Merge branch 'master' of github.com:ClickHouse/ClickHouse into disable-merge-filters-by-default

2024-11-26 01:22:04 +00:00 · 2024-07-06 21:30:02 +02:00 · 2024-07-06 21:30:02 +02:00 · d685c062a3
commit d685c062a3
parent 73ece96954 4eeb59ec3f
209 changed files with 2372 additions and 8937 deletions
--- a/cmake/target.cmake
+++ b/cmake/target.cmake
@ -84,5 +84,5 @@ if (CMAKE_CROSSCOMPILING)
        message (FATAL_ERROR "Trying to cross-compile to unsupported system: ${CMAKE_SYSTEM_NAME}!")
    endif ()

-    message (STATUS "Cross-compiling for target: ${CMAKE_CXX_COMPILE_TARGET}")
+    message (STATUS "Cross-compiling for target: ${CMAKE_CXX_COMPILER_TARGET}")
 endif ()
--- a/contrib/s2geometry
+++ b/contrib/s2geometry
@ -1 +1 @@
-Subproject commit 0146e2d1355828f8f633cb050948250ad7406c57
+Subproject commit 6522a40338d58752c2a4227a3fc2bc4107c73e43
--- a/contrib/s2geometry-cmake/CMakeLists.txt
+++ b/contrib/s2geometry-cmake/CMakeLists.txt
@ -1,7 +1,6 @@
 option(ENABLE_S2_GEOMETRY "Enable S2 Geometry" ${ENABLE_LIBRARIES})

-# ARCH_S390X broke upstream, it can be re-enabled once https://github.com/google/s2geometry/pull/372 is merged
-if (NOT ENABLE_S2_GEOMETRY OR ARCH_S390X)
+if (NOT ENABLE_S2_GEOMETRY)
    message(STATUS "Not using S2 Geometry")
    return()
 endif()
--- a/docker/keeper/Dockerfile
+++ b/docker/keeper/Dockerfile
@ -34,7 +34,7 @@ RUN arch=${TARGETARCH:-amd64} \
 # lts / testing / prestable / etc
 ARG REPO_CHANNEL="stable"
 ARG REPOSITORY="https://packages.clickhouse.com/tgz/${REPO_CHANNEL}"
-ARG VERSION="24.6.1.4423"
+ARG VERSION="24.6.2.17"
 ARG PACKAGES="clickhouse-keeper"
 ARG DIRECT_DOWNLOAD_URLS=""

--- a/docker/reqgenerator.py
+++ b/docker/reqgenerator.py
@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+# To run this script you must install docker and piddeptree python package
+#
+
+import subprocess
+import os
+import sys
+
+
+def build_docker_deps(image_name, imagedir):
+    cmd = f"""docker run  --entrypoint "/bin/bash" {image_name} -c "pip install pipdeptree 2>/dev/null 1>/dev/null && pipdeptree --freeze  --warn silence | sed 's/ \+//g' | sort | uniq" > {imagedir}/requirements.txt"""
+    subprocess.check_call(cmd, shell=True)
+
+
+def check_docker_file_install_with_pip(filepath):
+    image_name = None
+    with open(filepath, "r") as f:
+        for line in f:
+            if "docker build" in line:
+                arr = line.split(" ")
+                if len(arr) > 4:
+                    image_name = arr[4]
+            if "pip3 install" in line or "pip install" in line:
+                return image_name, True
+    return image_name, False
+
+
+def process_affected_images(images_dir):
+    for root, _dirs, files in os.walk(images_dir):
+        for f in files:
+            if f == "Dockerfile":
+                docker_file_path = os.path.join(root, f)
+                print("Checking image on path", docker_file_path)
+                image_name, has_pip = check_docker_file_install_with_pip(
+                    docker_file_path
+                )
+                if has_pip:
+                    print("Find pip in", image_name)
+                    try:
+                        build_docker_deps(image_name, root)
+                    except Exception as ex:
+                        print(ex)
+                else:
+                    print("Pip not found in", docker_file_path)
+
+
+process_affected_images(sys.argv[1])
--- a/docker/server/Dockerfile.alpine
+++ b/docker/server/Dockerfile.alpine
@ -32,7 +32,7 @@ RUN arch=${TARGETARCH:-amd64} \
 # lts / testing / prestable / etc
 ARG REPO_CHANNEL="stable"
 ARG REPOSITORY="https://packages.clickhouse.com/tgz/${REPO_CHANNEL}"
-ARG VERSION="24.6.1.4423"
+ARG VERSION="24.6.2.17"
 ARG PACKAGES="clickhouse-client clickhouse-server clickhouse-common-static"
 ARG DIRECT_DOWNLOAD_URLS=""

--- a/docker/server/Dockerfile.ubuntu
+++ b/docker/server/Dockerfile.ubuntu
@ -28,7 +28,7 @@ RUN sed -i "s|http://archive.ubuntu.com|${apt_archive}|g" /etc/apt/sources.list

 ARG REPO_CHANNEL="stable"
 ARG REPOSITORY="deb [signed-by=/usr/share/keyrings/clickhouse-keyring.gpg] https://packages.clickhouse.com/deb ${REPO_CHANNEL} main"
-ARG VERSION="24.6.1.4423"
+ARG VERSION="24.6.2.17"
 ARG PACKAGES="clickhouse-client clickhouse-server clickhouse-common-static"

 #docker-official-library:off
--- a/docker/test/fasttest/Dockerfile
+++ b/docker/test/fasttest/Dockerfile
@ -19,10 +19,7 @@ RUN apt-get update \
        odbcinst \
        psmisc \
        python3 \
-        python3-lxml \
        python3-pip \
-        python3-requests \
-        python3-termcolor \
        unixodbc \
        pv \
        jq \
@ -31,7 +28,8 @@ RUN apt-get update \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

-RUN pip3 install numpy==1.26.3 scipy==1.12.0 pandas==1.5.3 Jinja2==3.1.3
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r /requirements.txt

 # This symlink is required by gcc to find the lld linker
 RUN ln -s /usr/bin/lld-${LLVM_VERSION} /usr/bin/ld.lld
@ -39,6 +37,10 @@ RUN ln -s /usr/bin/lld-${LLVM_VERSION} /usr/bin/ld.lld
 # https://salsa.debian.org/pkg-llvm-team/llvm-toolchain/-/commit/992e52c0b156a5ba9c6a8a54f8c4857ddd3d371d
 RUN sed -i '/_IMPORT_CHECK_FILES_FOR_\(mlir-\|llvm-bolt\|merge-fdata\|MLIR\)/ {s|^|#|}' /usr/lib/llvm-${LLVM_VERSION}/lib/cmake/llvm/LLVMExports-*.cmake

+# LLVM changes paths for compiler-rt libraries. For some reason clang-18.1.8 cannot catch up libraries from default install path.
+# It's very dirty workaround, better to build compiler and LLVM ourself and use it. Details: https://github.com/llvm/llvm-project/issues/95792
+RUN test ! -d /usr/lib/llvm-18/lib/clang/18/lib/x86_64-pc-linux-gnu || ln -s /usr/lib/llvm-18/lib/clang/18/lib/x86_64-pc-linux-gnu /usr/lib/llvm-18/lib/clang/18/lib/x86_64-unknown-linux-gnu
+
 ARG CCACHE_VERSION=4.6.1
 RUN mkdir /tmp/ccache \
    && cd /tmp/ccache \
--- a/docker/test/fasttest/requirements.txt
+++ b/docker/test/fasttest/requirements.txt
@ -0,0 +1,41 @@
+Jinja2==3.1.3
+MarkupSafe==2.1.5
+PyJWT==2.3.0
+PyYAML==6.0.1
+Pygments==2.11.2
+SecretStorage==3.3.1
+blinker==1.4
+certifi==2020.6.20
+chardet==4.0.0
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+httplib2==0.20.2
+idna==3.3
+importlib-metadata==4.6.4
+jeepney==0.7.1
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+lxml==4.8.0
+more-itertools==8.10.0
+numpy==1.26.3
+oauthlib==3.2.0
+packaging==24.1
+pandas==1.5.3
+pip==24.1.1
+pipdeptree==2.23.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+python-dateutil==2.9.0.post0
+pytz==2024.1
+requests==2.32.3
+scipy==1.12.0
+setuptools==59.6.0
+six==1.16.0
+termcolor==1.1.0
+urllib3==1.26.5
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/fuzzer/Dockerfile
+++ b/docker/test/fuzzer/Dockerfile
@ -31,7 +31,8 @@ RUN apt-get update \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

-RUN pip3 install Jinja2
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r /requirements.txt

 COPY * /

--- a/docker/test/fuzzer/requirements.txt
+++ b/docker/test/fuzzer/requirements.txt
@ -0,0 +1,27 @@
+blinker==1.4
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+httplib2==0.20.2
+importlib-metadata==4.6.4
+jeepney==0.7.1
+Jinja2==3.1.4
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+MarkupSafe==2.1.5
+more-itertools==8.10.0
+oauthlib==3.2.0
+packaging==24.1
+pip==24.1.1
+pipdeptree==2.23.0
+PyJWT==2.3.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/integration/base/Dockerfile
+++ b/docker/test/integration/base/Dockerfile
@ -33,7 +33,8 @@ RUN apt-get update \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

-RUN pip3 install pycurl
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r requirements.txt && rm -rf /root/.cache/pip

 # Architecture of the image when BuildKit/buildx is used
 ARG TARGETARCH
--- a/docker/test/integration/base/requirements.txt
+++ b/docker/test/integration/base/requirements.txt
@ -0,0 +1,26 @@
+blinker==1.4
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+httplib2==0.20.2
+importlib-metadata==4.6.4
+jeepney==0.7.1
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+more-itertools==8.10.0
+oauthlib==3.2.0
+packaging==24.1
+pip==24.1.1
+pipdeptree==2.23.0
+pycurl==7.45.3
+PyJWT==2.3.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/integration/resolver/Dockerfile
+++ b/docker/test/integration/resolver/Dockerfile
@ -2,4 +2,5 @@
 # Helper docker container to run python bottle apps

 FROM python:3
-RUN python -m pip install bottle
+COPY requirements.txt /
+RUN python -m pip install --no-cache-dir -r requirements.txt
--- a/docker/test/integration/resolver/requirements.txt
+++ b/docker/test/integration/resolver/requirements.txt
@ -0,0 +1,6 @@
+bottle==0.12.25
+packaging==24.1
+pip==23.2.1
+pipdeptree==2.23.0
+setuptools==69.0.3
+wheel==0.42.0
--- a/docker/test/integration/runner/Dockerfile
+++ b/docker/test/integration/runner/Dockerfile
@ -26,7 +26,6 @@ RUN apt-get update \
    libicu-dev \
    bsdutils \
    curl \
-    python3-pika \
    liblua5.1-dev \
    luajit \
    libssl-dev \
@ -61,49 +60,8 @@ RUN curl -fsSL https://download.docker.com/linux/ubuntu/gpg | apt-key add - \

 # kazoo 2.10.0 is broken
 # https://s3.amazonaws.com/clickhouse-test-reports/59337/524625a1d2f4cc608a3f1059e3df2c30f353a649/integration_tests__asan__analyzer__[5_6].html
-RUN python3 -m pip install --no-cache-dir \
-    PyMySQL==1.1.0 \
-    asyncio==3.4.3 \
-    avro==1.10.2 \
-    azure-storage-blob==12.19.0 \
-    boto3==1.34.24 \
-    cassandra-driver==3.29.0 \
-    confluent-kafka==2.3.0 \
-    delta-spark==2.3.0 \
-    dict2xml==1.7.4 \
-    dicttoxml==1.7.16 \
-    docker==6.1.3 \
-    docker-compose==1.29.2 \
-    grpcio==1.60.0 \
-    grpcio-tools==1.60.0 \
-    kafka-python==2.0.2 \
-    lz4==4.3.3 \
-    minio==7.2.3 \
-    nats-py==2.6.0 \
-    protobuf==4.25.2 \
-    kazoo==2.9.0 \
-    psycopg2-binary==2.9.6 \
-    pyhdfs==0.3.1 \
-    pymongo==3.11.0 \
-    pyspark==3.3.2 \
-    pytest==7.4.4 \
-    pytest-order==1.0.0 \
-    pytest-random==0.2 \
-    pytest-repeat==0.9.3 \
-    pytest-timeout==2.2.0 \
-    pytest-xdist==3.5.0 \
-    pytest-reportlog==0.4.0 \
-    pytz==2023.3.post1 \
-    pyyaml==5.3.1 \
-    redis==5.0.1 \
-    requests-kerberos==0.14.0 \
-    tzlocal==2.1 \
-    retry==0.9.2 \
-    bs4==0.0.2 \
-    lxml==5.1.0 \
-    urllib3==2.0.7 \
-    jwcrypto==1.5.6
-# bs4, lxml are for cloud tests, do not delete
+COPY requirements.txt /
+RUN python3 -m pip install --no-cache-dir -r requirements.txt

 # Hudi supports only spark 3.3.*, not 3.4
 RUN curl -fsSL -O https://archive.apache.org/dist/spark/spark-3.3.2/spark-3.3.2-bin-hadoop3.tgz \
--- a/docker/test/integration/runner/requirements.txt
+++ b/docker/test/integration/runner/requirements.txt
@ -0,0 +1,113 @@
+PyHDFS==0.3.1
+PyJWT==2.3.0
+PyMySQL==1.1.0
+PyNaCl==1.5.0
+PyYAML==5.3.1
+SecretStorage==3.3.1
+argon2-cffi-bindings==21.2.0
+argon2-cffi==23.1.0
+async-timeout==4.0.3
+asyncio==3.4.3
+attrs==23.2.0
+avro==1.10.2
+azure-core==1.30.1
+azure-storage-blob==12.19.0
+bcrypt==4.1.3
+beautifulsoup4==4.12.3
+blinker==1.4
+boto3==1.34.24
+botocore==1.34.101
+bs4==0.0.2
+cassandra-driver==3.29.0
+certifi==2024.2.2
+cffi==1.16.0
+charset-normalizer==3.3.2
+click==8.1.7
+confluent-kafka==2.3.0
+cryptography==3.4.8
+dbus-python==1.2.18
+decorator==5.1.1
+delta-spark==2.3.0
+dict2xml==1.7.4
+dicttoxml==1.7.16
+distro-info==1.1+ubuntu0.2
+distro==1.7.0
+docker-compose==1.29.2
+docker==6.1.3
+dockerpty==0.4.1
+docopt==0.6.2
+exceptiongroup==1.2.1
+execnet==2.1.1
+geomet==0.2.1.post1
+grpcio-tools==1.60.0
+grpcio==1.60.0
+gssapi==1.8.3
+httplib2==0.20.2
+idna==3.7
+importlib-metadata==4.6.4
+iniconfig==2.0.0
+isodate==0.6.1
+jeepney==0.7.1
+jmespath==1.0.1
+jsonschema==3.2.0
+jwcrypto==1.5.6
+kafka-python==2.0.2
+kazoo==2.9.0
+keyring==23.5.0
+krb5==0.5.1
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+lxml==5.1.0
+lz4==4.3.3
+minio==7.2.3
+more-itertools==8.10.0
+nats-py==2.6.0
+oauthlib==3.2.0
+packaging==24.0
+paramiko==3.4.0
+pika==1.2.0
+pip==24.1.1
+pipdeptree==2.23.0
+pluggy==1.5.0
+protobuf==4.25.2
+psycopg2-binary==2.9.6
+py4j==0.10.9.5
+py==1.11.0
+pycparser==2.22
+pycryptodome==3.20.0
+pymongo==3.11.0
+pyparsing==2.4.7
+pyrsistent==0.20.0
+pyspark==3.3.2
+pyspnego==0.10.2
+pytest-order==1.0.0
+pytest-random==0.2
+pytest-repeat==0.9.3
+pytest-reportlog==0.4.0
+pytest-timeout==2.2.0
+pytest-xdist==3.5.0
+pytest==7.4.4
+python-apt==2.4.0+ubuntu3
+python-dateutil==2.9.0.post0
+python-dotenv==0.21.1
+pytz==2023.3.post1
+redis==5.0.1
+requests-kerberos==0.14.0
+requests==2.31.0
+retry==0.9.2
+s3transfer==0.10.1
+setuptools==59.6.0
+simplejson==3.19.2
+six==1.16.0
+soupsieve==2.5
+texttable==1.7.0
+tomli==2.0.1
+typing_extensions==4.11.0
+tzlocal==2.1
+unattended-upgrades==0.1
+urllib3==2.0.7
+wadllib==1.3.6
+websocket-client==0.59.0
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/libfuzzer/Dockerfile
+++ b/docker/test/libfuzzer/Dockerfile
@ -1,3 +1,4 @@
+# docker build -t clickhouse/libfuzzer .
 ARG FROM_TAG=latest
 FROM clickhouse/test-base:$FROM_TAG

@ -29,7 +30,8 @@ RUN apt-get update \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

-RUN pip3 install Jinja2
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r /requirements.txt

 COPY * /

--- a/docker/test/libfuzzer/requirements.txt
+++ b/docker/test/libfuzzer/requirements.txt
@ -0,0 +1,27 @@
+blinker==1.4
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+httplib2==0.20.2
+importlib-metadata==4.6.4
+jeepney==0.7.1
+Jinja2==3.1.4
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+MarkupSafe==2.1.5
+more-itertools==8.10.0
+oauthlib==3.2.0
+packaging==24.1
+pip==24.1.1
+pipdeptree==2.23.0
+PyJWT==2.3.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/performance-comparison/Dockerfile
+++ b/docker/test/performance-comparison/Dockerfile
@ -23,7 +23,6 @@ RUN apt-get update \
            python3 \
            python3-dev \
            python3-pip \
-            python3-setuptools \
            rsync \
            tree \
            tzdata \
@ -33,12 +32,14 @@ RUN apt-get update \
            cargo \
            ripgrep \
            zstd \
-    && pip3 --no-cache-dir install 'clickhouse-driver==0.2.1' scipy \
    && apt-get purge --yes python3-dev g++ \
    && apt-get autoremove --yes \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

