                func(
+                    std::integral_constant<JoinKind, KINDS[i]>(), std::integral_constant<JoinStrictness, STRICTNESSES[j]>(), v
+                    /*std::get<typename MapGetter<KINDS[i], STRICTNESSES[j]>::Map>(maps)*/);
+            }
+            else
+            {
+                using MapType = typename MapGetter<KINDS[i], STRICTNESSES[j], false>::Map;
+                std::vector<const MapType *> v;
+                v.reserve(mapsv.size());
+                for (const auto & el : mapsv)
+                    v.push_back(&std::get<MapType>(*el));
+
+                func(
+                    std::integral_constant<JoinKind, KINDS[i]>(), std::integral_constant<JoinStrictness, STRICTNESSES[j]>(), v
+                    /*std::get<typename MapGetter<KINDS[i], STRICTNESSES[j]>::Map>(maps)*/);
+
+            }
            return true;
        }
        return false;
--- a/src/Parsers/FieldFromAST.cpp
+++ b/src/Parsers/FieldFromAST.cpp
@ -1,5 +1,4 @@
 #include <Parsers/FieldFromAST.h>
-#include <Disks/getOrCreateDiskFromAST.h>
 #include <Parsers/formatAST.h>
 #include <Parsers/ASTIdentifier.h>
 #include <Parsers/ASTLiteral.h>
--- a/src/Processors/Formats/IOutputFormat.cpp
+++ b/src/Processors/Formats/IOutputFormat.cpp
@ -69,9 +69,10 @@ void IOutputFormat::work()

    if (finished && !finalized)
    {
-        if (rows_before_limit_counter && rows_before_limit_counter->hasAppliedLimit())
+        if (rows_before_limit_counter && rows_before_limit_counter->hasAppliedStep())
            setRowsBeforeLimit(rows_before_limit_counter->get());
-
+        if (rows_before_aggregation_counter && rows_before_aggregation_counter->hasAppliedStep())
+            setRowsBeforeAggregation(rows_before_aggregation_counter->get());
        finalize();
        if (auto_flush)
            flush();
--- a/src/Processors/Formats/IOutputFormat.h
+++ b/src/Processors/Formats/IOutputFormat.h
@ -1,9 +1,9 @@
 #pragma once

 #include <string>
-#include <Processors/IProcessor.h>
-#include <Processors/RowsBeforeLimitCounter.h>
 #include <IO/Progress.h>
+#include <Processors/IProcessor.h>
+#include <Processors/RowsBeforeStepCounter.h>
 #include <Common/Stopwatch.h>

 namespace DB
@ -36,14 +36,20 @@ public:
    void setAutoFlush() { auto_flush = true; }

    /// Value for rows_before_limit_at_least field.
-    virtual void setRowsBeforeLimit(size_t /*rows_before_limit*/) {}
+    virtual void setRowsBeforeLimit(size_t /*rows_before_limit*/) { }

    /// Counter to calculate rows_before_limit_at_least in processors pipeline.
-    void setRowsBeforeLimitCounter(RowsBeforeLimitCounterPtr counter) override { rows_before_limit_counter.swap(counter); }
+    void setRowsBeforeLimitCounter(RowsBeforeStepCounterPtr counter) override { rows_before_limit_counter.swap(counter); }
+
+    /// Value for rows_before_aggregation field.
+    virtual void setRowsBeforeAggregation(size_t /*rows_before_aggregation*/) { }
+
+    /// Counter to calculate rows_before_aggregation in processors pipeline.
+    void setRowsBeforeAggregationCounter(RowsBeforeStepCounterPtr counter) override { rows_before_aggregation_counter.swap(counter); }

    /// Notify about progress. Method could be called from different threads.
    /// Passed value are delta, that must be summarized.
-    virtual void onProgress(const Progress & /*progress*/) {}
+    virtual void onProgress(const Progress & /*progress*/) { }

    /// Content-Type to set when sending HTTP response.
    virtual std::string getContentType() const { return "text/plain; charset=UTF-8"; }
@ -151,6 +157,8 @@ protected:
        Progress progress;
        bool applied_limit = false;
        size_t rows_before_limit = 0;
+        bool applied_aggregation = false;
+        size_t rows_before_aggregation = 0;
        Chunk totals;
        Chunk extremes;
    };
@ -184,7 +192,8 @@ protected:
    bool need_write_prefix  = true;
    bool need_write_suffix = true;