+COPY requirements.txt /
+RUN pip3 --no-cache-dir install -r requirements.txt
+
 COPY run.sh /

 CMD ["bash", "/run.sh"]
--- a/docker/test/performance-comparison/requirements.txt
+++ b/docker/test/performance-comparison/requirements.txt
@ -0,0 +1,32 @@
+blinker==1.4
+clickhouse-driver==0.2.7
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+httplib2==0.20.2
+importlib-metadata==4.6.4
+jeepney==0.7.1
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+more-itertools==8.10.0
+numpy==1.26.3
+oauthlib==3.2.0
+packaging==24.1
+pip==24.1.1
+pipdeptree==2.23.0
+Pygments==2.11.2
+PyJWT==2.3.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+pytz==2023.4
+PyYAML==6.0.1
+scipy==1.12.0
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+tzlocal==2.1
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/sqllogic/Dockerfile
+++ b/docker/test/sqllogic/Dockerfile
@ -18,11 +18,8 @@ RUN apt-get update --yes \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

-RUN pip3 install \
-    numpy \
-    pyodbc \
-    deepdiff \
-    sqlglot
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r /requirements.txt

 ARG odbc_driver_url="https://github.com/ClickHouse/clickhouse-odbc/releases/download/v1.1.6.20200320/clickhouse-odbc-1.1.6-Linux.tar.gz"

--- a/docker/test/sqllogic/requirements.txt
+++ b/docker/test/sqllogic/requirements.txt
@ -0,0 +1,30 @@
+blinker==1.4
+cryptography==3.4.8
+dbus-python==1.2.18
+deepdiff==7.0.1
+distro==1.7.0
+httplib2==0.20.2
+importlib-metadata==4.6.4
+jeepney==0.7.1
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+more-itertools==8.10.0
+numpy==1.26.4
+oauthlib==3.2.0
+ordered-set==4.1.0
+packaging==24.1
+pip==24.1.1
+pipdeptree==2.23.0
+PyJWT==2.3.0
+pyodbc==5.1.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+sqlglot==23.16.0
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/sqltest/Dockerfile
+++ b/docker/test/sqltest/Dockerfile
@ -14,9 +14,8 @@ RUN apt-get update --yes \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

-RUN pip3 install \
-    pyyaml \
-    clickhouse-driver
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r /requirements.txt

 ARG sqltest_repo="https://github.com/elliotchance/sqltest/"

--- a/docker/test/sqltest/requirements.txt
+++ b/docker/test/sqltest/requirements.txt
@ -0,0 +1,29 @@
+blinker==1.4
+clickhouse-driver==0.2.7
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+httplib2==0.20.2
+importlib-metadata==4.6.4
+jeepney==0.7.1
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+more-itertools==8.10.0
+oauthlib==3.2.0
+packaging==24.1
+pip==24.1.1
+pipdeptree==2.23.0
+PyJWT==2.3.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+pytz==2024.1
+PyYAML==6.0.1
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+tzlocal==5.2
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/stateful/Dockerfile
+++ b/docker/test/stateful/Dockerfile
@ -6,7 +6,6 @@ FROM clickhouse/stateless-test:$FROM_TAG
 RUN apt-get update -y \
    && env DEBIAN_FRONTEND=noninteractive \
        apt-get install --yes --no-install-recommends \
-        python3-requests \
        nodejs \
        npm \
    && apt-get clean \
--- a/docker/test/stateless/Dockerfile
+++ b/docker/test/stateless/Dockerfile
@ -25,10 +25,7 @@ RUN apt-get update -y \
            openssl \
            postgresql-client \
            python3 \
-            python3-lxml \
            python3-pip \
-            python3-requests \
-            python3-termcolor \
            qemu-user-static \
            sqlite3 \
            sudo \
@ -51,7 +48,8 @@ RUN curl -OL https://github.com/protocolbuffers/protobuf/releases/download/v${PR
    && unzip protoc-${PROTOC_VERSION}-linux-x86_64.zip -d /usr/local \
    && rm protoc-${PROTOC_VERSION}-linux-x86_64.zip

-RUN pip3 install numpy==1.26.3 scipy==1.12.0 pandas==1.5.3 Jinja2==3.1.3 pyarrow==15.0.0
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r /requirements.txt

 RUN mkdir -p /tmp/clickhouse-odbc-tmp \
  && cd /tmp/clickhouse-odbc-tmp \
--- a/docker/test/stateless/requirements.txt
+++ b/docker/test/stateless/requirements.txt
@ -0,0 +1,51 @@
+awscli==1.22.34
+blinker==1.4
+botocore==1.23.34
+certifi==2020.6.20
+chardet==4.0.0
+colorama==0.4.4
+cryptography==3.4.8
+dbus-python==1.2.18
+distro==1.7.0
+docutils==0.17.1
+gyp==0.1
+httplib2==0.20.2
+idna==3.3
+importlib-metadata==4.6.4
+jeepney==0.7.1
+Jinja2==3.1.3
+jmespath==0.10.0
+keyring==23.5.0
+launchpadlib==1.10.16
+lazr.restfulclient==0.14.4
+lazr.uri==1.0.6
+lxml==4.8.0
+MarkupSafe==2.1.5
+more-itertools==8.10.0
+numpy==1.26.3
+oauthlib==3.2.0
+packaging==24.1
+pandas==1.5.3
+pip==24.1.1
+pipdeptree==2.23.0
+pyarrow==15.0.0
+pyasn1==0.4.8
+PyJWT==2.3.0
+pyparsing==2.4.7
+python-apt==2.4.0+ubuntu3
+python-dateutil==2.8.1
+pytz==2024.1
+PyYAML==6.0.1
+requests==2.32.3
+roman==3.3
+rsa==4.8
+s3transfer==0.5.0
+scipy==1.12.0
+SecretStorage==3.3.1
+setuptools==59.6.0
+six==1.16.0
+termcolor==1.1.0
+urllib3==1.26.5
+wadllib==1.3.6
+wheel==0.37.1
+zipp==1.0.0
--- a/docker/test/style/Dockerfile
+++ b/docker/test/style/Dockerfile
@ -23,22 +23,8 @@ RUN apt-get update && env DEBIAN_FRONTEND=noninteractive apt-get install --yes \
    && rm -rf /var/lib/apt/lists/* /var/cache/debconf /tmp/*

 # python-magic is the same version as in Ubuntu 22.04
-RUN pip3 install \
-        PyGithub \
-        black==23.12.0 \
-        boto3 \
-        codespell==2.2.1 \
-        mypy==1.8.0 \
-        pylint==3.1.0 \
-        python-magic==0.4.24 \
-        flake8==4.0.1 \
-        requests \
-        thefuzz \
-        tqdm==4.66.4 \
-        types-requests \
-        unidiff \
-        jwt \
-    && rm -rf /root/.cache/pip
+COPY requirements.txt /
+RUN pip3 install --no-cache-dir -r requirements.txt

 RUN echo "en_US.UTF-8 UTF-8" > /etc/locale.gen && locale-gen en_US.UTF-8
 ENV LC_ALL en_US.UTF-8
--- a/docker/test/style/requirements.txt
+++ b/docker/test/style/requirements.txt
@ -0,0 +1,58 @@
+aiohttp==3.9.5
+aiosignal==1.3.1
+astroid==3.1.0
+async-timeout==4.0.3
+attrs==23.2.0
+black==23.12.0
+boto3==1.34.131
+botocore==1.34.131
+certifi==2024.6.2
+cffi==1.16.0
+charset-normalizer==3.3.2
+click==8.1.7
+codespell==2.2.1
+cryptography==42.0.8
+Deprecated==1.2.14
+dill==0.3.8
+flake8==4.0.1
+frozenlist==1.4.1
+idna==3.7
+isort==5.13.2
+jmespath==1.0.1
+jwt==1.3.1
+mccabe==0.6.1
+multidict==6.0.5
+mypy==1.8.0
+mypy-extensions==1.0.0
+packaging==24.1
+pathspec==0.9.0
+pip==24.1.1
+pipdeptree==2.23.0
+platformdirs==4.2.2
+pycodestyle==2.8.0
+pycparser==2.22
+pyflakes==2.4.0
+PyGithub==2.3.0
+PyJWT==2.8.0
+pylint==3.1.0
+PyNaCl==1.5.0
+python-dateutil==2.9.0.post0
+python-magic==0.4.24
+PyYAML==6.0.1
+rapidfuzz==3.9.3
+requests==2.32.3
+s3transfer==0.10.1
+setuptools==59.6.0
+six==1.16.0
+thefuzz==0.22.1
+tomli==2.0.1
+tomlkit==0.12.5
+tqdm==4.66.4
+types-requests==2.32.0.20240622
+typing_extensions==4.12.2
+unidiff==0.7.5
+urllib3==2.2.2
+wheel==0.37.1
+wrapt==1.16.0
+yamllint==1.26.3
+yarl==1.9.4
--- a/docs/changelogs/v24.6.2.17-stable.md
+++ b/docs/changelogs/v24.6.2.17-stable.md
@ -0,0 +1,26 @@
+---
+sidebar_position: 1
+sidebar_label: 2024
+---
+
+# 2024 Changelog
+
+### ClickHouse release v24.6.2.17-stable (5710a8b5c0c) FIXME as compared to v24.6.1.4423-stable (dcced7c8478)
+
+#### New Feature
+* Backported in [#66002](https://github.com/ClickHouse/ClickHouse/issues/66002): Add AzureQueue storage. [#65458](https://github.com/ClickHouse/ClickHouse/pull/65458) ([Kseniia Sumarokova](https://github.com/kssenii)).
+
+#### Improvement
+* Backported in [#65898](https://github.com/ClickHouse/ClickHouse/issues/65898): Respect cgroup CPU limit in Keeper. [#65819](https://github.com/ClickHouse/ClickHouse/pull/65819) ([Antonio Andelic](https://github.com/antonio2368)).
+
+#### Bug Fix (user-visible misbehavior in an official stable release)
+* Backported in [#65935](https://github.com/ClickHouse/ClickHouse/issues/65935): For queries that read from `PostgreSQL`, cancel the internal `PostgreSQL` query if the ClickHouse query is finished. Otherwise, `ClickHouse` query cannot be canceled until the internal `PostgreSQL` query is finished. [#65771](https://github.com/ClickHouse/ClickHouse/pull/65771) ([Maksim Kita](https://github.com/kitaisreal)).
+
+#### NOT FOR CHANGELOG / INSIGNIFICANT
+
+* Backported in [#65907](https://github.com/ClickHouse/ClickHouse/issues/65907): Fix bug with session closing in Keeper. [#65735](https://github.com/ClickHouse/ClickHouse/pull/65735) ([Antonio Andelic](https://github.com/antonio2368)).
+* Backported in [#65962](https://github.com/ClickHouse/ClickHouse/issues/65962): Add missing workload identity changes. [#65848](https://github.com/ClickHouse/ClickHouse/pull/65848) ([SmitaRKulkarni](https://github.com/SmitaRKulkarni)).
+* Backported in [#66033](https://github.com/ClickHouse/ClickHouse/issues/66033): Follow up to [#65046](https://github.com/ClickHouse/ClickHouse/issues/65046). [#65928](https://github.com/ClickHouse/ClickHouse/pull/65928) ([Kseniia Sumarokova](https://github.com/kssenii)).
+* Backported in [#66076](https://github.com/ClickHouse/ClickHouse/issues/66076): Fix support of non-const scale arguments in rounding functions. [#65983](https://github.com/ClickHouse/ClickHouse/pull/65983) ([Mikhail Gorshkov](https://github.com/mgorshkov)).
+* Backported in [#66017](https://github.com/ClickHouse/ClickHouse/issues/66017): Fix race in s3queue. [#65986](https://github.com/ClickHouse/ClickHouse/pull/65986) ([Kseniia Sumarokova](https://github.com/kssenii)).
+
--- a/docs/en/operations/backup.md
+++ b/docs/en/operations/backup.md
@ -84,6 +84,7 @@ The BACKUP and RESTORE statements take a list of DATABASE and TABLE names, a des
    - [`compression_method`](/docs/en/sql-reference/statements/create/table.md/#column-compression-codecs) and compression_level
    - `password` for the file on disk
    - `base_backup`: the destination of the previous backup of this source.  For example, `Disk('backups', '1.zip')`
+    - `use_same_s3_credentials_for_base_backup`: whether base backup to S3 should inherit credentials from the query. Only works with `S3`.
    - `structure_only`: if enabled, allows to only backup or restore the CREATE statements without the data of tables
    - `storage_policy`: storage policy for the tables being restored. See [Using Multiple Block Devices for Data Storage](../engines/table-engines/mergetree-family/mergetree.md#table_engine-mergetree-multiple-volumes). This setting is only applicable to the `RESTORE` command. The specified storage policy applies only to tables with an engine from the `MergeTree` family.
    - `s3_storage_class`: the storage class used for S3 backup. For example, `STANDARD`
--- a/docs/en/operations/settings/merge-tree-settings.md
+++ b/docs/en/operations/settings/merge-tree-settings.md
@ -974,6 +974,13 @@ Default value: false

 - [exclude_deleted_rows_for_part_size_in_merge](#exclude_deleted_rows_for_part_size_in_merge) setting

+## use_compact_variant_discriminators_serialization {#use_compact_variant_discriminators_serialization}
+
+Enables compact mode for binary serialization of discriminators in Variant data type.
+This mode allows to use significantly less memory for storing discriminators in parts when there is mostly one variant or a lot of NULL values.
+
+Default value: true
+
 ## merge_workload

 Used to regulate how resources are utilized and shared between merges and other workloads. Specified value is used as `workload` setting value for background merges of this table. If not specified (empty string), then server setting `merge_workload` is used instead.
--- a/docs/en/sql-reference/aggregate-functions/reference/aggthrow.md
+++ b/docs/en/sql-reference/aggregate-functions/reference/aggthrow.md
@ -0,0 +1,37 @@
+---
+slug: /en/sql-reference/aggregate-functions/reference/aggthrow
+sidebar_position: 101
+---
+
+# aggThrow
+
+This function can be used for the purpose of testing exception safety. It will throw an exception on creation with the specified probability.
+
+**Syntax**
+
+```sql
+aggThrow(throw_prob)
+```
+
+**Arguments**
+
+- `throw_prob` — Probability to throw on creation. [Float64](../../data-types/float.md).
+
+**Returned value**
+
+- An exception: `Code: 503. DB::Exception: Aggregate function aggThrow has thrown exception successfully`.
+
+**Example**
+
+Query:
+
+```sql
+SELECT number % 2 AS even, aggThrow(number) FROM numbers(10) GROUP BY even;
+```
+
+Result:
+
+```response
+Received exception:
+Code: 503. DB::Exception: Aggregate function aggThrow has thrown exception successfully: While executing AggregatingTransform. (AGGREGATE_FUNCTION_THROW)
+```
--- a/docs/en/sql-reference/aggregate-functions/reference/index.md
+++ b/docs/en/sql-reference/aggregate-functions/reference/index.md
@ -43,6 +43,7 @@ Standard aggregate functions:

 ClickHouse-specific aggregate functions:

+- [aggThrow](../reference/aggthrow.md)
 - [analysisOfVariance](../reference/analysis_of_variance.md)
 - [any](../reference/any_respect_nulls.md)
 - [anyHeavy](../reference/anyheavy.md)
--- a/docs/en/sql-reference/functions/other-functions.md
+++ b/docs/en/sql-reference/functions/other-functions.md
@ -86,7 +86,7 @@ Returns the fully qualified domain name of the ClickHouse server.
 fqdn();
 ```

-This function is case-insensitive.
+Aliases: `fullHostName`, 'FQDN'. 

 **Returned value**

--- a/docs/en/sql-reference/window-functions/index.md
+++ b/docs/en/sql-reference/window-functions/index.md
@ -23,6 +23,7 @@ ClickHouse supports the standard grammar for defining windows and window functio
 | `GROUPS` frame                                                                     | ❌                                                                                                                                                                               |
 | Calculating aggregate functions over a frame (`sum(value) over (order by time)`)   | ✅ (All aggregate functions are supported)                                                                                                                                                       |
 | `rank()`, `dense_rank()`, `row_number()`                                           | ✅                                                                                                                                                                                   |
+| `percent_rank()` | ✅  Efficiently computes the relative standing of a value within a partition in a dataset. This function effectively replaces the more verbose and computationally intensive manual SQL calculation expressed as `ifNull((rank() OVER(PARTITION BY x ORDER BY y) - 1) / nullif(count(1) OVER(PARTITION BY x) - 1, 0), 0)`| 
 | `lag/lead(value, offset)`                                                          | ❌ <br/> You can use one of the following workarounds:<br/> 1) `any(value) over (.... rows between <offset> preceding and <offset> preceding)`, or `following` for `lead` <br/> 2) `lagInFrame/leadInFrame`, which are analogous, but respect the window frame. To get behavior identical to `lag/lead`, use `rows between unbounded preceding and unbounded following`                                                                 |
 | ntile(buckets) | ✅ <br/> Specify window like, (partition by x order by y rows between unbounded preceding and unrounded following). |

--- a/programs/server/Server.cpp
+++ b/programs/server/Server.cpp
@ -626,6 +626,28 @@ static void initializeAzureSDKLogger(
 #endif
 }

+#if defined(SANITIZER)
+static std::vector<String> getSanitizerNames()
+{
+    std::vector<String> names;
+
+#if defined(ADDRESS_SANITIZER)
+    names.push_back("address");
+#endif
+#if defined(THREAD_SANITIZER)
+    names.push_back("thread");
+#endif
+#if defined(MEMORY_SANITIZER)
+    names.push_back("memory");
+#endif
+#if defined(UNDEFINED_BEHAVIOR_SANITIZER)
+    names.push_back("undefined behavior");
+#endif
+
+    return names;
+}
+#endif
+
 int Server::main(const std::vector<std::string> & /*args*/)
 try
 {
@ -716,7 +738,17 @@ try
        global_context->addWarningMessage("ThreadFuzzer is enabled. Application will run slowly and unstable.");

 #if defined(SANITIZER)
-    global_context->addWarningMessage("Server was built with sanitizer. It will work slowly.");
+    auto sanitizers = getSanitizerNames();
+
+    String log_message;
+    if (sanitizers.empty())
+        log_message = "sanitizer";
+    else if (sanitizers.size() == 1)
+        log_message = fmt::format("{} sanitizer", sanitizers.front());
+    else
+        log_message = fmt::format("sanitizers ({})", fmt::join(sanitizers, ", "));
+
+    global_context->addWarningMessage(fmt::format("Server was built with {}. It will work slowly.", log_message));
 #endif

 #if defined(SANITIZE_COVERAGE) || WITH_COVERAGE
--- a/src/Columns/ColumnObject.cpp
+++ b/src/Columns/ColumnObject.cpp
@ -1093,10 +1093,4 @@ void ColumnObject::finalize()
    checkObjectHasNoAmbiguosPaths(getKeys());
 }

-void ColumnObject::updateHashFast(SipHash & hash) const
-{
-    for (const auto & entry : subcolumns)
-        for (auto & part : entry->data.data)
-            part->updateHashFast(hash);
-}
 }
--- a/src/Columns/ColumnObject.h
+++ b/src/Columns/ColumnObject.h
@ -242,7 +242,7 @@ public:
    const char * skipSerializedInArena(const char *) const override { throwMustBeConcrete(); }
    void updateHashWithValue(size_t, SipHash &) const override { throwMustBeConcrete(); }
    void updateWeakHash32(WeakHash32 &) const override { throwMustBeConcrete(); }
-    void updateHashFast(SipHash & hash) const override;
+    void updateHashFast(SipHash &) const override { throwMustBeConcrete(); }
    void expand(const Filter &, bool) override { throwMustBeConcrete(); }
    bool hasEqualValues() const override { throwMustBeConcrete(); }
    size_t byteSizeAt(size_t) const override { throwMustBeConcrete(); }
--- a/src/Columns/ColumnTuple.cpp
+++ b/src/Columns/ColumnTuple.cpp
@ -711,7 +711,13 @@ void ColumnTuple::takeDynamicStructureFromSourceColumns(const Columns & source_c
 ColumnPtr ColumnTuple::compress() const
 {
    if (columns.empty())
-        return Ptr();
+    {
+        return ColumnCompressed::create(size(), 0,
+            [n = column_length]
+            {
+                return ColumnTuple::create(n);
+            });
+    }

    size_t byte_size = 0;
    Columns compressed;
--- a/src/Common/BinStringDecodeHelper.h
+++ b/src/Common/BinStringDecodeHelper.h
@ -5,7 +5,7 @@
 namespace DB
 {

-static void inline hexStringDecode(const char * pos, const char * end, char *& out, size_t word_size = 2)
+static void inline hexStringDecode(const char * pos, const char * end, char *& out, size_t word_size)
 {
    if ((end - pos) & 1)
    {
@ -23,7 +23,7 @@ static void inline hexStringDecode(const char * pos, const char * end, char *& o
    ++out;
 }

-static void inline binStringDecode(const char * pos, const char * end, char *& out)
+static void inline binStringDecode(const char * pos, const char * end, char *& out, size_t word_size)
 {
    if (pos == end)
    {
@ -53,7 +53,7 @@ static void inline binStringDecode(const char * pos, const char * end, char *& o
        ++out;
    }

-    assert((end - pos) % 8 == 0);
+    chassert((end - pos) % word_size == 0);

    while (end - pos != 0)
    {
--- a/src/Common/CollectionOfDerived.h
+++ b/src/Common/CollectionOfDerived.h
@ -1,184 +0,0 @@
-#pragma once
-
-#include <base/defines.h>
-
-#include <Common/Exception.h>
-
-#include <algorithm>
-#include <memory>
-#include <typeindex>
-#include <vector>
-#include <string>
-
-
-namespace DB
-{
-
-namespace ErrorCodes
-{
-    extern const int LOGICAL_ERROR;
-}
-
-/* This is a collections of objects derived from ItemBase.
-*  Collection contains no more than one instance for each derived type.
-*  The derived type is used to access the instance.
-*/
-
-template<class ItemBase>
-class CollectionOfDerivedItems
-{
-public:
-    using Self = CollectionOfDerivedItems<ItemBase>;
-    using ItemPtr = std::shared_ptr<ItemBase>;
-
-private:
-    struct Rec
-    {
-        std::type_index type_idx;
-        ItemPtr ptr;
-
-        bool operator<(const Rec & other) const
-        {
-            return type_idx < other.type_idx;
-        }
-
-        bool operator<(const std::type_index & value) const
-        {
-            return type_idx < value;
-        }
-
-        bool operator==(const Rec & other) const
-        {
-            return type_idx == other.type_idx;
-        }
-    };
-    using Records = std::vector<Rec>;
-
-public:
-    void swap(Self & other) noexcept
-    {
-        records.swap(other.records);
-    }
-
-    void clear()
-    {
-        records.clear();
-    }
-
-    bool empty() const
-    {
-        return records.empty();
-    }
-
-    size_t size() const
-    {
-        return records.size();
-    }
-
-    Self clone() const
-    {
-        Self result;
-        result.records.reserve(records.size());
-        for (const auto & rec : records)
-            result.records.emplace_back(rec.type_idx, rec.ptr->clone());
-        return result;
-    }
-
-    void append(Self && other)
-    {
-        auto middle_idx = records.size();
-        std::move(other.records.begin(), other.records.end(), std::back_inserter(records));
-        std::inplace_merge(records.begin(), records.begin() + middle_idx, records.end());
-        chassert(isUniqTypes());
-    }
-
-    template <class T>
-    void add(std::shared_ptr<T> info)
-    {
-        static_assert(std::is_base_of_v<ItemBase, T>, "Template parameter must inherit items base class");
-        return addImpl(std::type_index(typeid(T)), std::move(info));
-    }
-
-    template <class T>
-    std::shared_ptr<T> get() const
-    {
-        static_assert(std::is_base_of_v<ItemBase, T>, "Template parameter must inherit items base class");
-        auto it = getImpl(std::type_index(typeid(T)));
-        if (it == records.cend())
-            return nullptr;
-        auto cast = std::dynamic_pointer_cast<T>(it->ptr);
-        chassert(cast);
-        return cast;
-    }
-
-    template <class T>
-    std::shared_ptr<T> extract()
-    {
-        static_assert(std::is_base_of_v<ItemBase, T>, "Template parameter must inherit items base class");
-        auto it = getImpl(std::type_index(typeid(T)));
-        if (it == records.cend())
-            return nullptr;
-        auto cast = std::dynamic_pointer_cast<T>(it->ptr);
-        chassert(cast);
-
-        records.erase(it);
-        return cast;
-    }
-
-    std::string debug() const
-    {
-        std::string result;
-
-        for (auto & rec : records)
-        {
-            result.append(rec.type_idx.name());
-            result.append(" ");
-        }
-
-        return result;
-    }
-
-private:
-    bool isUniqTypes() const
-    {
-        auto uniq_it = std::adjacent_find(records.begin(), records.end());
-
-        return uniq_it == records.end();
-    }
-
-    void addImpl(std::type_index type_idx, ItemPtr item)
-    {
-        auto it = std::lower_bound(records.begin(), records.end(), type_idx);
-
-        if (it == records.end())
-        {
-            records.emplace_back(type_idx, item);
-            return;
-        }
-
-        if (it->type_idx == type_idx)
-            throw Exception(ErrorCodes::LOGICAL_ERROR, "inserted items must be unique by their type, type {} is inserted twice", type_idx.name());
-
-
-        records.emplace(it, type_idx, item);
-
-        chassert(isUniqTypes());
-    }
-
-    Records::const_iterator getImpl(std::type_index type_idx) const
-    {
-        auto it = std::lower_bound(records.cbegin(), records.cend(), type_idx);
-
-        if (it == records.cend())
-            return records.cend();
-
-        if (it->type_idx != type_idx)
-            return records.cend();
-
-        return it;
-    }
-
-    Records records;
-};
-
-}
--- a/src/Common/GWPAsan.cpp
+++ b/src/Common/GWPAsan.cpp
@ -57,9 +57,12 @@ static bool guarded_alloc_initialized = []
        opts.MaxSimultaneousAllocations = 1024;

    if (!env_options_raw || !std::string_view{env_options_raw}.contains("SampleRate"))
-        opts.SampleRate = 50000;
+        opts.SampleRate = 10000;
+
+    const char * collect_stacktraces = std::getenv("GWP_ASAN_COLLECT_STACKTRACES"); // NOLINT(concurrency-mt-unsafe)
+    if (collect_stacktraces && std::string_view{collect_stacktraces} == "1")
+        opts.Backtrace = getBackTrace;

-    opts.Backtrace = getBackTrace;
    GuardedAlloc.init(opts);

    return true;
--- a/src/Core/ServerSettings.h
+++ b/src/Core/ServerSettings.h
@ -153,7 +153,7 @@ namespace DB
    M(Bool, enable_azure_sdk_logging, false, "Enables logging from Azure sdk", 0) \
    M(String, merge_workload, "default", "Name of workload to be used to access resources for all merges (may be overridden by a merge tree setting)", 0) \
    M(String, mutation_workload, "default", "Name of workload to be used to access resources for all mutations (may be overridden by a merge tree setting)", 0) \
-    M(Double, gwp_asan_force_sample_probability, 0, "Probability that an allocation from specific places will be sampled by GWP Asan (i.e. PODArray allocations)", 0) \
+    M(Double, gwp_asan_force_sample_probability, 0.0003, "Probability that an allocation from specific places will be sampled by GWP Asan (i.e. PODArray allocations)", 0) \
    M(UInt64, config_reload_interval_ms, 2000, "How often clickhouse will reload config and check for new changes", 0) \

 /// If you add a setting which can be updated at runtime, please update 'changeable_settings' map in StorageSystemServerSettings.cpp
--- a/src/Core/Settings.h
+++ b/src/Core/Settings.h
@ -36,7 +36,7 @@ class IColumn;
    M(Dialect, dialect, Dialect::clickhouse, "Which dialect will be used to parse query", 0)\
    M(UInt64, min_compress_block_size, 65536, "The actual size of the block to compress, if the uncompressed data less than max_compress_block_size is no less than this value and no less than the volume of data for one mark.", 0) \
    M(UInt64, max_compress_block_size, 1048576, "The maximum size of blocks of uncompressed data before compressing for writing to a table.", 0) \
-    M(UInt64, max_block_size, DEFAULT_BLOCK_SIZE, "Maximum block size in rows for reading", 0) \
+    M(UInt64, max_block_size, DEFAULT_BLOCK_SIZE, "Maximum block size for reading", 0) \
    M(UInt64, max_insert_block_size, DEFAULT_INSERT_BLOCK_SIZE, "The maximum block size for insertion, if we control the creation of blocks for insertion.", 0) \
    M(UInt64, min_insert_block_size_rows, DEFAULT_INSERT_BLOCK_SIZE, "Squash blocks passed to INSERT query to specified size in rows, if blocks are not big enough.", 0) \
    M(UInt64, min_insert_block_size_bytes, (DEFAULT_INSERT_BLOCK_SIZE * 256), "Squash blocks passed to INSERT query to specified size in bytes, if blocks are not big enough.", 0) \
@ -634,8 +634,9 @@ class IColumn;
    M(Bool, optimize_time_filter_with_preimage, true, "Optimize Date and DateTime predicates by converting functions into equivalent comparisons without conversions (e.g. toYear(col) = 2023 -> col >= '2023-01-01' AND col <= '2023-12-31')", 0) \
    M(Bool, normalize_function_names, true, "Normalize function names to their canonical names", 0) \
    M(Bool, enable_early_constant_folding, true, "Enable query optimization where we analyze function and subqueries results and rewrite query if there are constants there", 0) \
-    M(Bool, deduplicate_blocks_in_dependent_materialized_views, false, "Should deduplicate blocks for materialized views. Use true to always deduplicate in dependent tables.", 0) \
+    M(Bool, deduplicate_blocks_in_dependent_materialized_views, false, "Should deduplicate blocks for materialized views if the block is not a duplicate for the table. Use true to always deduplicate in dependent tables.", 0) \
    M(Bool, throw_if_deduplication_in_dependent_materialized_views_enabled_with_async_insert, true, "Throw exception on INSERT query when the setting `deduplicate_blocks_in_dependent_materialized_views` is enabled along with `async_insert`. It guarantees correctness, because these features can't work together.", 0) \
+    M(Bool, update_insert_deduplication_token_in_dependent_materialized_views, false, "Should update insert deduplication token with table identifier during insert in dependent materialized views.", 0) \
    M(Bool, materialized_views_ignore_errors, false, "Allows to ignore errors for MATERIALIZED VIEW, and deliver original block to the table regardless of MVs", 0) \
    M(Bool, ignore_materialized_views_with_dropped_target_table, false, "Ignore MVs with dropped target table during pushing to views", 0) \
    M(Bool, allow_experimental_refreshable_materialized_view, false, "Allow refreshable materialized views (CREATE MATERIALIZED VIEW <name> REFRESH ...).", 0) \
@ -953,7 +954,6 @@ class IColumn;

 #define OBSOLETE_SETTINGS(M, ALIAS) \
    /** Obsolete settings that do nothing but left for compatibility reasons. Remove each one after half a year of obsolescence. */ \
-    MAKE_OBSOLETE(M, Bool, update_insert_deduplication_token_in_dependent_materialized_views, 1) \
    MAKE_OBSOLETE(M, UInt64, max_memory_usage_for_all_queries, 0) \
    MAKE_OBSOLETE(M, UInt64, multiple_joins_rewriter_version, 0) \
    MAKE_OBSOLETE(M, Bool, enable_debug_queries, false) \
--- a/src/DataTypes/Serializations/ISerialization.h
+++ b/src/DataTypes/Serializations/ISerialization.h
@ -256,6 +256,8 @@ public:

        bool position_independent_encoding = true;

+        bool use_compact_variant_discriminators_serialization = false;
+
        enum class DynamicStatisticsMode
        {
            NONE,   /// Don't write statistics.
--- a/src/DataTypes/Serializations/SerializationVariant.cpp
+++ b/src/DataTypes/Serializations/SerializationVariant.cpp
@ -30,12 +30,18 @@ namespace ErrorCodes

 struct SerializeBinaryBulkStateVariant : public ISerialization::SerializeBinaryBulkState
 {
-    std::vector<ISerialization::SerializeBinaryBulkStatePtr> states;
+    explicit SerializeBinaryBulkStateVariant(UInt64 mode) : discriminators_mode(mode)
+    {
+    }
+
+    SerializationVariant::DiscriminatorsSerializationMode discriminators_mode;
+    std::vector<ISerialization::SerializeBinaryBulkStatePtr> variant_states;
 };

 struct DeserializeBinaryBulkStateVariant : public ISerialization::DeserializeBinaryBulkState
 {
-    std::vector<ISerialization::DeserializeBinaryBulkStatePtr> states;
+    ISerialization::DeserializeBinaryBulkStatePtr discriminators_state;
+    std::vector<ISerialization::DeserializeBinaryBulkStatePtr> variant_states;
 };

 void SerializationVariant::enumerateStreams(
@ -71,7 +77,7 @@ void SerializationVariant::enumerateStreams(
                             .withType(type_variant ? type_variant->getVariant(i) : nullptr)
                             .withColumn(column_variant ? column_variant->getVariantPtrByGlobalDiscriminator(i) : nullptr)
                             .withSerializationInfo(data.serialization_info)
-                             .withDeserializeState(variant_deserialize_state ? variant_deserialize_state->states[i] : nullptr);
+                             .withDeserializeState(variant_deserialize_state ? variant_deserialize_state->variant_states[i] : nullptr);

        addVariantElementToPath(settings.path, i);
        settings.path.back().data = variant_data;
@ -87,17 +93,26 @@ void SerializationVariant::serializeBinaryBulkStatePrefix(
    SerializeBinaryBulkSettings & settings,
    SerializeBinaryBulkStatePtr & state) const
 {