-    RowsBeforeLimitCounterPtr rows_before_limit_counter;
+    RowsBeforeStepCounterPtr rows_before_limit_counter;
+    RowsBeforeStepCounterPtr rows_before_aggregation_counter;
    Statistics statistics;

 private:
--- a/src/Processors/Formats/Impl/JSONColumnsWithMetadataBlockOutputFormat.cpp
+++ b/src/Processors/Formats/Impl/JSONColumnsWithMetadataBlockOutputFormat.cpp
@ -81,6 +81,8 @@ void JSONColumnsWithMetadataBlockOutputFormat::finalizeImpl()
        rows,
        statistics.rows_before_limit,
        statistics.applied_limit,
+        statistics.rows_before_aggregation,
+        statistics.applied_aggregation,
        statistics.watch,
        statistics.progress,
        format_settings.write_statistics,
--- a/src/Processors/Formats/Impl/JSONColumnsWithMetadataBlockOutputFormat.h
+++ b/src/Processors/Formats/Impl/JSONColumnsWithMetadataBlockOutputFormat.h
@ -44,6 +44,11 @@ public:
    String getName() const override { return "JSONCompactColumnsBlockOutputFormat"; }

    void setRowsBeforeLimit(size_t rows_before_limit_) override { statistics.rows_before_limit = rows_before_limit_; statistics.applied_limit = true; }
+    void setRowsBeforeAggregation(size_t rows_before_aggregation_) override
+    {
+        statistics.rows_before_aggregation = rows_before_aggregation_;
+        statistics.applied_aggregation = true;
+    }
    void onProgress(const Progress & progress_) override { statistics.progress.incrementPiecewiseAtomically(progress_); }

 protected:
--- a/src/Processors/Formats/Impl/JSONRowOutputFormat.cpp
+++ b/src/Processors/Formats/Impl/JSONRowOutputFormat.cpp
@ -116,6 +116,8 @@ void JSONRowOutputFormat::finalizeImpl()
        row_count,
        statistics.rows_before_limit,
        statistics.applied_limit,
+        statistics.rows_before_aggregation,
+        statistics.applied_aggregation,
        statistics.watch,
        statistics.progress,
        settings.write_statistics && exception_message.empty(),
--- a/src/Processors/Formats/Impl/JSONRowOutputFormat.h
+++ b/src/Processors/Formats/Impl/JSONRowOutputFormat.h
@ -35,6 +35,11 @@ public:
        statistics.applied_limit = true;
        statistics.rows_before_limit = rows_before_limit_;
    }
+    void setRowsBeforeAggregation(size_t rows_before_aggregation_) override
+    {
+        statistics.applied_aggregation = true;
+        statistics.rows_before_aggregation = rows_before_aggregation_;
+    }

 protected:
    void writeField(const IColumn & column, const ISerialization & serialization, size_t row_num) override;
--- a/src/Processors/Formats/Impl/ParallelFormattingOutputFormat.h
+++ b/src/Processors/Formats/Impl/ParallelFormattingOutputFormat.h
@ -313,6 +313,12 @@ private:
        statistics.rows_before_limit = rows_before_limit;
        statistics.applied_limit = true;
    }
+    void setRowsBeforeAggregation(size_t rows_before_aggregation) override
+    {
+        std::lock_guard lock(statistics_mutex);
+        statistics.rows_before_aggregation = rows_before_aggregation;
+        statistics.applied_aggregation = true;
+    }
 };

 }
--- a/src/Processors/Formats/Impl/TemplateBlockOutputFormat.cpp
+++ b/src/Processors/Formats/Impl/TemplateBlockOutputFormat.cpp
@ -42,9 +42,11 @@ TemplateBlockOutputFormat::TemplateBlockOutputFormat(const Block & header_, Writ
            case static_cast<size_t>(ResultsetPart::TimeElapsed):
            case static_cast<size_t>(ResultsetPart::RowsRead):
            case static_cast<size_t>(ResultsetPart::BytesRead):
+            case static_cast<size_t>(ResultsetPart::RowsBeforeAggregation):
                if (format.escaping_rules[i] == EscapingRule::None)
-                    format.throwInvalidFormat("Serialization type for output part rows, rows_before_limit, time, "
-                                              "rows_read or bytes_read is not specified", i);
+                    format.throwInvalidFormat(
+                        "Serialization type for output part rows, rows, time, "
+                        "rows_read or bytes_read is not specified", i);
                break;
            default:
                format.throwInvalidFormat("Invalid output part", i);
@ -88,6 +90,8 @@ TemplateBlockOutputFormat::ResultsetPart TemplateBlockOutputFormat::stringToResu
        return ResultsetPart::RowsRead;
    else if (part == "bytes_read")
        return ResultsetPart::BytesRead;
+    else if (part == "rows_before_aggregation")
+        return ResultsetPart::RowsBeforeAggregation;
    else
        throw Exception(ErrorCodes::SYNTAX_ERROR, "Unknown output part {}", part);
 }
@ -173,6 +177,11 @@ void TemplateBlockOutputFormat::finalizeImpl()
            case ResultsetPart::BytesRead:
                writeValue<size_t, DataTypeUInt64>(statistics.progress.read_bytes.load(), format.escaping_rules[i]);
                break;
+            case ResultsetPart::RowsBeforeAggregation:
+                if (!statistics.applied_aggregation)
+                    format.throwInvalidFormat("Cannot print rows_before_aggregation for this request", i);
+                writeValue<size_t, DataTypeUInt64>(statistics.rows_before_aggregation, format.escaping_rules[i]);
+                break;
            default:
                break;
        }
--- a/src/Processors/Formats/Impl/TemplateBlockOutputFormat.h
+++ b/src/Processors/Formats/Impl/TemplateBlockOutputFormat.h
@ -21,6 +21,11 @@ public:
    String getName() const override { return "TemplateBlockOutputFormat"; }

    void setRowsBeforeLimit(size_t rows_before_limit_) override { statistics.rows_before_limit = rows_before_limit_; statistics.applied_limit = true; }
+    void setRowsBeforeAggregation(size_t rows_before_aggregation_) override
+    {
+        statistics.rows_before_aggregation = rows_before_aggregation_;
+        statistics.applied_aggregation = true;
+    }
    void onProgress(const Progress & progress_) override { statistics.progress.incrementPiecewiseAtomically(progress_); }

    enum class ResultsetPart : size_t
@ -33,7 +38,8 @@ public:
        RowsBeforeLimit,
        TimeElapsed,
        RowsRead,
-        BytesRead
+        BytesRead,
+        RowsBeforeAggregation
    };

    static ResultsetPart stringToResultsetPart(const String & part);
--- a/src/Processors/Formats/Impl/XMLRowOutputFormat.cpp
+++ b/src/Processors/Formats/Impl/XMLRowOutputFormat.cpp
@ -191,6 +191,7 @@ void XMLRowOutputFormat::finalizeImpl()


    writeRowsBeforeLimitAtLeast();
+    writeRowsBeforeAggregationAtLeast();

    if (!exception_message.empty())
        writeException();
@ -219,6 +220,16 @@ void XMLRowOutputFormat::writeRowsBeforeLimitAtLeast()
    }
 }

+void XMLRowOutputFormat::writeRowsBeforeAggregationAtLeast()
+{
+    if (statistics.applied_aggregation)
+    {
+        writeCString("\t<rows_before_aggregation>", *ostr);
+        writeIntText(statistics.rows_before_aggregation, *ostr);
+        writeCString("</rows_before_aggregation>\n", *ostr);
+    }
+}
+
 void XMLRowOutputFormat::writeStatistics()
 {
    writeCString("\t<statistics>\n", *ostr);
--- a/src/Processors/Formats/Impl/XMLRowOutputFormat.h
+++ b/src/Processors/Formats/Impl/XMLRowOutputFormat.h
@ -48,6 +48,11 @@ private:
        statistics.rows_before_limit = rows_before_limit_;
    }