-    const ColumnVariant & col = assert_cast<const ColumnVariant &>(column);
+    settings.path.push_back(Substream::VariantDiscriminators);
+    auto * discriminators_stream = settings.getter(settings.path);
+    settings.path.pop_back();

-    auto variant_state = std::make_shared<SerializeBinaryBulkStateVariant>();
-    variant_state->states.resize(variants.size());
+    if (!discriminators_stream)
+        throw Exception(ErrorCodes::LOGICAL_ERROR, "Got empty stream for VariantDiscriminators in SerializationVariant::serializeBinaryBulkStatePrefix");
+
+    UInt64 mode = settings.use_compact_variant_discriminators_serialization ? DiscriminatorsSerializationMode::COMPACT : DiscriminatorsSerializationMode::BASIC;
+    writeBinaryLittleEndian(mode, *discriminators_stream);
+
+    const ColumnVariant & col = assert_cast<const ColumnVariant &>(column);
+    auto variant_state = std::make_shared<SerializeBinaryBulkStateVariant>(mode);
+    variant_state->variant_states.resize(variants.size());

    settings.path.push_back(Substream::VariantElements);

    for (size_t i = 0; i < variants.size(); ++i)
    {
        addVariantElementToPath(settings.path, i);
-        variants[i]->serializeBinaryBulkStatePrefix(col.getVariantByGlobalDiscriminator(i), settings, variant_state->states[i]);
+        variants[i]->serializeBinaryBulkStatePrefix(col.getVariantByGlobalDiscriminator(i), settings, variant_state->variant_states[i]);
        settings.path.pop_back();
    }

@ -116,7 +131,7 @@ void SerializationVariant::serializeBinaryBulkStateSuffix(
    for (size_t i = 0; i < variants.size(); ++i)
    {
        addVariantElementToPath(settings.path, i);
-        variants[i]->serializeBinaryBulkStateSuffix(settings, variant_state->states[i]);
+        variants[i]->serializeBinaryBulkStateSuffix(settings, variant_state->variant_states[i]);
        settings.path.pop_back();
    }
    settings.path.pop_back();
@ -128,14 +143,19 @@ void SerializationVariant::deserializeBinaryBulkStatePrefix(
    DeserializeBinaryBulkStatePtr & state,
    SubstreamsDeserializeStatesCache * cache) const
 {
+    DeserializeBinaryBulkStatePtr discriminators_state = deserializeDiscriminatorsStatePrefix(settings, cache);
+    if (!discriminators_state)
+        return;
+
    auto variant_state = std::make_shared<DeserializeBinaryBulkStateVariant>();
-    variant_state->states.resize(variants.size());
+    variant_state->discriminators_state = discriminators_state;
+    variant_state->variant_states.resize(variants.size());

    settings.path.push_back(Substream::VariantElements);
    for (size_t i = 0; i < variants.size(); ++i)
    {
        addVariantElementToPath(settings.path, i);
-        variants[i]->deserializeBinaryBulkStatePrefix(settings, variant_state->states[i], cache);
+        variants[i]->deserializeBinaryBulkStatePrefix(settings, variant_state->variant_states[i], cache);
        settings.path.pop_back();
    }

@ -143,6 +163,29 @@ void SerializationVariant::deserializeBinaryBulkStatePrefix(
    state = std::move(variant_state);
 }

+ISerialization::DeserializeBinaryBulkStatePtr SerializationVariant::deserializeDiscriminatorsStatePrefix(
+    DeserializeBinaryBulkSettings & settings,
+    SubstreamsDeserializeStatesCache * cache)
+{
+    settings.path.push_back(Substream::VariantDiscriminators);
+
+    DeserializeBinaryBulkStatePtr discriminators_state = nullptr;
+    if (auto cached_state = getFromSubstreamsDeserializeStatesCache(cache, settings.path))
+    {
+        discriminators_state = cached_state;
+    }
+    else if (auto * discriminators_stream = settings.getter(settings.path))
+    {
+        UInt64 mode;
+        readBinaryLittleEndian(mode, *discriminators_stream);
+        discriminators_state = std::make_shared<DeserializeBinaryBulkStateVariantDiscriminators>(mode);
+        addToSubstreamsDeserializeStatesCache(cache, settings.path, discriminators_state);
+    }
+
+    settings.path.pop_back();
+    return discriminators_state;
+}
+

 void SerializationVariant::serializeBinaryBulkWithMultipleStreamsAndUpdateVariantStatistics(
    const IColumn & column,
@ -165,13 +208,71 @@ void SerializationVariant::serializeBinaryBulkWithMultipleStreamsAndUpdateVarian

    auto * variant_state = checkAndGetState<SerializeBinaryBulkStateVariant>(state);

-    /// If offset = 0 and limit == col.size() or we have only NULLs, we don't need to calculate
+    /// Don't write anything if column is empty.
+    if (limit == 0)
+        return;
+
+    /// Write number of rows in this granule in compact mode.
+    if (variant_state->discriminators_mode.value == DiscriminatorsSerializationMode::COMPACT)
+        writeVarUInt(UInt64(limit), *discriminators_stream);
+
+    /// If column has only one none empty discriminators and no NULLs we don't need to
+    /// calculate limits for variants and use provided offset/limit.
+    if (auto non_empty_local_discr = col.getLocalDiscriminatorOfOneNoneEmptyVariantNoNulls())
+    {
+        auto non_empty_global_discr = col.globalDiscriminatorByLocal(*non_empty_local_discr);
+
+        /// In compact mode write the format of the granule and single non-empty discriminator.
+        if (variant_state->discriminators_mode.value == DiscriminatorsSerializationMode::COMPACT)
+        {
+            writeBinaryLittleEndian(UInt8(CompactDiscriminatorsGranuleFormat::COMPACT), *discriminators_stream);
+            writeBinaryLittleEndian(non_empty_global_discr, *discriminators_stream);
+        }
+        /// For basic mode just serialize this discriminator limit times.
+        else
+        {
+            for (size_t i = 0; i < limit; ++i)
+                writeBinaryLittleEndian(non_empty_global_discr, *discriminators_stream);
+        }
+
+        settings.path.push_back(Substream::VariantElements);
+        addVariantElementToPath(settings.path, non_empty_global_discr);
+        /// We can use the same offset/limit as for whole Variant column
+        variants[non_empty_global_discr]->serializeBinaryBulkWithMultipleStreams(col.getVariantByGlobalDiscriminator(non_empty_global_discr), offset, limit, settings, variant_state->variant_states[non_empty_global_discr]);
+        variants_statistics[variant_names[non_empty_global_discr]] += limit;
+        settings.path.pop_back();
+        settings.path.pop_back();
+        return;
+    }
+    /// If column has only NULLs, just serialize NULL discriminators.
+    else if (col.hasOnlyNulls())
+    {
+        /// In compact mode write single NULL_DISCRIMINATOR.
+        if (variant_state->discriminators_mode.value == DiscriminatorsSerializationMode::COMPACT)
+        {
+            writeBinaryLittleEndian(UInt8(CompactDiscriminatorsGranuleFormat::COMPACT), *discriminators_stream);
+            writeBinaryLittleEndian(ColumnVariant::NULL_DISCRIMINATOR, *discriminators_stream);
+        }
+        /// In basic mode write NULL_DISCRIMINATOR limit times.
+        else
+        {
+            for (size_t i = 0; i < limit; ++i)
+                writeBinaryLittleEndian(ColumnVariant::NULL_DISCRIMINATOR, *discriminators_stream);
+        }
+        return;
+    }
+
+    /// If offset = 0 and limit == col.size() we don't need to calculate
    /// offsets and limits for variants and need to just serialize whole columns.
-    if ((offset == 0 && limit == col.size()) || col.hasOnlyNulls())
+    if ((offset == 0 && limit == col.size()))
    {
        /// First, serialize discriminators.
-        /// If we have only NULLs or local and global discriminators are the same, just serialize the column as is.
-        if (col.hasOnlyNulls() || col.hasGlobalVariantsOrder())
+        /// Here we are sure that column contains different discriminators, use plain granule format in compact mode.
+        if (variant_state->discriminators_mode.value == DiscriminatorsSerializationMode::COMPACT)
+            writeBinaryLittleEndian(UInt8(CompactDiscriminatorsGranuleFormat::PLAIN), *discriminators_stream);
+
+        /// If local and global discriminators are the same, just serialize the column as is.
+        if (col.hasGlobalVariantsOrder())
        {
            SerializationNumber<ColumnVariant::Discriminator>().serializeBinaryBulk(col.getLocalDiscriminatorsColumn(), *discriminators_stream, offset, limit);
        }
@ -188,7 +289,7 @@ void SerializationVariant::serializeBinaryBulkWithMultipleStreamsAndUpdateVarian
        for (size_t i = 0; i != variants.size(); ++i)
        {
            addVariantElementToPath(settings.path, i);
-            variants[i]->serializeBinaryBulkWithMultipleStreams(col.getVariantByGlobalDiscriminator(i), 0, 0, settings, variant_state->states[i]);
+            variants[i]->serializeBinaryBulkWithMultipleStreams(col.getVariantByGlobalDiscriminator(i), 0, 0, settings, variant_state->variant_states[i]);
            variants_statistics[variant_names[i]] += col.getVariantByGlobalDiscriminator(i).size();
            settings.path.pop_back();
        }
@ -196,36 +297,16 @@ void SerializationVariant::serializeBinaryBulkWithMultipleStreamsAndUpdateVarian
        return;
    }

-    /// If we have only one non empty variant and no NULLs, we can use the same limit offset for this variant.
-    if (auto non_empty_local_discr = col.getLocalDiscriminatorOfOneNoneEmptyVariantNoNulls())
-    {
-        /// First, serialize discriminators.
-        /// We know that all discriminators are the same, so we just need to serialize this discriminator limit times.
-        auto non_empty_global_discr = col.globalDiscriminatorByLocal(*non_empty_local_discr);
-        for (size_t i = 0; i != limit; ++i)
-            writeBinaryLittleEndian(non_empty_global_discr, *discriminators_stream);
-
-        /// Second, serialize non-empty variant (other variants are empty and we can skip their serialization).
-        settings.path.push_back(Substream::VariantElements);
-        addVariantElementToPath(settings.path, non_empty_global_discr);
-        /// We can use the same offset/limit as for whole Variant column
-        variants[non_empty_global_discr]->serializeBinaryBulkWithMultipleStreams(col.getVariantByGlobalDiscriminator(non_empty_global_discr), offset, limit, settings, variant_state->states[non_empty_global_discr]);
-        variants_statistics[variant_names[non_empty_global_discr]] += limit;
-        settings.path.pop_back();
-        settings.path.pop_back();
-        return;
-    }
-
    /// In general case we should iterate through local discriminators in range [offset, offset + limit] to serialize global discriminators and calculate offset/limit pair for each variant.
    const auto & local_discriminators = col.getLocalDiscriminators();
    const auto & offsets = col.getOffsets();
    std::vector<std::pair<size_t, size_t>> variant_offsets_and_limits(variants.size(), {0, 0});
    size_t end = offset + limit;
+    size_t num_non_empty_variants_in_range = 0;
+    ColumnVariant::Discriminator last_non_empty_variant_discr = 0;
    for (size_t i = offset; i < end; ++i)
    {
        auto global_discr = col.globalDiscriminatorByLocal(local_discriminators[i]);
-        writeBinaryLittleEndian(global_discr, *discriminators_stream);
-
        if (global_discr != ColumnVariant::NULL_DISCRIMINATOR)
        {
            /// If we see this discriminator for the first time, update offset
@ -233,9 +314,38 @@ void SerializationVariant::serializeBinaryBulkWithMultipleStreamsAndUpdateVarian
                variant_offsets_and_limits[global_discr].first = offsets[i];
            /// Update limit for this discriminator.
            ++variant_offsets_and_limits[global_discr].second;
+            ++num_non_empty_variants_in_range;
+            last_non_empty_variant_discr = global_discr;
        }
    }

+    /// In basic mode just serialize discriminators as is row by row.
+    if (variant_state->discriminators_mode.value == DiscriminatorsSerializationMode::BASIC)
+    {
+        for (size_t i = offset; i < end; ++i)
+            writeBinaryLittleEndian(col.globalDiscriminatorByLocal(local_discriminators[i]), *discriminators_stream);
+    }
+    /// In compact mode check if we have the same discriminator for all rows in this granule.
+    /// First, check if all values in granule are NULLs.
+    else if (num_non_empty_variants_in_range == 0)
+    {
+        writeBinaryLittleEndian(UInt8(CompactDiscriminatorsGranuleFormat::COMPACT), *discriminators_stream);
+        writeBinaryLittleEndian(ColumnVariant::NULL_DISCRIMINATOR, *discriminators_stream);
+    }
+    /// Then, check if there is only 1 variant and no NULLs in this granule.
+    else if (num_non_empty_variants_in_range == 1 && variant_offsets_and_limits[last_non_empty_variant_discr].second == limit)
+    {
+        writeBinaryLittleEndian(UInt8(CompactDiscriminatorsGranuleFormat::COMPACT), *discriminators_stream);
+        writeBinaryLittleEndian(last_non_empty_variant_discr, *discriminators_stream);
+    }
+    /// Otherwise there are different discriminators in this granule.
+    else
+    {
+        writeBinaryLittleEndian(UInt8(CompactDiscriminatorsGranuleFormat::PLAIN), *discriminators_stream);
+        for (size_t i = offset; i < end; ++i)
+            writeBinaryLittleEndian(col.globalDiscriminatorByLocal(local_discriminators[i]), *discriminators_stream);
+    }
+
    /// Serialize variants in global order.
    settings.path.push_back(Substream::VariantElements);
    for (size_t i = 0; i != variants.size(); ++i)
@ -249,7 +359,7 @@ void SerializationVariant::serializeBinaryBulkWithMultipleStreamsAndUpdateVarian
                variant_offsets_and_limits[i].first,
                variant_offsets_and_limits[i].second,
                settings,
-                variant_state->states[i]);
+                variant_state->variant_states[i]);
            variants_statistics[variant_names[i]] += variant_offsets_and_limits[i].second;
            settings.path.pop_back();
        }
@ -284,39 +394,68 @@ void SerializationVariant::deserializeBinaryBulkWithMultipleStreams(

    /// First, deserialize discriminators.
    settings.path.push_back(Substream::VariantDiscriminators);
+
+    DeserializeBinaryBulkStateVariant * variant_state = nullptr;
+    std::vector<size_t> variant_limits;
    if (auto cached_discriminators = getFromSubstreamsCache(cache, settings.path))
    {
+        variant_state = checkAndGetState<DeserializeBinaryBulkStateVariant>(state);
        col.getLocalDiscriminatorsPtr() = cached_discriminators;
    }
-    else
+    else if (auto * discriminators_stream = settings.getter(settings.path))
    {
-        auto * discriminators_stream = settings.getter(settings.path);
-        if (!discriminators_stream)
-            return;
+        variant_state = checkAndGetState<DeserializeBinaryBulkStateVariant>(state);
+        auto * discriminators_state = checkAndGetState<DeserializeBinaryBulkStateVariantDiscriminators>(variant_state->discriminators_state);
+
+        /// Deserialize discriminators according to serialization mode.
+        if (discriminators_state->mode.value == DiscriminatorsSerializationMode::BASIC)
+            SerializationNumber<ColumnVariant::Discriminator>().deserializeBinaryBulk(*col.getLocalDiscriminatorsPtr()->assumeMutable(), *discriminators_stream, limit, 0);
+        else
+            variant_limits = deserializeCompactDiscriminators(col.getLocalDiscriminatorsPtr(), limit, discriminators_stream, settings.continuous_reading, *discriminators_state);

-        SerializationNumber<ColumnVariant::Discriminator>().deserializeBinaryBulk(*col.getLocalDiscriminatorsPtr()->assumeMutable(), *discriminators_stream, limit, 0);
        addToSubstreamsCache(cache, settings.path, col.getLocalDiscriminatorsPtr());
    }
+    /// It may happen that there is no such stream, in this case just do nothing.
+    else
+    {
+        settings.path.pop_back();
+        return;
+    }
+
    settings.path.pop_back();

-    /// Second, calculate limits for each variant by iterating through new discriminators.
-    std::vector<size_t> variant_limits(variants.size(), 0);
-    auto & discriminators_data = col.getLocalDiscriminators();
-    size_t discriminators_offset = discriminators_data.size() - limit;
-    for (size_t i = discriminators_offset; i != discriminators_data.size(); ++i)
+    /// Second, calculate limits for each variant by iterating through new discriminators
+    /// if we didn't do it during discriminators deserialization.
+    if (variant_limits.empty())
    {
-        ColumnVariant::Discriminator discr = discriminators_data[i];
-        if (discr != ColumnVariant::NULL_DISCRIMINATOR)
-            ++variant_limits[discr];
+        variant_limits.resize(variants.size(), 0);
+        auto & discriminators_data = col.getLocalDiscriminators();
+
+        /// We can actually read less than limit discriminators and we cannot determine the actual number of read rows
+        /// by discriminators column as it could be taken from the substreams cache. And we need actual number of read
+        /// rows to fill offsets correctly later if they are not in the cache. We can determine if offsets column is in cache
+        /// or not by comparing it with discriminators column size (they should be the same when offsets are in cache).
+        /// If offsets are not in the cache, we can use it's size to determine the actual number of read rows.
+        size_t num_new_discriminators = limit;
+        size_t offsets_size = col.getOffsetsPtr()->size();
+        if (discriminators_data.size() > offsets_size)
+            num_new_discriminators = discriminators_data.size() - offsets_size;
+        size_t discriminators_offset = discriminators_data.size() - num_new_discriminators;
+
+        for (size_t i = discriminators_offset; i != discriminators_data.size(); ++i)
+        {
+            ColumnVariant::Discriminator discr = discriminators_data[i];
+            if (discr != ColumnVariant::NULL_DISCRIMINATOR)
+                ++variant_limits[discr];
+        }
    }

    /// Now we can deserialize variants according to their limits.
-    auto * variant_state = checkAndGetState<DeserializeBinaryBulkStateVariant>(state);
    settings.path.push_back(Substream::VariantElements);
    for (size_t i = 0; i != variants.size(); ++i)
    {
        addVariantElementToPath(settings.path, i);
-        variants[i]->deserializeBinaryBulkWithMultipleStreams(col.getVariantPtrByLocalDiscriminator(i), variant_limits[i], settings, variant_state->states[i], cache);
+        variants[i]->deserializeBinaryBulkWithMultipleStreams(col.getVariantPtrByLocalDiscriminator(i), variant_limits[i], settings, variant_state->variant_states[i], cache);
        settings.path.pop_back();
    }
    settings.path.pop_back();