+    void setRowsBeforeAggregation(size_t rows_before_aggregation_) override
+    {
+        statistics.applied_aggregation = true;
+        statistics.rows_before_aggregation = rows_before_aggregation_;
+    }
    void onRowsReadBeforeUpdate() override { row_count = getRowsReadBefore(); }

    void onProgress(const Progress & value) override;
@ -56,6 +61,7 @@ private:

    void writeExtremesElement(const char * title, const Columns & columns, size_t row_num);
    void writeRowsBeforeLimitAtLeast();
+    void writeRowsBeforeAggregationAtLeast();
    void writeStatistics();
    void writeException();

--- a/src/Processors/Formats/LazyOutputFormat.cpp
+++ b/src/Processors/Formats/LazyOutputFormat.cpp
@ -45,4 +45,8 @@ void LazyOutputFormat::setRowsBeforeLimit(size_t rows_before_limit)
    info.setRowsBeforeLimit(rows_before_limit);
 }

+void LazyOutputFormat::setRowsBeforeAggregation(size_t rows_before_aggregation)
+{
+    info.setRowsBeforeAggregation(rows_before_aggregation);
+}
 }
--- a/src/Processors/Formats/LazyOutputFormat.h
+++ b/src/Processors/Formats/LazyOutputFormat.h
@ -28,6 +28,7 @@ public:
    ProfileInfo & getProfileInfo() { return info; }

    void setRowsBeforeLimit(size_t rows_before_limit) override;
+    void setRowsBeforeAggregation(size_t rows_before_aggregation) override;

    void onCancel() noexcept override
    {
--- a/src/Processors/Formats/PullingOutputFormat.cpp
+++ b/src/Processors/Formats/PullingOutputFormat.cpp
@ -42,5 +42,8 @@ void PullingOutputFormat::setRowsBeforeLimit(size_t rows_before_limit)
 {
    info.setRowsBeforeLimit(rows_before_limit);
 }
-
+void PullingOutputFormat::setRowsBeforeAggregation(size_t rows_before_aggregation)
+{
+    info.setRowsBeforeAggregation(rows_before_aggregation);
+}
 }
--- a/src/Processors/Formats/PullingOutputFormat.h
+++ b/src/Processors/Formats/PullingOutputFormat.h
@ -22,6 +22,7 @@ public:
    ProfileInfo & getProfileInfo() { return info; }

    void setRowsBeforeLimit(size_t rows_before_limit) override;
+    void setRowsBeforeAggregation(size_t rows_before_aggregation) override;

    bool expectMaterializedColumns() const override { return false; }

--- a/src/Processors/IProcessor.h
+++ b/src/Processors/IProcessor.h
@ -21,8 +21,8 @@ class IQueryPlanStep;
 struct StorageLimits;
 using StorageLimitsList = std::list<StorageLimits>;

-class RowsBeforeLimitCounter;
-using RowsBeforeLimitCounterPtr = std::shared_ptr<RowsBeforeLimitCounter>;
+class RowsBeforeStepCounter;
+using RowsBeforeStepCounterPtr = std::shared_ptr<RowsBeforeStepCounter>;

 class IProcessor;
 using ProcessorPtr = std::shared_ptr<IProcessor>;
@ -377,7 +377,11 @@ public:

    /// Set rows_before_limit counter for current processor.
    /// This counter is used to calculate the number of rows right before any filtration of LimitTransform.
-    virtual void setRowsBeforeLimitCounter(RowsBeforeLimitCounterPtr /* counter */) {}
+    virtual void setRowsBeforeLimitCounter(RowsBeforeStepCounterPtr /* counter */) { }
+
+    /// Set rows_before_aggregation counter for current processor.
+    /// This counter is used to calculate the number of rows right before AggregatingTransform.
+    virtual void setRowsBeforeAggregationCounter(RowsBeforeStepCounterPtr /* counter */) { }

 protected:
    virtual void onCancel() noexcept {}
--- a/src/Processors/LimitTransform.h
+++ b/src/Processors/LimitTransform.h
@ -1,8 +1,8 @@
 #pragma once

-#include <Processors/IProcessor.h>
-#include <Processors/RowsBeforeLimitCounter.h>
 #include <Core/SortDescription.h>
+#include <Processors/IProcessor.h>
+#include <Processors/RowsBeforeStepCounter.h>

 namespace DB
 {
@ -30,7 +30,7 @@ private:
    std::vector<size_t> sort_column_positions;