@ -336,20 +475,49 @@ void SerializationVariant::deserializeBinaryBulkWithMultipleStreams(
    }
    else
    {
-        auto & offsets = col.getOffsets();
-        offsets.reserve(offsets.size() + limit);
        std::vector<size_t> variant_offsets;
        variant_offsets.reserve(variants.size());
+        size_t num_non_empty_variants = 0;
+        ColumnVariant::Discriminator last_non_empty_discr = 0;
        for (size_t i = 0; i != variants.size(); ++i)
-            variant_offsets.push_back(col.getVariantByLocalDiscriminator(i).size() - variant_limits[i]);
-
-        for (size_t i = discriminators_offset; i != discriminators_data.size(); ++i)
        {
-            ColumnVariant::Discriminator discr = discriminators_data[i];
-            if (discr == ColumnVariant::NULL_DISCRIMINATOR)
-                offsets.emplace_back();
-            else
-                offsets.push_back(variant_offsets[discr]++);
+            if (variant_limits[i])
+            {
+                ++num_non_empty_variants;
+                last_non_empty_discr = i;
+            }
+
+            variant_offsets.push_back(col.getVariantByLocalDiscriminator(i).size() - variant_limits[i]);
+        }
+
+        auto & discriminators_data = col.getLocalDiscriminators();
+        auto & offsets = col.getOffsets();
+        size_t num_new_offsets = discriminators_data.size() - offsets.size();
+        offsets.reserve(offsets.size() + num_new_offsets);
+        /// If there are only NULLs were read, fill offsets with 0.
+        if (num_non_empty_variants == 0)
+        {
+            offsets.resize_fill(discriminators_data.size(), 0);
+        }
+        /// If there is only 1 variant and no NULLs was read, fill offsets with sequential offsets of this variant.
+        else if (num_non_empty_variants == 1 && variant_limits[last_non_empty_discr] == num_new_offsets)
+        {
+            size_t first_offset = col.getVariantByLocalDiscriminator(last_non_empty_discr).size() - num_new_offsets;
+            for (size_t i = 0; i != num_new_offsets; ++i)
+                offsets.push_back(first_offset + i);
+        }
+        /// Otherwise iterate through discriminators and fill offsets accordingly.
+        else
+        {
+            size_t start = offsets.size();
+            for (size_t i = start; i != discriminators_data.size(); ++i)
+            {
+                ColumnVariant::Discriminator discr = discriminators_data[i];
+                if (discr == ColumnVariant::NULL_DISCRIMINATOR)
+                    offsets.emplace_back();
+                else
+                    offsets.push_back(variant_offsets[discr]++);
+            }
        }

        addToSubstreamsCache(cache, settings.path, col.getOffsetsPtr());
@ -357,6 +525,72 @@ void SerializationVariant::deserializeBinaryBulkWithMultipleStreams(
    settings.path.pop_back();
 }

+std::vector<size_t> SerializationVariant::deserializeCompactDiscriminators(
+    DB::ColumnPtr & discriminators_column,
+    size_t limit,
+    ReadBuffer * stream,
+    bool continuous_reading,
+    DeserializeBinaryBulkStateVariantDiscriminators & state) const
+{
+    auto & discriminators = assert_cast<ColumnVariant::ColumnDiscriminators &>(*discriminators_column->assumeMutable());
+    auto & discriminators_data = discriminators.getData();
+
+    /// Reset state if we are reading from the start of the granule and not from the previous position in the file.
+    if (!continuous_reading)
+        state.remaining_rows_in_granule = 0;
+
+    /// Calculate limits for variants during discriminators deserialization.
+    std::vector<size_t> variant_limits(variants.size(), 0);
+    while (limit)
+    {
+        /// If we read all rows from current granule, start reading the next one.
+        if (state.remaining_rows_in_granule == 0)
+        {
+            if (stream->eof())
+                return variant_limits;
+
+            readDiscriminatorsGranuleStart(state, stream);
+        }
+
+        size_t limit_in_granule = std::min(limit, state.remaining_rows_in_granule);
+        if (state.granule_format == CompactDiscriminatorsGranuleFormat::COMPACT)
+        {
+            auto & data = discriminators.getData();
+            data.resize_fill(data.size() + limit_in_granule, state.compact_discr);
+            if (state.compact_discr != ColumnVariant::NULL_DISCRIMINATOR)
+                variant_limits[state.compact_discr] += limit_in_granule;
+        }
+        else
+        {
+            SerializationNumber<ColumnVariant::Discriminator>().deserializeBinaryBulk(discriminators, *stream, limit_in_granule, 0);
+            size_t start = discriminators_data.size() - limit_in_granule;
+            for (size_t i = start; i != discriminators_data.size(); ++i)
+            {
+                ColumnVariant::Discriminator discr = discriminators_data[i];
+                if (discr != ColumnVariant::NULL_DISCRIMINATOR)
+                    ++variant_limits[discr];
+            }
+        }
+
+        state.remaining_rows_in_granule -= limit_in_granule;
+        limit -= limit_in_granule;
+    }
+
+    return variant_limits;
+}
+
+void SerializationVariant::readDiscriminatorsGranuleStart(DeserializeBinaryBulkStateVariantDiscriminators & state, DB::ReadBuffer * stream)
+{
+    UInt64 granule_size;
+    readVarUInt(granule_size, *stream);
+    state.remaining_rows_in_granule = granule_size;
+    UInt8 granule_format;
+    readBinaryLittleEndian(granule_format, *stream);
+    state.granule_format = static_cast<CompactDiscriminatorsGranuleFormat>(granule_format);
+    if (granule_format == CompactDiscriminatorsGranuleFormat::COMPACT)
+        readBinaryLittleEndian(state.compact_discr, *stream);
+}
+
 void SerializationVariant::addVariantElementToPath(DB::ISerialization::SubstreamPath & path, size_t i) const
 {
    path.push_back(Substream::VariantElement);
--- a/src/DataTypes/Serializations/SerializationVariant.h
+++ b/src/DataTypes/Serializations/SerializationVariant.h
@ -6,6 +6,13 @@
 namespace DB
 {

+
+namespace ErrorCodes
+{
+    extern const int INCORRECT_DATA;
+}
+
+
 /// Class for serializing/deserializing column with Variant type.
 /// It supports both text and binary bulk serializations/deserializations.
 ///
@ -18,6 +25,17 @@ namespace DB
 ///
 /// During binary bulk serialization it transforms local discriminators
 /// to global and serializes them into a separate stream VariantDiscriminators.
+/// There are 2 modes of serialising discriminators:
+/// Basic mode, when all discriminators are serialized as is row by row.
+/// Compact mode, when we avoid writing the same discriminators in granules when there is
+/// only one variant (or only NULLs) in the granule.
+/// In compact mode we serialize granules in the following format:
+/// <number of rows in granule><granule format><granule data>
+/// There are 2 different formats of granule - plain and compact.
+/// Plain format is used when there are different discriminators in this granule,
+/// in this format all discriminators are serialized as is row by row.
+/// Compact format is used when all discriminators are the same in this granule,
+/// in this case only this single discriminator is serialized.
 /// Each variant is serialized into a separate stream with path VariantElements/VariantElement
 /// (VariantElements stream is needed for correct sub-columns creation). We store and serialize
 /// variants in a sparse form (the size of a variant column equals to the number of its discriminator
@ -32,6 +50,25 @@ namespace DB
 class SerializationVariant : public ISerialization
 {
 public:
+    struct DiscriminatorsSerializationMode
+    {
+        enum Value
+        {
+            BASIC = 0,    /// Store the whole discriminators column.
+            COMPACT = 1,  /// Don't write discriminators in granule if all of them are the same.
+        };
+
+        static void checkMode(UInt64 mode)
+        {
+            if (mode > Value::COMPACT)
+                throw Exception(ErrorCodes::INCORRECT_DATA, "Invalid version for SerializationVariant discriminators column.");
+        }
+
+        explicit DiscriminatorsSerializationMode(UInt64 mode) : value(static_cast<Value>(mode)) { checkMode(mode); }
+
+        Value value;
+    };
+
    using VariantSerializations = std::vector<SerializationPtr>;

    explicit SerializationVariant(
@ -123,8 +160,43 @@ public:
    static std::vector<size_t> getVariantsDeserializeTextOrder(const DataTypes & variant_types);

 private:
+    friend SerializationVariantElement;
+
    void addVariantElementToPath(SubstreamPath & path, size_t i) const;

+    enum CompactDiscriminatorsGranuleFormat
+    {
+        PLAIN = 0,   /// Granule has different discriminators and they are serialized as is row by row.
+        COMPACT = 1, /// Granule has single discriminator for all rows and it is serialized as single value.
+    };
+
+    struct DeserializeBinaryBulkStateVariantDiscriminators : public ISerialization::DeserializeBinaryBulkState
+    {
+        explicit DeserializeBinaryBulkStateVariantDiscriminators(UInt64 mode_) : mode(mode_)
+        {
+        }
+
+        DiscriminatorsSerializationMode mode;
+
+        /// Deserialize state of currently read granule in compact mode.
+        CompactDiscriminatorsGranuleFormat granule_format = CompactDiscriminatorsGranuleFormat::PLAIN;
+        size_t remaining_rows_in_granule = 0;
+        ColumnVariant::Discriminator compact_discr = 0;
+    };
+
+    static DeserializeBinaryBulkStatePtr deserializeDiscriminatorsStatePrefix(
+        DeserializeBinaryBulkSettings & settings,
+        SubstreamsDeserializeStatesCache * cache);
+
+    std::vector<size_t> deserializeCompactDiscriminators(
+        ColumnPtr & discriminators_column,
+        size_t limit,
+        ReadBuffer * stream,
+        bool continuous_reading,
+        DeserializeBinaryBulkStateVariantDiscriminators & state) const;
+
+    static void readDiscriminatorsGranuleStart(DeserializeBinaryBulkStateVariantDiscriminators & state, ReadBuffer * stream);
+
    bool tryDeserializeTextEscapedImpl(IColumn & column, const String & field, const FormatSettings & settings) const;
    bool tryDeserializeTextQuotedImpl(IColumn & column, const String & field, const FormatSettings & settings) const;
    bool tryDeserializeWholeTextImpl(IColumn & column, const String & field, const FormatSettings & settings) const;
--- a/src/DataTypes/Serializations/SerializationVariantElement.cpp
+++ b/src/DataTypes/Serializations/SerializationVariantElement.cpp
@ -1,5 +1,6 @@
 #include <DataTypes/Serializations/SerializationVariantElement.h>
 #include <DataTypes/Serializations/SerializationNumber.h>
+#include <DataTypes/Serializations/SerializationVariant.h>
 #include <Columns/ColumnLowCardinality.h>
 #include <Columns/ColumnNullable.h>
 #include <IO/ReadHelpers.h>
@ -12,7 +13,7 @@ namespace ErrorCodes
    extern const int NOT_IMPLEMENTED;
 }

-struct DeserializeBinaryBulkStateVariantElement : public ISerialization::DeserializeBinaryBulkState
+struct SerializationVariantElement::DeserializeBinaryBulkStateVariantElement : public ISerialization::DeserializeBinaryBulkState
 {
    /// During deserialization discriminators and variant streams can be shared.
    /// For example we can read several variant elements together: "select v.UInt32, v.String from table",
@ -24,7 +25,7 @@ struct DeserializeBinaryBulkStateVariantElement : public ISerialization::Deseria
    /// substream cache correctly.
    ColumnPtr discriminators;
    ColumnPtr variant;
-
+    ISerialization::DeserializeBinaryBulkStatePtr discriminators_state;
    ISerialization::DeserializeBinaryBulkStatePtr variant_element_state;
 };

@ -65,7 +66,12 @@ void SerializationVariantElement::serializeBinaryBulkStateSuffix(SerializeBinary
 void SerializationVariantElement::deserializeBinaryBulkStatePrefix(
    DeserializeBinaryBulkSettings & settings, DeserializeBinaryBulkStatePtr & state, SubstreamsDeserializeStatesCache * cache) const
 {
+    DeserializeBinaryBulkStatePtr discriminators_state = SerializationVariant::deserializeDiscriminatorsStatePrefix(settings, cache);
+    if (!discriminators_state)
+        return;
+
    auto variant_element_state = std::make_shared<DeserializeBinaryBulkStateVariantElement>();
+    variant_element_state->discriminators_state = discriminators_state;

    addVariantToPath(settings.path);
    nested_serialization->deserializeBinaryBulkStatePrefix(settings, variant_element_state->variant_element_state, cache);
@ -86,35 +92,54 @@ void SerializationVariantElement::deserializeBinaryBulkWithMultipleStreams(
    DeserializeBinaryBulkStatePtr & state,
    SubstreamsCache * cache) const
 {
-    auto * variant_element_state = checkAndGetState<DeserializeBinaryBulkStateVariantElement>(state);
-
    /// First, deserialize discriminators from Variant column.
    settings.path.push_back(Substream::VariantDiscriminators);
+
+    DeserializeBinaryBulkStateVariantElement * variant_element_state = nullptr;
+    std::optional<size_t> variant_limit;
    if (auto cached_discriminators = getFromSubstreamsCache(cache, settings.path))
    {
+        variant_element_state = checkAndGetState<DeserializeBinaryBulkStateVariantElement>(state);
        variant_element_state->discriminators = cached_discriminators;
    }
-    else
+    else if (auto * discriminators_stream = settings.getter(settings.path))
    {
-        auto * discriminators_stream = settings.getter(settings.path);
-        if (!discriminators_stream)
-            return;
+        variant_element_state = checkAndGetState<DeserializeBinaryBulkStateVariantElement>(state);
+        auto * discriminators_state = checkAndGetState<SerializationVariant::DeserializeBinaryBulkStateVariantDiscriminators>(variant_element_state->discriminators_state);

        /// If we started to read a new column, reinitialize discriminators column in deserialization state.
        if (!variant_element_state->discriminators || result_column->empty())
            variant_element_state->discriminators = ColumnVariant::ColumnDiscriminators::create();

-        SerializationNumber<ColumnVariant::Discriminator>().deserializeBinaryBulk(*variant_element_state->discriminators->assumeMutable(), *discriminators_stream, limit, 0);
+        /// Deserialize discriminators according to serialization mode.
+        if (discriminators_state->mode.value == SerializationVariant::DiscriminatorsSerializationMode::BASIC)
+            SerializationNumber<ColumnVariant::Discriminator>().deserializeBinaryBulk(*variant_element_state->discriminators->assumeMutable(), *discriminators_stream, limit, 0);
+        else
+            variant_limit = deserializeCompactDiscriminators(variant_element_state->discriminators, limit, discriminators_stream, settings.continuous_reading, *variant_element_state);
+
        addToSubstreamsCache(cache, settings.path, variant_element_state->discriminators);
    }
+    else
+    {
+        settings.path.pop_back();
+        return;
+    }
+
    settings.path.pop_back();

-    /// Iterate through new discriminators to calculate the limit for our variant.
+    /// We could read less than limit discriminators, but we will need actual number of read rows later.
+    size_t num_new_discriminators = variant_element_state->discriminators->size() - result_column->size();
+
+    /// Iterate through new discriminators to calculate the limit for our variant
+    /// if we didn't do it during discriminators deserialization.
    const auto & discriminators_data = assert_cast<const ColumnVariant::ColumnDiscriminators &>(*variant_element_state->discriminators).getData();
-    size_t discriminators_offset = variant_element_state->discriminators->size() - limit;
-    size_t variant_limit = 0;
-    for (size_t i = discriminators_offset; i != discriminators_data.size(); ++i)
-        variant_limit += (discriminators_data[i] == variant_discriminator);
+    size_t discriminators_offset = variant_element_state->discriminators->size() - num_new_discriminators;
+    if (!variant_limit)
+    {
+        variant_limit = 0;
+        for (size_t i = discriminators_offset; i != discriminators_data.size(); ++i)
+            *variant_limit += (discriminators_data[i] == variant_discriminator);
+    }

    /// Now we know the limit for our variant and can deserialize it.

@ -125,19 +150,19 @@ void SerializationVariantElement::deserializeBinaryBulkWithMultipleStreams(
        auto & nullable_column = assert_cast<ColumnNullable &>(*mutable_column);
        NullMap & null_map = nullable_column.getNullMapData();
        /// If we have only our discriminator in range, fill null map with 0.
-        if (variant_limit == limit)
+        if (variant_limit == num_new_discriminators)
        {
-            null_map.resize_fill(null_map.size() + limit, 0);
+            null_map.resize_fill(null_map.size() + num_new_discriminators, 0);
        }
        /// If no our discriminator in current range, fill null map with 1.
        else if (variant_limit == 0)
        {
-            null_map.resize_fill(null_map.size() + limit, 1);
+            null_map.resize_fill(null_map.size() + num_new_discriminators, 1);
        }
        /// Otherwise we should iterate through discriminators to fill null map.
        else
        {
-            null_map.reserve(null_map.size() + limit);
+            null_map.reserve(null_map.size() + num_new_discriminators);
            for (size_t i = discriminators_offset; i != discriminators_data.size(); ++i)
                null_map.push_back(discriminators_data[i] != variant_discriminator);
        }
@ -159,12 +184,12 @@ void SerializationVariantElement::deserializeBinaryBulkWithMultipleStreams(
    /// If nothing to deserialize, just insert defaults.
    if (variant_limit == 0)
    {
-        mutable_column->insertManyDefaults(limit);
+        mutable_column->insertManyDefaults(num_new_discriminators);
        return;
    }

    addVariantToPath(settings.path);
-    nested_serialization->deserializeBinaryBulkWithMultipleStreams(variant_element_state->variant, variant_limit, settings, variant_element_state->variant_element_state, cache);
+    nested_serialization->deserializeBinaryBulkWithMultipleStreams(variant_element_state->variant, *variant_limit, settings, variant_element_state->variant_element_state, cache);
    removeVariantFromPath(settings.path);

    /// If nothing was deserialized when variant_limit > 0
@ -173,16 +198,16 @@ void SerializationVariantElement::deserializeBinaryBulkWithMultipleStreams(
    /// In this case we should just insert default values.
    if (variant_element_state->variant->empty())
    {
-        mutable_column->insertManyDefaults(limit);
+        mutable_column->insertManyDefaults(num_new_discriminators);