    UInt64 rows_read = 0; /// including the last read block
-    RowsBeforeLimitCounterPtr rows_before_limit_at_least;
+    RowsBeforeStepCounterPtr rows_before_limit_at_least;

    /// State of port's pair.
    /// Chunks from different port pairs are not mixed for better cache locality.
@ -71,7 +71,7 @@ public:
    InputPort & getInputPort() { return inputs.front(); }
    OutputPort & getOutputPort() { return outputs.front(); }

-    void setRowsBeforeLimitCounter(RowsBeforeLimitCounterPtr counter) override { rows_before_limit_at_least.swap(counter); }
+    void setRowsBeforeLimitCounter(RowsBeforeStepCounterPtr counter) override { rows_before_limit_at_least.swap(counter); }
    void setInputPortHasCounter(size_t pos) { ports_data[pos].input_port_has_counter = true; }
 };

--- a/src/Processors/OffsetTransform.h
+++ b/src/Processors/OffsetTransform.h
@ -1,8 +1,8 @@
 #pragma once

-#include <Processors/IProcessor.h>
-#include <Processors/RowsBeforeLimitCounter.h>
 #include <Core/SortDescription.h>
+#include <Processors/IProcessor.h>
+#include <Processors/RowsBeforeStepCounter.h>

 namespace DB
 {
@ -16,7 +16,7 @@ private:
    UInt64 offset;
    UInt64 rows_read = 0; /// including the last read block

-    RowsBeforeLimitCounterPtr rows_before_limit_at_least;
+    RowsBeforeStepCounterPtr rows_before_limit_at_least;

    /// State of port's pair.
    /// Chunks from different port pairs are not mixed for better cache locality.
@ -45,7 +45,7 @@ public:
    InputPort & getInputPort() { return inputs.front(); }
    OutputPort & getOutputPort() { return outputs.front(); }

-    void setRowsBeforeLimitCounter(RowsBeforeLimitCounterPtr counter) override { rows_before_limit_at_least.swap(counter); }
+    void setRowsBeforeLimitCounter(RowsBeforeStepCounterPtr counter) override { rows_before_limit_at_least.swap(counter); }
 };

 }
--- a/src/Processors/QueryPlan/Optimizations/optimizeUseAggregateProjection.cpp
+++ b/src/Processors/QueryPlan/Optimizations/optimizeUseAggregateProjection.cpp
@ -255,20 +255,13 @@ static void appendAggregateFunctions(

        const auto * node = input;

-        if (node->result_name != aggregate.column_name)
-        {
-            if (DataTypeAggregateFunction::strictEquals(type, node->result_type))
-            {
-                node = &proj_dag.addAlias(*node, aggregate.column_name);
-            }
-            else
-            {
-                /// Cast to aggregate types specified in query if it's not
-                /// strictly the same as the one specified in projection. This
-                /// is required to generate correct results during finalization.
-                node = &proj_dag.addCast(*node, type, aggregate.column_name);
-            }
-        }
+        if (!DataTypeAggregateFunction::strictEquals(type, node->result_type))
+            /// Cast to aggregate types specified in query if it's not
+            /// strictly the same as the one specified in projection. This
+            /// is required to generate correct results during finalization.
+            node = &proj_dag.addCast(*node, type, aggregate.column_name);
+        else if (node->result_name != aggregate.column_name)
+            node = &proj_dag.addAlias(*node, aggregate.column_name);

        proj_dag_outputs.push_back(node);
    }
--- a/src/Processors/QueryPlan/Optimizations/removeRedundantSorting.cpp
+++ b/src/Processors/QueryPlan/Optimizations/removeRedundantSorting.cpp
@ -8,6 +8,7 @@
 #include <Processors/QueryPlan/JoinStep.h>
 #include <Processors/QueryPlan/LimitByStep.h>
 #include <Processors/QueryPlan/LimitStep.h>
+#include <Processors/QueryPlan/OffsetStep.h>
 #include <Processors/QueryPlan/Optimizations/Optimizations.h>
 #include <Processors/QueryPlan/QueryPlanVisitor.h>
 #include <Processors/QueryPlan/ReadFromMergeTree.h>
@ -59,9 +60,10 @@ public:

        if (typeid_cast<LimitStep *>(current_step)
            || typeid_cast<LimitByStep *>(current_step) /// (1) if there are LIMITs on top of ORDER BY, the ORDER BY is non-removable
-            || typeid_cast<FillingStep *>(current_step) /// (2) if ORDER BY is with FILL WITH, it is non-removable
-            || typeid_cast<SortingStep *>(current_step) /// (3) ORDER BY will change order of previous sorting
-            || typeid_cast<AggregatingStep *>(current_step)) /// (4) aggregation change order
+            || typeid_cast<OffsetStep *>(current_step) /// (2) OFFSET on top of ORDER BY, the ORDER BY is non-removable
+            || typeid_cast<FillingStep *>(current_step) /// (3) if ORDER BY is with FILL WITH, it is non-removable
+            || typeid_cast<SortingStep *>(current_step) /// (4) ORDER BY will change order of previous sorting
+            || typeid_cast<AggregatingStep *>(current_step)) /// (5) aggregation change order
        {
            logStep("nodes_affect_order/push", current_node);
            nodes_affect_order.push_back(current_node);
--- a/src/Processors/RowsBeforeLimitCounter.h
+++ b/src/Processors/RowsBeforeLimitCounter.h
@ -1,36 +0,0 @@
-#pragma once
-#include <atomic>
-#include <memory>
-
-namespace DB
-{
-
-/// This class helps to calculate rows_before_limit_at_least.
-class RowsBeforeLimitCounter
-{
-public:
-    void add(uint64_t rows)
-    {
-        setAppliedLimit();
-        rows_before_limit.fetch_add(rows, std::memory_order_release);
-    }
-
-    void set(uint64_t rows)
-    {
-        setAppliedLimit();
-        rows_before_limit.store(rows, std::memory_order_release);
-    }
-
-    uint64_t get() const { return rows_before_limit.load(std::memory_order_acquire); }
-
-    void setAppliedLimit() { has_applied_limit.store(true, std::memory_order_release); }
-    bool hasAppliedLimit() const { return has_applied_limit.load(std::memory_order_acquire); }
-
-private:
-    std::atomic<uint64_t> rows_before_limit = 0;
-    std::atomic_bool has_applied_limit = false;
-};
-
-using RowsBeforeLimitCounterPtr = std::shared_ptr<RowsBeforeLimitCounter>;
-
-}
--- a/src/Processors/RowsBeforeStepCounter.h
+++ b/src/Processors/RowsBeforeStepCounter.h
@ -0,0 +1,36 @@
+#pragma once
+#include <atomic>
+#include <memory>
+
+namespace DB
+{
+
+/// This class helps to calculate rows_before_limit_at_least and rows_before_aggregation.
+class RowsBeforeStepCounter
+{
+public:
+    void add(uint64_t rows)
+    {
+        setAppliedStep();
+        rows_before_step.fetch_add(rows, std::memory_order_release);
+    }
+
+    void set(uint64_t rows)
+    {
+        setAppliedStep();
+        rows_before_step.store(rows, std::memory_order_release);
+    }
+
+    uint64_t get() const { return rows_before_step.load(std::memory_order_acquire); }
+
+    void setAppliedStep() { has_applied_step.store(true, std::memory_order_release); }
+    bool hasAppliedStep() const { return has_applied_step.load(std::memory_order_acquire); }
+
+private:
+    std::atomic<uint64_t> rows_before_step = 0;
+    std::atomic_bool has_applied_step = false;
+};
+
+using RowsBeforeStepCounterPtr = std::shared_ptr<RowsBeforeStepCounter>;
+
+}
--- a/src/Processors/Sources/DelayedSource.cpp
+++ b/src/Processors/Sources/DelayedSource.cpp
@ -139,6 +139,12 @@ void DelayedSource::work()
            processor->setRowsBeforeLimitCounter(rows_before_limit);
    }