        return;
    }

-    size_t variant_offset = variant_element_state->variant->size() - variant_limit;
+    size_t variant_offset = variant_element_state->variant->size() - *variant_limit;

    /// If we have only our discriminator in range, insert the whole range to result column.
-    if (variant_limit == limit)
+    if (variant_limit == num_new_discriminators)
    {
-        mutable_column->insertRangeFrom(*variant_element_state->variant, variant_offset, variant_limit);
+        mutable_column->insertRangeFrom(*variant_element_state->variant, variant_offset, *variant_limit);
    }
    /// Otherwise iterate through discriminators and insert value from variant or default value depending on the discriminator.
    else
@ -197,6 +222,57 @@ void SerializationVariantElement::deserializeBinaryBulkWithMultipleStreams(
    }
 }

+size_t SerializationVariantElement::deserializeCompactDiscriminators(
+    DB::ColumnPtr & discriminators_column,
+    size_t limit,
+    DB::ReadBuffer * stream,
+    bool continuous_reading,
+    DeserializeBinaryBulkStateVariantElement & variant_element_state) const
+{
+    auto * discriminators_state = checkAndGetState<SerializationVariant::DeserializeBinaryBulkStateVariantDiscriminators>(variant_element_state.discriminators_state);
+    auto & discriminators = assert_cast<ColumnVariant::ColumnDiscriminators &>(*discriminators_column->assumeMutable());
+    auto & discriminators_data = discriminators.getData();
+
+    /// Reset state if we are reading from the start of the granule and not from the previous position in the file.
+    if (!continuous_reading)
+        discriminators_state->remaining_rows_in_granule = 0;
+
+    /// Calculate our variant limit during discriminators deserialization.
+    size_t variant_limit = 0;
+    while (limit)
+    {
+        /// If we read all rows from current granule, start reading the next one.
+        if (discriminators_state->remaining_rows_in_granule == 0)
+        {
+            if (stream->eof())
+                return variant_limit;
+
+            SerializationVariant::readDiscriminatorsGranuleStart(*discriminators_state, stream);
+        }
+
+        size_t limit_in_granule = std::min(limit, discriminators_state->remaining_rows_in_granule);
+        if (discriminators_state->granule_format == SerializationVariant::CompactDiscriminatorsGranuleFormat::COMPACT)
+        {
+            auto & data = discriminators.getData();
+            data.resize_fill(data.size() + limit_in_granule, discriminators_state->compact_discr);
+            if (discriminators_state->compact_discr == variant_discriminator)
+                variant_limit += limit_in_granule;
+        }
+        else
+        {
+            SerializationNumber<ColumnVariant::Discriminator>().deserializeBinaryBulk(discriminators, *stream, limit_in_granule, 0);
+            size_t start = discriminators_data.size() - limit_in_granule;
+            for (size_t i = start; i != discriminators_data.size(); ++i)
+                variant_limit += (discriminators_data[i] == variant_discriminator);
+        }
+
+        discriminators_state->remaining_rows_in_granule -= limit_in_granule;
+        limit -= limit_in_granule;
+    }
+
+    return variant_limit;
+}
+
 void SerializationVariantElement::addVariantToPath(DB::ISerialization::SubstreamPath & path) const
 {
    path.push_back(Substream::VariantElements);
--- a/src/DataTypes/Serializations/SerializationVariantElement.h
+++ b/src/DataTypes/Serializations/SerializationVariantElement.h
@ -80,6 +80,15 @@ public:
 private:
    friend SerializationVariant;

+    struct DeserializeBinaryBulkStateVariantElement;
+
+    size_t deserializeCompactDiscriminators(
+        ColumnPtr & discriminators_column,
+        size_t limit,
+        ReadBuffer * stream,
+        bool continuous_reading,
+        DeserializeBinaryBulkStateVariantElement & variant_element_state) const;
+
    void addVariantToPath(SubstreamPath & path) const;
    void removeVariantFromPath(SubstreamPath & path) const;
 };
--- a/src/Functions/FunctionsBinaryRepresentation.cpp
+++ b/src/Functions/FunctionsBinaryRepresentation.cpp
@ -3,14 +3,14 @@
 #include <Columns/ColumnString.h>
 #include <Columns/ColumnVector.h>
 #include <Columns/ColumnsNumber.h>
-#include <Common/BitHelpers.h>
-#include <Common/BinStringDecodeHelper.h>
 #include <DataTypes/DataTypeString.h>
 #include <Functions/FunctionFactory.h>
 #include <Functions/IFunction.h>
 #include <IO/WriteHelpers.h>
 #include <Interpreters/Context_fwd.h>
 #include <Interpreters/castColumn.h>
+#include <Common/BinStringDecodeHelper.h>
+#include <Common/BitHelpers.h>

 namespace DB
 {
@ -218,10 +218,7 @@ struct UnbinImpl
    static constexpr auto name = "unbin";
    static constexpr size_t word_size = 8;

-    static void decode(const char * pos, const char * end, char *& out)
-    {
-        binStringDecode(pos, end, out);
-    }
+    static void decode(const char * pos, const char * end, char *& out) { binStringDecode(pos, end, out, word_size); }
 };

 /// Encode number or string to string with binary or hexadecimal representation
@ -651,7 +648,15 @@ public:

            size_t size = in_offsets.size();
            out_offsets.resize(size);
-            out_vec.resize(in_vec.size() / word_size + size);
+
+            size_t max_out_len = 0;
+            for (size_t i = 0; i < in_offsets.size(); ++i)
+            {
+                const size_t len = in_offsets[i] - (i == 0 ? 0 : in_offsets[i - 1])
+                    - /* trailing zero symbol that is always added in ColumnString and that is ignored while decoding */ 1;
+                max_out_len += (len + word_size - 1) / word_size + /* trailing zero symbol that is always added by Impl::decode */ 1;
+            }
+            out_vec.resize(max_out_len);

            char * begin = reinterpret_cast<char *>(out_vec.data());
            char * pos = begin;
@ -661,6 +666,7 @@ public:
            {
                size_t new_offset = in_offsets[i];

+                /// `new_offset - 1` because in ColumnString each string is stored with trailing zero byte
                Impl::decode(reinterpret_cast<const char *>(&in_vec[prev_offset]), reinterpret_cast<const char *>(&in_vec[new_offset - 1]), pos);

                out_offsets[i] = pos - begin;
@ -668,6 +674,9 @@ public:
                prev_offset = new_offset;
            }

+            chassert(
+                static_cast<size_t>(pos - begin) <= out_vec.size(),
+                fmt::format("too small amount of memory was preallocated: needed {}, but have only {}", pos - begin, out_vec.size()));
            out_vec.resize(pos - begin);

            return col_res;
@ -680,11 +689,11 @@ public:
            ColumnString::Offsets & out_offsets = col_res->getOffsets();

            const ColumnString::Chars & in_vec = col_fix_string->getChars();
-            size_t n = col_fix_string->getN();
+            const size_t n = col_fix_string->getN();

            size_t size = col_fix_string->size();
            out_offsets.resize(size);
-            out_vec.resize(in_vec.size() / word_size + size);
+            out_vec.resize(((n + word_size - 1) / word_size + /* trailing zero symbol that is always added by Impl::decode */ 1) * size);

            char * begin = reinterpret_cast<char *>(out_vec.data());
            char * pos = begin;
@ -694,6 +703,7 @@ public:
            {
                size_t new_offset = prev_offset + n;

+                /// here we don't subtract 1 from `new_offset` because in ColumnFixedString strings are stored without trailing zero byte
                Impl::decode(reinterpret_cast<const char *>(&in_vec[prev_offset]), reinterpret_cast<const char *>(&in_vec[new_offset]), pos);

                out_offsets[i] = pos - begin;
@ -701,6 +711,9 @@ public:
                prev_offset = new_offset;
            }

+            chassert(
+                static_cast<size_t>(pos - begin) <= out_vec.size(),
+                fmt::format("too small amount of memory was preallocated: needed {}, but have only {}", pos - begin, out_vec.size()));
            out_vec.resize(pos - begin);

            return col_res;
--- a/src/IO/BufferWithOwnMemory.h
+++ b/src/IO/BufferWithOwnMemory.h
@ -4,12 +4,15 @@

 #include <Common/ProfileEvents.h>
 #include <Common/Allocator.h>
+#include <Common/GWPAsan.h>

 #include <Common/Exception.h>
 #include <Core/Defines.h>

 #include <base/arithmeticOverflow.h>

+#include "config.h"
+

 namespace ProfileEvents
 {
@ -41,10 +44,13 @@ struct Memory : boost::noncopyable, Allocator
    char * m_data = nullptr;
    size_t alignment = 0;

+    [[maybe_unused]] bool allow_gwp_asan_force_sample{false};
+
    Memory() = default;

    /// If alignment != 0, then allocate memory aligned to specified value.
-    explicit Memory(size_t size_, size_t alignment_ = 0) : alignment(alignment_)
+    explicit Memory(size_t size_, size_t alignment_ = 0, bool allow_gwp_asan_force_sample_ = false)
+        : alignment(alignment_), allow_gwp_asan_force_sample(allow_gwp_asan_force_sample_)
    {
        alloc(size_);
    }
@ -127,6 +133,11 @@ private:
        ProfileEvents::increment(ProfileEvents::IOBufferAllocs);
        ProfileEvents::increment(ProfileEvents::IOBufferAllocBytes, new_capacity);

+#if USE_GWP_ASAN
+        if (unlikely(allow_gwp_asan_force_sample && GWPAsan::shouldForceSample()))
+            gwp_asan::getThreadLocals()->NextSampleCounter = 1;
+#endif
+
        m_data = static_cast<char *>(Allocator::alloc(new_capacity, alignment));
        m_capacity = new_capacity;
        m_size = new_size;
@ -154,7 +165,7 @@ protected:
 public:
    /// If non-nullptr 'existing_memory' is passed, then buffer will not create its own memory and will use existing_memory without ownership.
    explicit BufferWithOwnMemory(size_t size = DBMS_DEFAULT_BUFFER_SIZE, char * existing_memory = nullptr, size_t alignment = 0)
-        : Base(nullptr, 0), memory(existing_memory ? 0 : size, alignment)
+        : Base(nullptr, 0), memory(existing_memory ? 0 : size, alignment, /*allow_gwp_asan_force_sample_=*/true)
    {
        Base::set(existing_memory ? existing_memory : memory.data(), size);
        Base::padded = !existing_memory;
--- a/src/Interpreters/AsynchronousInsertQueue.cpp
+++ b/src/Interpreters/AsynchronousInsertQueue.cpp
@ -301,13 +301,7 @@ void AsynchronousInsertQueue::preprocessInsertQuery(const ASTPtr & query, const
    auto & insert_query = query->as<ASTInsertQuery &>();
    insert_query.async_insert_flush = true;

-    InterpreterInsertQuery interpreter(
-        query,
-        query_context,
-        query_context->getSettingsRef().insert_allow_materialized_columns,
-        /* no_squash */ false,
-        /* no_destination */ false,
-        /* async_insert */ false);
+    InterpreterInsertQuery interpreter(query, query_context, query_context->getSettingsRef().insert_allow_materialized_columns);
    auto table = interpreter.getTable(insert_query);
    auto sample_block = InterpreterInsertQuery::getSampleBlock(insert_query, table, table->getInMemoryMetadataPtr(), query_context);

@ -787,12 +781,7 @@ try
    try
    {
        interpreter = std::make_unique<InterpreterInsertQuery>(
-            key.query,
-            insert_context,
-            key.settings.insert_allow_materialized_columns,
-            false,
-            false,
-            true);
+            key.query, insert_context, key.settings.insert_allow_materialized_columns, false, false, true);

        pipeline = interpreter->execute().pipeline;
        chassert(pipeline.pushing());
@ -1011,7 +1000,7 @@ Chunk AsynchronousInsertQueue::processEntriesWithParsing(
    }

    Chunk chunk(executor.getResultColumns(), total_rows);
-    chunk.getChunkInfos().add(std::move(chunk_info));
+    chunk.setChunkInfo(std::move(chunk_info));
    return chunk;
 }

@ -1063,7 +1052,7 @@ Chunk AsynchronousInsertQueue::processPreprocessedEntries(
    }

    Chunk chunk(std::move(result_columns), total_rows);
-    chunk.getChunkInfos().add(std::move(chunk_info));
+    chunk.setChunkInfo(std::move(chunk_info));
    return chunk;
 }

--- a/src/Interpreters/InterpreterCheckQuery.cpp
+++ b/src/Interpreters/InterpreterCheckQuery.cpp
@ -2,7 +2,6 @@
 #include <Interpreters/InterpreterFactory.h>

 #include <algorithm>
-#include <memory>

 #include <Access/Common/AccessFlags.h>

@ -23,7 +22,6 @@
 #include <Parsers/ASTCheckQuery.h>
 #include <Parsers/ASTSetQuery.h>

-#include <Processors/Chunk.h>
 #include <Processors/IAccumulatingTransform.h>
 #include <Processors/IInflatingTransform.h>
 #include <Processors/ISimpleTransform.h>
@ -93,7 +91,7 @@ Chunk getChunkFromCheckResult(const String & database, const String & table, con
    return Chunk(std::move(columns), 1);
 }

-class TableCheckTask : public ChunkInfoCloneable<TableCheckTask>
+class TableCheckTask : public ChunkInfo
 {
 public:
    TableCheckTask(StorageID table_id, const std::variant<std::monostate, ASTPtr, String> & partition_or_part, ContextPtr context)
@ -112,12 +110,6 @@ public:
        context->checkAccess(AccessType::SHOW_TABLES, table_->getStorageID());
    }

-    TableCheckTask(const TableCheckTask & other)
-        : table(other.table)
-        , check_data_tasks(other.check_data_tasks)
-        , is_finished(other.is_finished.load())
-    {}
-
    std::optional<CheckResult> checkNext() const
    {
        if (isFinished())
@ -129,8 +121,8 @@ public:
            std::this_thread::sleep_for(sleep_time);
        });

-        IStorage::DataValidationTasksPtr tmp = check_data_tasks;
-        auto result = table->checkDataNext(tmp);
+        IStorage::DataValidationTasksPtr check_data_tasks_ = check_data_tasks;
+        auto result = table->checkDataNext(check_data_tasks_);
        is_finished = !result.has_value();
        return result;
    }
@ -188,7 +180,7 @@ protected:
        /// source should return at least one row to start pipeline
        result.addColumn(ColumnUInt8::create(1, 1));
        /// actual data stored in chunk info
-        result.getChunkInfos().add(std::move(current_check_task));
+        result.setChunkInfo(std::move(current_check_task));
        return result;
    }

@ -288,7 +280,7 @@ public:
 protected:
    void transform(Chunk & chunk) override
    {
-        auto table_check_task = chunk.getChunkInfos().get<TableCheckTask>();
+        auto table_check_task = std::dynamic_pointer_cast<const TableCheckTask>(chunk.getChunkInfo());
        auto check_result = table_check_task->checkNext();
        if (!check_result)
        {
--- a/src/Interpreters/InterpreterCreateQuery.cpp
+++ b/src/Interpreters/InterpreterCreateQuery.cpp
@ -1776,13 +1776,8 @@ BlockIO InterpreterCreateQuery::fillTableIfNeeded(const ASTCreateQuery & create)
        else
            insert->select = create.select->clone();

-        return InterpreterInsertQuery(
-            insert,
-            getContext(),
-            getContext()->getSettingsRef().insert_allow_materialized_columns,
-            /* no_squash */ false,
-            /* no_destination */ false,
-            /* async_isnert */ false).execute();
+        return InterpreterInsertQuery(insert, getContext(),
+            getContext()->getSettingsRef().insert_allow_materialized_columns).execute();
    }

    return {};
--- a/src/Interpreters/InterpreterExplainQuery.cpp
+++ b/src/Interpreters/InterpreterExplainQuery.cpp
@ -534,13 +534,7 @@ QueryPipeline InterpreterExplainQuery::executeImpl()
            }
            else if (dynamic_cast<const ASTInsertQuery *>(ast.getExplainedQuery().get()))
            {
-                InterpreterInsertQuery insert(
-                    ast.getExplainedQuery(),
-                    getContext(),
-                    /* allow_materialized */ false,
-                    /* no_squash */ false,
-                    /* no_destination */ false,
-                    /* async_isnert */ false);
+                InterpreterInsertQuery insert(ast.getExplainedQuery(), getContext());
                auto io = insert.execute();
                printPipeline(io.pipeline.getProcessors(), buf);
            }
--- a/src/Interpreters/InterpreterInsertQuery.cpp
+++ b/src/Interpreters/InterpreterInsertQuery.cpp
@ -16,7 +16,6 @@
 #include <Interpreters/getTableExpressions.h>
 #include <Interpreters/processColumnTransformers.h>
 #include <Interpreters/InterpreterSelectQueryAnalyzer.h>
-#include <Interpreters/Context_fwd.h>
 #include <Parsers/ASTFunction.h>
 #include <Parsers/ASTInsertQuery.h>
 #include <Parsers/ASTSelectQuery.h>
@ -27,7 +26,6 @@
 #include <Processors/Transforms/CountingTransform.h>
 #include <Processors/Transforms/ExpressionTransform.h>
 #include <Processors/Transforms/MaterializingTransform.h>
-#include <Processors/Transforms/DeduplicationTokenTransforms.h>
 #include <Processors/Transforms/SquashingTransform.h>
 #include <Processors/Transforms/PlanSquashingTransform.h>
 #include <Processors/Transforms/getSourceFromASTInsertQuery.h>
@ -40,7 +38,6 @@
 #include <Common/ThreadStatus.h>
 #include <Common/checkStackSize.h>
 #include <Common/ProfileEvents.h>
-#include "base/defines.h"


 namespace ProfileEvents
@ -397,358 +394,28 @@ Chain InterpreterInsertQuery::buildPreSinkChain(
    return out;
 }

-std::pair<std::vector<Chain>, std::vector<Chain>> InterpreterInsertQuery::buildPreAndSinkChains(size_t presink_streams, size_t sink_streams, StoragePtr table, const StorageMetadataPtr & metadata_snapshot, const Block & query_sample_block)
-{
-    chassert(presink_streams > 0);
-    chassert(sink_streams > 0);
-
-    ThreadGroupPtr running_group;
-    if (current_thread)
-        running_group = current_thread->getThreadGroup();
-    if (!running_group)
-        running_group = std::make_shared<ThreadGroup>(getContext());
-
-    std::vector<Chain> sink_chains;
-    std::vector<Chain> presink_chains;
-
-    for (size_t i = 0; i < sink_streams; ++i)
-    {
-        auto out = buildSink(table, metadata_snapshot, /* thread_status_holder= */ nullptr,
-            running_group, /* elapsed_counter_ms= */ nullptr);
-
-        sink_chains.emplace_back(std::move(out));
-    }
-
-    for (size_t i = 0; i < presink_streams; ++i)
-    {
-        auto out = buildPreSinkChain(sink_chains[0].getInputHeader(), table, metadata_snapshot, query_sample_block);
-        presink_chains.emplace_back(std::move(out));
-    }
-
-    return {std::move(presink_chains), std::move(sink_chains)};
-}
-
-
-QueryPipeline InterpreterInsertQuery::buildInsertSelectPipeline(ASTInsertQuery & query, StoragePtr table)
-{
-    const Settings & settings = getContext()->getSettingsRef();
-
-    auto metadata_snapshot = table->getInMemoryMetadataPtr();
-    auto query_sample_block = getSampleBlock(query, table, metadata_snapshot, getContext(), no_destination, allow_materialized);
-
-    bool is_trivial_insert_select = false;
-
-    if (settings.optimize_trivial_insert_select)
-    {
-        const auto & select_query = query.select->as<ASTSelectWithUnionQuery &>();
-        const auto & selects = select_query.list_of_selects->children;
-        const auto & union_modes = select_query.list_of_modes;
-
-        /// ASTSelectWithUnionQuery is not normalized now, so it may pass some queries which can be Trivial select queries
-        const auto mode_is_all = [](const auto & mode) { return mode == SelectUnionMode::UNION_ALL; };
-
-        is_trivial_insert_select =
-            std::all_of(union_modes.begin(), union_modes.end(), std::move(mode_is_all))
-            && std::all_of(selects.begin(), selects.end(), isTrivialSelect);
-    }
-
-    ContextPtr select_context = getContext();
-
-    if (is_trivial_insert_select)
-    {
-        /** When doing trivial INSERT INTO ... SELECT ... FROM table,
-            * don't need to process SELECT with more than max_insert_threads
-            * and it's reasonable to set block size for SELECT to the desired block size for INSERT
-            * to avoid unnecessary squashing.
-            */
-
-        Settings new_settings = select_context->getSettings();
-
-        new_settings.max_threads = std::max<UInt64>(1, settings.max_insert_threads);
-
-        if (table->prefersLargeBlocks())
-        {
-            if (settings.min_insert_block_size_rows)
-                new_settings.max_block_size = settings.min_insert_block_size_rows;
-            if (settings.min_insert_block_size_bytes)
-                new_settings.preferred_block_size_bytes = settings.min_insert_block_size_bytes;
-        }
-
-        auto context_for_trivial_select = Context::createCopy(context);
-        context_for_trivial_select->setSettings(new_settings);
-        context_for_trivial_select->setInsertionTable(getContext()->getInsertionTable(), getContext()->getInsertionTableColumnNames());
-
-        select_context = context_for_trivial_select;
-    }
-
-    QueryPipelineBuilder pipeline;
-
-    {
-        auto select_query_options = SelectQueryOptions(QueryProcessingStage::Complete, 1);
-
-        if (settings.allow_experimental_analyzer)
-        {
-            InterpreterSelectQueryAnalyzer interpreter_select_analyzer(query.select, select_context, select_query_options);
-            pipeline = interpreter_select_analyzer.buildQueryPipeline();
-        }
-        else
-        {
-            InterpreterSelectWithUnionQuery interpreter_select(query.select, select_context, select_query_options);
-            pipeline = interpreter_select.buildQueryPipeline();
-        }
-    }
-
-    pipeline.dropTotalsAndExtremes();
-
-    /// Allow to insert Nullable into non-Nullable columns, NULL values will be added as defaults values.
-    if (getContext()->getSettingsRef().insert_null_as_default)
-    {
-        const auto & input_columns = pipeline.getHeader().getColumnsWithTypeAndName();
-        const auto & query_columns = query_sample_block.getColumnsWithTypeAndName();
-        const auto & output_columns = metadata_snapshot->getColumns();
-
-        if (input_columns.size() == query_columns.size())
-        {
-            for (size_t col_idx = 0; col_idx < query_columns.size(); ++col_idx)
-            {
-                /// Change query sample block columns to Nullable to allow inserting nullable columns, where NULL values will be substituted with
-                /// default column values (in AddingDefaultsTransform), so all values will be cast correctly.
-                if (isNullableOrLowCardinalityNullable(input_columns[col_idx].type)
-                    && !isNullableOrLowCardinalityNullable(query_columns[col_idx].type)
-                    && !isVariant(query_columns[col_idx].type)
-                    && !isDynamic(query_columns[col_idx].type)
-                    && output_columns.has(query_columns[col_idx].name))
-                {
-                    query_sample_block.setColumn(
-                        col_idx,
-                        ColumnWithTypeAndName(
-                            makeNullableOrLowCardinalityNullable(query_columns[col_idx].column),
-                            makeNullableOrLowCardinalityNullable(query_columns[col_idx].type),
-                            query_columns[col_idx].name));
-                }
-            }
-        }
-    }
-
-    auto actions_dag = ActionsDAG::makeConvertingActions(
-            pipeline.getHeader().getColumnsWithTypeAndName(),
-            query_sample_block.getColumnsWithTypeAndName(),
-            ActionsDAG::MatchColumnsMode::Position);
-    auto actions = std::make_shared<ExpressionActions>(actions_dag, ExpressionActionsSettings::fromContext(getContext(), CompileExpressions::yes));
-
-    pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-    {
-        return std::make_shared<ExpressionTransform>(in_header, actions);
-    });
-
-    /// We need to convert Sparse columns to full, because it's destination storage
-    /// may not support it or may have different settings for applying Sparse serialization.
-    pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-    {
-        return std::make_shared<MaterializingTransform>(in_header);
-    });
-
-    pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-    {
-        auto context_ptr = getContext();
-        auto counting = std::make_shared<CountingTransform>(in_header, nullptr, context_ptr->getQuota());
-        counting->setProcessListElement(context_ptr->getProcessListElement());
-        counting->setProgressCallback(context_ptr->getProgressCallback());
-
-        return counting;
-    });
-
-    size_t num_select_threads = pipeline.getNumThreads();
-
-    pipeline.resize(1);
-
-    if (shouldAddSquashingFroStorage(table))
-    {
-        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-        {
-            return std::make_shared<PlanSquashingTransform>(
-                in_header,
-                table->prefersLargeBlocks() ? settings.min_insert_block_size_rows : settings.max_block_size,
-                table->prefersLargeBlocks() ? settings.min_insert_block_size_bytes : 0ULL);
-        });
-    }
-
-    pipeline.addSimpleTransform([&](const Block &in_header) -> ProcessorPtr
-    {
-        return std::make_shared<DeduplicationToken::AddTokenInfoTransform>(in_header);
-    });
-
-    if (!settings.insert_deduplication_token.value.empty())
-    {
-        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-        {
-            return std::make_shared<DeduplicationToken::SetUserTokenTransform>(settings.insert_deduplication_token.value, in_header);
-        });
-
-        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-        {
-            return std::make_shared<DeduplicationToken::SetSourceBlockNumberTransform>(in_header);
-        });
-    }
-
-    /// Number of streams works like this:
-    ///  * For the SELECT, use `max_threads`, or `max_insert_threads`, or whatever
-    ///    InterpreterSelectQuery ends up with.
-    ///  * Use `max_insert_threads` streams for various insert-preparation steps, e.g.
-    ///    materializing and squashing (too slow to do in one thread). That's `presink_chains`.
-    ///  * If the table supports parallel inserts, use max_insert_threads for writing to IStorage.
-    ///    Otherwise ResizeProcessor them down to 1 stream.
-
-    size_t presink_streams_size = std::max<size_t>(settings.max_insert_threads, pipeline.getNumStreams());
-
-    size_t sink_streams_size = table->supportsParallelInsert() ? std::max<size_t>(1, settings.max_insert_threads) : 1;
-
-    if (!settings.parallel_view_processing)
-    {
-        auto table_id = table->getStorageID();
-        auto views = DatabaseCatalog::instance().getDependentViews(table_id);
-
-        if (table->isView() || !views.empty())
-            sink_streams_size = 1;
-    }
-
-    auto [presink_chains, sink_chains] = buildPreAndSinkChains(
-        presink_streams_size, sink_streams_size,
-        table, metadata_snapshot, query_sample_block);
-
-    pipeline.resize(presink_chains.size());
-
-    if (shouldAddSquashingFroStorage(table))
-    {
-        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
-        {
-            return std::make_shared<ApplySquashingTransform>(
-                in_header,
-                table->prefersLargeBlocks() ? settings.min_insert_block_size_rows : settings.max_block_size,
-                table->prefersLargeBlocks() ? settings.min_insert_block_size_bytes : 0ULL);
-        });
-    }
-
-    for (auto & chain : presink_chains)
-        pipeline.addResources(chain.detachResources());
-    pipeline.addChains(std::move(presink_chains));
-
-    pipeline.resize(sink_streams_size);
-
-    for (auto & chain : sink_chains)
-        pipeline.addResources(chain.detachResources());
-    pipeline.addChains(std::move(sink_chains));
-
-    if (!settings.parallel_view_processing)
-    {
-        /// Don't use more threads for INSERT than for SELECT to reduce memory consumption.
-        if (pipeline.getNumThreads() > num_select_threads)
-            pipeline.setMaxThreads(num_select_threads);
-    }
-    else if (pipeline.getNumThreads() < settings.max_threads)
-    {
-        /// It is possible for query to have max_threads=1, due to optimize_trivial_insert_select,
-        /// however in case of parallel_view_processing and multiple views, views can still be processed in parallel.
-        ///
-        /// Note, number of threads will be limited by buildPushingToViewsChain() to max_threads.
-        pipeline.setMaxThreads(settings.max_threads);
-    }
-
-    pipeline.setSinks([&](const Block & cur_header, QueryPipelineBuilder::StreamType) -> ProcessorPtr
-    {
-        return std::make_shared<EmptySink>(cur_header);
-    });
-
-    return QueryPipelineBuilder::getPipeline(std::move(pipeline));
-}
-
-
-QueryPipeline InterpreterInsertQuery::buildInsertPipeline(ASTInsertQuery & query, StoragePtr table)
-{
-    const Settings & settings = getContext()->getSettingsRef();
-
-    auto metadata_snapshot = table->getInMemoryMetadataPtr();
-    auto query_sample_block = getSampleBlock(query, table, metadata_snapshot, getContext(), no_destination, allow_materialized);
-
-    Chain chain;
-
-    {
-        auto [presink_chains, sink_chains] = buildPreAndSinkChains(
-            /* presink_streams */1, /* sink_streams */1,
-            table, metadata_snapshot, query_sample_block);
-
-        chain = std::move(presink_chains.front());
-        chain.appendChain(std::move(sink_chains.front()));
-    }
-
-    if (!settings.insert_deduplication_token.value.empty())
-    {
-        chain.addSource(std::make_shared<DeduplicationToken::SetSourceBlockNumberTransform>(chain.getInputHeader()));
-        chain.addSource(std::make_shared<DeduplicationToken::SetUserTokenTransform>(settings.insert_deduplication_token.value, chain.getInputHeader()));
-    }
-
-    chain.addSource(std::make_shared<DeduplicationToken::AddTokenInfoTransform>(chain.getInputHeader()));
-
-    if (shouldAddSquashingFroStorage(table))
-    {
-        bool table_prefers_large_blocks = table->prefersLargeBlocks();
-
-        auto squashing = std::make_shared<ApplySquashingTransform>(
-                    chain.getInputHeader(),
-                    table_prefers_large_blocks ? settings.min_insert_block_size_rows : settings.max_block_size,
-                    table_prefers_large_blocks ? settings.min_insert_block_size_bytes : 0ULL);
-
-        chain.addSource(std::move(squashing));
-
-        auto balancing = std::make_shared<PlanSquashingTransform>(
-                    chain.getInputHeader(),
-                    table_prefers_large_blocks ? settings.min_insert_block_size_rows : settings.max_block_size,
-                    table_prefers_large_blocks ? settings.min_insert_block_size_bytes : 0ULL);
-
-        chain.addSource(std::move(balancing));
-    }
-
-    auto context_ptr = getContext();
-    auto counting = std::make_shared<CountingTransform>(chain.getInputHeader(), nullptr, context_ptr->getQuota());
-    counting->setProcessListElement(context_ptr->getProcessListElement());
-    counting->setProgressCallback(context_ptr->getProgressCallback());
-    chain.addSource(std::move(counting));
-
-    QueryPipeline pipeline = QueryPipeline(std::move(chain));
-
-    pipeline.setNumThreads(std::min<size_t>(pipeline.getNumThreads(), settings.max_threads));
-    pipeline.setConcurrencyControl(settings.use_concurrency_control);
-
-    if (query.hasInlinedData() && !async_insert)
-    {
-        /// can execute without additional data
-        auto format = getInputFormatFromASTInsertQuery(query_ptr, true, query_sample_block, getContext(), nullptr);
-        for (auto && buffer : owned_buffers)
-            format->addBuffer(std::move(buffer));
-
-        auto pipe = getSourceFromInputFormat(query_ptr, std::move(format), getContext(), nullptr);
-        pipeline.complete(std::move(pipe));
-    }
-
-    return pipeline;
-}
-
-
 BlockIO InterpreterInsertQuery::execute()
 {
    const Settings & settings = getContext()->getSettingsRef();
    auto & query = query_ptr->as<ASTInsertQuery &>();

+    QueryPipelineBuilder pipeline;
+    std::optional<QueryPipeline> distributed_pipeline;
+    QueryPlanResourceHolder resources;

    StoragePtr table = getTable(query);
    checkStorageSupportsTransactionsIfNeeded(table, getContext());

+    StoragePtr inner_table;
+    if (const auto * mv = dynamic_cast<const StorageMaterializedView *>(table.get()))
+        inner_table = mv->getTargetTable();
+
    if (query.partition_by && !table->supportsPartitionBy())
        throw Exception(ErrorCodes::NOT_IMPLEMENTED, "PARTITION BY clause is not supported by storage");

    auto table_lock = table->lockForShare(getContext()->getInitialQueryId(), settings.lock_acquire_timeout);
-
    auto metadata_snapshot = table->getInMemoryMetadataPtr();
+
    auto query_sample_block = getSampleBlock(query, table, metadata_snapshot, getContext(), no_destination, allow_materialized);

    /// For table functions we check access while executing
@ -756,43 +423,320 @@ BlockIO InterpreterInsertQuery::execute()
    if (!query.table_function)
        getContext()->checkAccess(AccessType::INSERT, query.table_id, query_sample_block.getNames());

-    if (!allow_materialized)
+    if (query.select && settings.parallel_distributed_insert_select)
+        // Distributed INSERT SELECT
+        distributed_pipeline = table->distributedWrite(query, getContext());
+
+    std::vector<Chain> presink_chains;
+    std::vector<Chain> sink_chains;
+    if (!distributed_pipeline)
    {
-        for (const auto & column : metadata_snapshot->getColumns())
-            if (column.default_desc.kind == ColumnDefaultKind::Materialized && query_sample_block.has(column.name))
-                throw Exception(ErrorCodes::ILLEGAL_COLUMN, "Cannot insert column {}, because it is MATERIALIZED column.", column.name);
+        /// Number of streams works like this:
+        ///  * For the SELECT, use `max_threads`, or `max_insert_threads`, or whatever
+        ///    InterpreterSelectQuery ends up with.
+        ///  * Use `max_insert_threads` streams for various insert-preparation steps, e.g.
+        ///    materializing and squashing (too slow to do in one thread). That's `presink_chains`.
+        ///  * If the table supports parallel inserts, use the same streams for writing to IStorage.
+        ///    Otherwise ResizeProcessor them down to 1 stream.
+        ///  * If it's not an INSERT SELECT, forget all that and use one stream.
+        size_t pre_streams_size = 1;
+        size_t sink_streams_size = 1;
+
+        if (query.select)
+        {
+            bool is_trivial_insert_select = false;
+
+            if (settings.optimize_trivial_insert_select)
+            {
+                const auto & select_query = query.select->as<ASTSelectWithUnionQuery &>();
+                const auto & selects = select_query.list_of_selects->children;
+                const auto & union_modes = select_query.list_of_modes;
+
+                /// ASTSelectWithUnionQuery is not normalized now, so it may pass some queries which can be Trivial select queries
+                const auto mode_is_all = [](const auto & mode) { return mode == SelectUnionMode::UNION_ALL; };
+
+                is_trivial_insert_select =
+                    std::all_of(union_modes.begin(), union_modes.end(), std::move(mode_is_all))
+                    && std::all_of(selects.begin(), selects.end(), isTrivialSelect);
+            }
+
+            if (is_trivial_insert_select)
+            {
+                /** When doing trivial INSERT INTO ... SELECT ... FROM table,
+                  * don't need to process SELECT with more than max_insert_threads
+                  * and it's reasonable to set block size for SELECT to the desired block size for INSERT
+                  * to avoid unnecessary squashing.
+                  */
+
+                Settings new_settings = getContext()->getSettings();
+
+                new_settings.max_threads = std::max<UInt64>(1, settings.max_insert_threads);
+
+                if (table->prefersLargeBlocks())
+                {
+                    if (settings.min_insert_block_size_rows)
+                        new_settings.max_block_size = settings.min_insert_block_size_rows;
+                    if (settings.min_insert_block_size_bytes)
+                        new_settings.preferred_block_size_bytes = settings.min_insert_block_size_bytes;
+                }
+
+                auto new_context = Context::createCopy(context);
+                new_context->setSettings(new_settings);
+                new_context->setInsertionTable(getContext()->getInsertionTable(), getContext()->getInsertionTableColumnNames());
+
+                auto select_query_options = SelectQueryOptions(QueryProcessingStage::Complete, 1);
+
+                if (settings.allow_experimental_analyzer)
+                {
+                    InterpreterSelectQueryAnalyzer interpreter_select_analyzer(query.select, new_context, select_query_options);
+                    pipeline = interpreter_select_analyzer.buildQueryPipeline();
+                }
+                else
+                {
+                    InterpreterSelectWithUnionQuery interpreter_select(query.select, new_context, select_query_options);
+                    pipeline = interpreter_select.buildQueryPipeline();
+                }
+            }
+            else
+            {
+                /// Passing 1 as subquery_depth will disable limiting size of intermediate result.