+    if (rows_before_aggregation)
+    {
+        for (auto & processor : processors)
+            processor->setRowsBeforeAggregationCounter(rows_before_aggregation);
+    }
+
    synchronizePorts(totals_output, totals, header, processors);
    synchronizePorts(extremes_output, extremes, header, processors);
 }
--- a/src/Processors/Sources/DelayedSource.h
+++ b/src/Processors/Sources/DelayedSource.h
@ -30,13 +30,15 @@ public:
    OutputPort * getTotalsPort() { return totals; }
    OutputPort * getExtremesPort() { return extremes; }

-    void setRowsBeforeLimitCounter(RowsBeforeLimitCounterPtr counter) override { rows_before_limit.swap(counter); }
+    void setRowsBeforeLimitCounter(RowsBeforeStepCounterPtr counter) override { rows_before_limit.swap(counter); }
+    void setRowsBeforeAggregationCounter(RowsBeforeStepCounterPtr counter) override { rows_before_aggregation.swap(counter); }

 private:
    QueryPlanResourceHolder resources;
    Creator creator;
    Processors processors;
-    RowsBeforeLimitCounterPtr rows_before_limit;
+    RowsBeforeStepCounterPtr rows_before_limit;
+    RowsBeforeStepCounterPtr rows_before_aggregation;

    /// Outputs for DelayedSource.
    OutputPort * main = nullptr;
--- a/src/Processors/Sources/RemoteSource.cpp
+++ b/src/Processors/Sources/RemoteSource.cpp
@ -37,16 +37,23 @@ RemoteSource::RemoteSource(RemoteQueryExecutorPtr executor, bool add_aggregation
        progress(value.read_rows, value.read_bytes);
    });

-    query_executor->setProfileInfoCallback([this](const ProfileInfo & info)
-    {
-        if (rows_before_limit)
+    query_executor->setProfileInfoCallback(
+        [this](const ProfileInfo & info)
        {
-            if (info.hasAppliedLimit())
-                rows_before_limit->add(info.getRowsBeforeLimit());
-            else
-                manually_add_rows_before_limit_counter = true; /// Remote subquery doesn't contain a limit
-        }
-    });
+            if (rows_before_limit)
+            {
+                if (info.hasAppliedLimit())
+                    rows_before_limit->add(info.getRowsBeforeLimit());
+                else
+                    manually_add_rows_before_limit_counter = true; /// Remote subquery doesn't contain a limit
+            }
+
+            if (rows_before_aggregation)
+            {
+                if (info.hasAppliedAggregation())
+                    rows_before_aggregation->add(info.getRowsBeforeAggregation());
+            }
+        });
 }

 RemoteSource::~RemoteSource() = default;
@ -184,7 +191,6 @@ std::optional<Chunk> RemoteSource::tryGenerate()
    {
        if (manually_add_rows_before_limit_counter)
            rows_before_limit->add(rows);
-
        query_executor->finish();
        return {};
    }
--- a/src/Processors/Sources/RemoteSource.h
+++ b/src/Processors/Sources/RemoteSource.h
@ -1,10 +1,10 @@
 #pragma once

 #include <Processors/ISource.h>
-#include <Processors/RowsBeforeLimitCounter.h>
+#include <Processors/RowsBeforeStepCounter.h>
 #include <QueryPipeline/Pipe.h>
-#include <Core/UUID.h>

+#include <Core/UUID.h>
 namespace DB
 {

@ -25,7 +25,8 @@ public:
    void work() override;
    String getName() const override { return "Remote"; }

-    void setRowsBeforeLimitCounter(RowsBeforeLimitCounterPtr counter) override { rows_before_limit.swap(counter); }
+    void setRowsBeforeLimitCounter(RowsBeforeStepCounterPtr counter) override { rows_before_limit.swap(counter); }
+    void setRowsBeforeAggregationCounter(RowsBeforeStepCounterPtr counter) override { rows_before_aggregation.swap(counter); }

    /// Stop reading from stream if output port is finished.
    void onUpdatePorts() override;
@ -46,7 +47,8 @@ private:
    bool executor_finished = false;
    bool add_aggregation_info = false;
    RemoteQueryExecutorPtr query_executor;
-    RowsBeforeLimitCounterPtr rows_before_limit;
+    RowsBeforeStepCounterPtr rows_before_limit;
+    RowsBeforeStepCounterPtr rows_before_aggregation;

    const bool async_read;
    const bool async_query_sending;
--- a/src/Processors/Sources/ShellCommandSource.cpp
+++ b/src/Processors/Sources/ShellCommandSource.cpp
@ -8,13 +8,15 @@
 #include <IO/WriteHelpers.h>
 #include <IO/ReadHelpers.h>

-#include <QueryPipeline/Pipe.h>
-#include <Processors/ISimpleTransform.h>
-#include <Processors/Formats/IOutputFormat.h>
-#include <Processors/Executors/CompletedPipelineExecutor.h>
 #include <Interpreters/Context.h>
+#include <Processors/Executors/CompletedPipelineExecutor.h>
+#include <Processors/Formats/IOutputFormat.h>
+#include <Processors/ISimpleTransform.h>
+#include <QueryPipeline/Pipe.h>
+
 #include <boost/circular_buffer.hpp>

+#include <ranges>

 namespace DB
 {
@ -68,11 +70,17 @@ static void makeFdBlocking(int fd)

 static int pollWithTimeout(pollfd * pfds, size_t num, size_t timeout_milliseconds)
 {
+    auto logger = getLogger("TimeoutReadBufferFromFileDescriptor");
+    auto describe_fd = [](const auto & pollfd) { return fmt::format("(fd={}, flags={})", pollfd.fd, fcntl(pollfd.fd, F_GETFL)); };
+
    int res;

    while (true)
    {
        Stopwatch watch;
+
+        LOG_TEST(logger, "Polling descriptors: {}", fmt::join(std::span(pfds, pfds + num) | std::views::transform(describe_fd), ", "));
+
        res = poll(pfds, static_cast<nfds_t>(num), static_cast<int>(timeout_milliseconds));

        if (res < 0)
@ -82,7 +90,10 @@ static int pollWithTimeout(pollfd * pfds, size_t num, size_t timeout_millisecond

            const auto elapsed = watch.elapsedMilliseconds();
            if (timeout_milliseconds <= elapsed)
+            {
+                LOG_TEST(logger, "Timeout exceeded: elapsed={}, timeout={}", elapsed, timeout_milliseconds);
                break;
+            }
            timeout_milliseconds -= elapsed;
        }
        else
@ -91,6 +102,12 @@ static int pollWithTimeout(pollfd * pfds, size_t num, size_t timeout_millisecond
        }
    }

+    LOG_TEST(
+        logger,
+        "Poll for descriptors: {} returned {}",
+        fmt::join(std::span(pfds, pfds + num) | std::views::transform(describe_fd), ", "),
+        res);
+
    return res;
 }

@ -200,12 +217,6 @@ public:
        return true;
    }

-    void reset() const
-    {
-        makeFdBlocking(stdout_fd);
-        makeFdBlocking(stderr_fd);
-    }
-
    ~TimeoutReadBufferFromFileDescriptor() override
    {
        tryMakeFdBlocking(stdout_fd);
--- a/src/Processors/Transforms/AggregatingInOrderTransform.cpp
+++ b/src/Processors/Transforms/AggregatingInOrderTransform.cpp
@ -81,6 +81,8 @@ void AggregatingInOrderTransform::consume(Chunk chunk)
        is_consume_started = true;
    }

+    if (rows_before_aggregation)
+        rows_before_aggregation->add(rows);
    src_rows += rows;
    src_bytes += chunk.bytes();

--- a/src/Processors/Transforms/AggregatingInOrderTransform.h
+++ b/src/Processors/Transforms/AggregatingInOrderTransform.h
@ -45,6 +45,7 @@ public:
    void work() override;

    void consume(Chunk chunk);
+    void setRowsBeforeAggregationCounter(RowsBeforeStepCounterPtr counter) override { rows_before_aggregation.swap(counter); }

 private:
    void generate();
@ -86,6 +87,8 @@ private:
    Chunk current_chunk;
    Chunk to_push_chunk;

+    RowsBeforeStepCounterPtr rows_before_aggregation;
+
    LoggerPtr log = getLogger("AggregatingInOrderTransform");
 };

--- a/Show More
+++ b/Show More