+                auto select_query_options = SelectQueryOptions(QueryProcessingStage::Complete, 1);
+
+                if (settings.allow_experimental_analyzer)
+                {
+                    InterpreterSelectQueryAnalyzer interpreter_select_analyzer(query.select, getContext(), select_query_options);
+                    pipeline = interpreter_select_analyzer.buildQueryPipeline();
+                }
+                else
+                {
+                    InterpreterSelectWithUnionQuery interpreter_select(query.select, getContext(), select_query_options);
+                    pipeline = interpreter_select.buildQueryPipeline();
+                }
+            }
+
+            pipeline.dropTotalsAndExtremes();
+
+            if (settings.max_insert_threads > 1)
+            {
+                auto table_id = table->getStorageID();
+                auto views = DatabaseCatalog::instance().getDependentViews(table_id);
+
+                /// It breaks some views-related tests and we have dedicated `parallel_view_processing` for views, so let's just skip them.
+                /// Also it doesn't make sense to reshuffle data if storage doesn't support parallel inserts.
+                const bool resize_to_max_insert_threads = !table->isView() && views.empty() && table->supportsParallelInsert();
+                pre_streams_size = resize_to_max_insert_threads ? settings.max_insert_threads
+                                                                : std::min<size_t>(settings.max_insert_threads, pipeline.getNumStreams());
+
+                /// Deduplication when passing insert_deduplication_token breaks if using more than one thread
+                if (!settings.insert_deduplication_token.toString().empty())
+                {
+                    LOG_DEBUG(
+                        getLogger("InsertQuery"),
+                        "Insert-select query using insert_deduplication_token, setting streams to 1 to avoid deduplication issues");
+                    pre_streams_size = 1;
+                }
+
+                if (table->supportsParallelInsert())
+                    sink_streams_size = pre_streams_size;
+            }
+
+            pipeline.resize(pre_streams_size);
+
+            /// Allow to insert Nullable into non-Nullable columns, NULL values will be added as defaults values.
+            if (getContext()->getSettingsRef().insert_null_as_default)
+            {
+                const auto & input_columns = pipeline.getHeader().getColumnsWithTypeAndName();
+                const auto & query_columns = query_sample_block.getColumnsWithTypeAndName();
+                const auto & output_columns = metadata_snapshot->getColumns();
+
+                if (input_columns.size() == query_columns.size())
+                {
+                    for (size_t col_idx = 0; col_idx < query_columns.size(); ++col_idx)
+                    {
+                        /// Change query sample block columns to Nullable to allow inserting nullable columns, where NULL values will be substituted with
+                        /// default column values (in AddingDefaultsTransform), so all values will be cast correctly.
+                        if (isNullableOrLowCardinalityNullable(input_columns[col_idx].type)
+                            && !isNullableOrLowCardinalityNullable(query_columns[col_idx].type)
+                            && !isVariant(query_columns[col_idx].type)
+                            && !isDynamic(query_columns[col_idx].type)
+                            && output_columns.has(query_columns[col_idx].name))
+                            query_sample_block.setColumn(col_idx, ColumnWithTypeAndName(makeNullableOrLowCardinalityNullable(query_columns[col_idx].column), makeNullableOrLowCardinalityNullable(query_columns[col_idx].type), query_columns[col_idx].name));
+                    }
+                }
+            }
+        }
+
+        ThreadGroupPtr running_group;
+        if (current_thread)
+            running_group = current_thread->getThreadGroup();
+        if (!running_group)
+            running_group = std::make_shared<ThreadGroup>(getContext());
+        for (size_t i = 0; i < sink_streams_size; ++i)
+        {
+            auto out = buildSink(table, metadata_snapshot, /* thread_status_holder= */ nullptr,
+                running_group, /* elapsed_counter_ms= */ nullptr);
+            sink_chains.emplace_back(std::move(out));
+        }
+        for (size_t i = 0; i < pre_streams_size; ++i)
+        {
+            auto out = buildPreSinkChain(sink_chains[0].getInputHeader(), table, metadata_snapshot, query_sample_block);
+            presink_chains.emplace_back(std::move(out));
+        }
    }

    BlockIO res;

-    if (query.select)
+    /// What type of query: INSERT or INSERT SELECT or INSERT WATCH?
+    if (distributed_pipeline)
    {
-        if (settings.parallel_distributed_insert_select)
+        res.pipeline = std::move(*distributed_pipeline);
+    }
+    else if (query.select)
+    {
+        const auto & header = presink_chains.at(0).getInputHeader();
+        auto actions_dag = ActionsDAG::makeConvertingActions(
+                pipeline.getHeader().getColumnsWithTypeAndName(),
+                header.getColumnsWithTypeAndName(),
+                ActionsDAG::MatchColumnsMode::Position);
+        auto actions = std::make_shared<ExpressionActions>(actions_dag, ExpressionActionsSettings::fromContext(getContext(), CompileExpressions::yes));
+
+        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
        {
-            auto distributed = table->distributedWrite(query, getContext());
-            if (distributed)
-            {
-                res.pipeline = std::move(*distributed);
-            }
-            else
-            {
-                res.pipeline = buildInsertSelectPipeline(query, table);
-            }
-        }
-        else
+            return std::make_shared<ExpressionTransform>(in_header, actions);
+        });
+
+        /// We need to convert Sparse columns to full, because it's destination storage
+        /// may not support it or may have different settings for applying Sparse serialization.
+        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
        {
-            res.pipeline = buildInsertSelectPipeline(query, table);
+            return std::make_shared<MaterializingTransform>(in_header);
+        });
+
+        pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
+        {
+            auto context_ptr = getContext();
+            auto counting = std::make_shared<CountingTransform>(in_header, nullptr, context_ptr->getQuota());
+            counting->setProcessListElement(context_ptr->getProcessListElement());
+            counting->setProgressCallback(context_ptr->getProgressCallback());
+
+            return counting;
+        });
+
+        if (shouldAddSquashingFroStorage(table))
+        {
+            bool table_prefers_large_blocks = table->prefersLargeBlocks();
+
+            size_t threads = presink_chains.size();
+
+            pipeline.resize(1);
+
+            pipeline.addTransform(std::make_shared<PlanSquashingTransform>(
+                    header,
+                    table_prefers_large_blocks ? settings.min_insert_block_size_rows : settings.max_block_size,
+                    table_prefers_large_blocks ? settings.min_insert_block_size_bytes : 0ULL));
+
+            pipeline.resize(threads);
+
+            pipeline.addSimpleTransform([&](const Block & in_header) -> ProcessorPtr
+            {
+                return std::make_shared<ApplySquashingTransform>(
+                    in_header,
+                    table_prefers_large_blocks ? settings.min_insert_block_size_rows : settings.max_block_size,
+                    table_prefers_large_blocks ? settings.min_insert_block_size_bytes : 0ULL);
+            });
        }
+
+        size_t num_select_threads = pipeline.getNumThreads();
+
+        for (auto & chain : presink_chains)
+            resources = chain.detachResources();
+        for (auto & chain : sink_chains)
+            resources = chain.detachResources();
+
+        pipeline.addChains(std::move(presink_chains));
+        pipeline.resize(sink_chains.size());
+        pipeline.addChains(std::move(sink_chains));
+
+        if (!settings.parallel_view_processing)
+        {
+            /// Don't use more threads for INSERT than for SELECT to reduce memory consumption.
+            if (pipeline.getNumThreads() > num_select_threads)
+                pipeline.setMaxThreads(num_select_threads);
+        }
+        else if (pipeline.getNumThreads() < settings.max_threads)
+        {
+            /// It is possible for query to have max_threads=1, due to optimize_trivial_insert_select,
+            /// however in case of parallel_view_processing and multiple views, views can still be processed in parallel.
+            ///
+            /// Note, number of threads will be limited by buildPushingToViewsChain() to max_threads.
+            pipeline.setMaxThreads(settings.max_threads);
+        }
+
+        pipeline.setSinks([&](const Block & cur_header, QueryPipelineBuilder::StreamType) -> ProcessorPtr
+        {
+            return std::make_shared<EmptySink>(cur_header);
+        });
+
+        if (!allow_materialized)
+        {
+            for (const auto & column : metadata_snapshot->getColumns())
+                if (column.default_desc.kind == ColumnDefaultKind::Materialized && header.has(column.name))
+                    throw Exception(ErrorCodes::ILLEGAL_COLUMN, "Cannot insert column {}, because it is MATERIALIZED column.", column.name);
+        }
+
+        res.pipeline = QueryPipelineBuilder::getPipeline(std::move(pipeline));
    }
    else
    {
-        res.pipeline = buildInsertPipeline(query, table);
+        auto & chain = presink_chains.at(0);
+        chain.appendChain(std::move(sink_chains.at(0)));
+
+        if (shouldAddSquashingFroStorage(table))
+        {
+            bool table_prefers_large_blocks = table->prefersLargeBlocks();
+
+            auto squashing = std::make_shared<ApplySquashingTransform>(
+                    chain.getInputHeader(),
+                    table_prefers_large_blocks ? settings.min_insert_block_size_rows : settings.max_block_size,
+                    table_prefers_large_blocks ? settings.min_insert_block_size_bytes : 0ULL);
+
+            chain.addSource(std::move(squashing));
+
+            auto balancing = std::make_shared<PlanSquashingTransform>(
+                    chain.getInputHeader(),
+                    table_prefers_large_blocks ? settings.min_insert_block_size_rows : settings.max_block_size,
+                    table_prefers_large_blocks ? settings.min_insert_block_size_bytes : 0ULL);
+
+            chain.addSource(std::move(balancing));
+        }
+
+        auto context_ptr = getContext();
+        auto counting = std::make_shared<CountingTransform>(chain.getInputHeader(), nullptr, context_ptr->getQuota());
+        counting->setProcessListElement(context_ptr->getProcessListElement());
+        counting->setProgressCallback(context_ptr->getProgressCallback());
+        chain.addSource(std::move(counting));
+
+        res.pipeline = QueryPipeline(std::move(presink_chains[0]));
+        res.pipeline.setNumThreads(std::min<size_t>(res.pipeline.getNumThreads(), settings.max_threads));
+        res.pipeline.setConcurrencyControl(settings.use_concurrency_control);
+
+        if (query.hasInlinedData() && !async_insert)
+        {
+            /// can execute without additional data
+            auto format = getInputFormatFromASTInsertQuery(query_ptr, true, query_sample_block, getContext(), nullptr);
+            for (auto && buffer : owned_buffers)
+                format->addBuffer(std::move(buffer));
+
+            auto pipe = getSourceFromInputFormat(query_ptr, std::move(format), getContext(), nullptr);
+            res.pipeline.complete(std::move(pipe));
+        }
    }

-    res.pipeline.addStorageHolder(table);
+    res.pipeline.addResources(std::move(resources));

-    if (const auto * mv = dynamic_cast<const StorageMaterializedView *>(table.get()))
-        res.pipeline.addStorageHolder(mv->getTargetTable());
+    res.pipeline.addStorageHolder(table);
+    if (inner_table)
+        res.pipeline.addStorageHolder(inner_table);

    return res;
 }
@ -813,27 +757,17 @@ void InterpreterInsertQuery::extendQueryLogElemImpl(QueryLogElement & elem, Cont
    }
 }

-
 void InterpreterInsertQuery::extendQueryLogElemImpl(QueryLogElement & elem, const ASTPtr &, ContextPtr context_) const
 {
    extendQueryLogElemImpl(elem, context_);
 }

-
 void registerInterpreterInsertQuery(InterpreterFactory & factory)
 {
    auto create_fn = [] (const InterpreterFactory::Arguments & args)
    {
-        return std::make_unique<InterpreterInsertQuery>(
-            args.query,
-            args.context,
-            args.allow_materialized,
-            /* no_squash */false,
-            /* no_destination */false,
-            /* async_insert */false);
+        return std::make_unique<InterpreterInsertQuery>(args.query, args.context, args.allow_materialized);
    };
    factory.registerInterpreter("InterpreterInsertQuery", create_fn);
 }
-
-
 }
--- a/src/Interpreters/InterpreterInsertQuery.h
+++ b/src/Interpreters/InterpreterInsertQuery.h
@ -23,10 +23,10 @@ public:
    InterpreterInsertQuery(
        const ASTPtr & query_ptr_,
        ContextPtr context_,
-        bool allow_materialized_,
-        bool no_squash_,
-        bool no_destination,
-        bool async_insert_);
+        bool allow_materialized_ = false,
+        bool no_squash_ = false,
+        bool no_destination_ = false,
+        bool async_insert_ = false);

    /** Prepare a request for execution. Return block streams
      * - the stream into which you can write data to execute the query, if INSERT;
@ -73,17 +73,12 @@ private:

    ASTPtr query_ptr;
    const bool allow_materialized;
-    bool no_squash = false;
-    bool no_destination = false;
+    const bool no_squash;
+    const bool no_destination;
    const bool async_insert;

    std::vector<std::unique_ptr<ReadBuffer>> owned_buffers;

-    std::pair<std::vector<Chain>, std::vector<Chain>> buildPreAndSinkChains(size_t presink_streams, size_t sink_streams, StoragePtr table, const StorageMetadataPtr & metadata_snapshot, const Block & query_sample_block);
-
-    QueryPipeline buildInsertSelectPipeline(ASTInsertQuery & query, StoragePtr table);
-    QueryPipeline buildInsertPipeline(ASTInsertQuery & query, StoragePtr table);
-
    Chain buildSink(
        const StoragePtr & table,
        const StorageMetadataPtr & metadata_snapshot,
--- a/src/Interpreters/Squashing.cpp
+++ b/src/Interpreters/Squashing.cpp
@ -1,7 +1,6 @@
 #include <vector>
 #include <Interpreters/Squashing.h>
 #include <Common/CurrentThread.h>
-#include <base/defines.h>


 namespace DB
@ -12,33 +11,24 @@ namespace ErrorCodes
 }

 Squashing::Squashing(Block header_, size_t min_block_size_rows_, size_t min_block_size_bytes_)
-    : min_block_size_rows(min_block_size_rows_)
+    : header(header_)
+    , min_block_size_rows(min_block_size_rows_)
    , min_block_size_bytes(min_block_size_bytes_)
-    , header(header_)
 {
 }

 Chunk Squashing::flush()
 {
-    if (!accumulated)
-        return {};
-
-    auto result = convertToChunk(accumulated.extract());
-    chassert(result);
-    return result;
+    return convertToChunk(std::move(chunks_to_merge_vec));
 }

 Chunk Squashing::squash(Chunk && input_chunk)
 {
-    if (!input_chunk)
+    if (!input_chunk.hasChunkInfo())
        return Chunk();

-    auto squash_info = input_chunk.getChunkInfos().extract<ChunksToSquash>();
-
-    if (!squash_info)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "There is no ChunksToSquash in ChunkInfoPtr");
-
-    return squash(std::move(squash_info->chunks), std::move(input_chunk.getChunkInfos()));
+    const auto *info = getInfoFromChunk(input_chunk);
+    return squash(info->chunks);
 }

 Chunk Squashing::add(Chunk && input_chunk)
@ -47,37 +37,48 @@ Chunk Squashing::add(Chunk && input_chunk)
        return {};

    /// Just read block is already enough.
-    if (isEnoughSize(input_chunk))
+    if (isEnoughSize(input_chunk.getNumRows(), input_chunk.bytes()))
    {
        /// If no accumulated data, return just read block.
-        if (!accumulated)
+        if (chunks_to_merge_vec.empty())
        {
-            accumulated.add(std::move(input_chunk));
-            return convertToChunk(accumulated.extract());
+            chunks_to_merge_vec.push_back(std::move(input_chunk));
+            Chunk res_chunk = convertToChunk(std::move(chunks_to_merge_vec));
+            chunks_to_merge_vec.clear();
+            return res_chunk;
        }

        /// Return accumulated data (maybe it has small size) and place new block to accumulated data.
-        Chunk res_chunk = convertToChunk(accumulated.extract());
-        accumulated.add(std::move(input_chunk));
+        Chunk res_chunk = convertToChunk(std::move(chunks_to_merge_vec));
+        chunks_to_merge_vec.clear();
+        changeCurrentSize(input_chunk.getNumRows(), input_chunk.bytes());
+        chunks_to_merge_vec.push_back(std::move(input_chunk));
        return res_chunk;
    }

    /// Accumulated block is already enough.
-    if (isEnoughSize())
+    if (isEnoughSize(accumulated_size.rows, accumulated_size.bytes))
    {
        /// Return accumulated data and place new block to accumulated data.
-        Chunk res_chunk = convertToChunk(accumulated.extract());
-        accumulated.add(std::move(input_chunk));
+        Chunk res_chunk = convertToChunk(std::move(chunks_to_merge_vec));
+        chunks_to_merge_vec.clear();
+        changeCurrentSize(input_chunk.getNumRows(), input_chunk.bytes());
+        chunks_to_merge_vec.push_back(std::move(input_chunk));
        return res_chunk;
    }

    /// Pushing data into accumulating vector
-    accumulated.add(std::move(input_chunk));
+    expandCurrentSize(input_chunk.getNumRows(), input_chunk.bytes());
+    chunks_to_merge_vec.push_back(std::move(input_chunk));

    /// If accumulated data is big enough, we send it
-    if (isEnoughSize())
-        return convertToChunk(accumulated.extract());
-
+    if (isEnoughSize(accumulated_size.rows, accumulated_size.bytes))
+    {
+        Chunk res_chunk = convertToChunk(std::move(chunks_to_merge_vec));
+        changeCurrentSize(0, 0);
+        chunks_to_merge_vec.clear();
+        return res_chunk;
+    }
    return {};
 }

@ -89,15 +90,14 @@ Chunk Squashing::convertToChunk(std::vector<Chunk> && chunks) const
    auto info = std::make_shared<ChunksToSquash>();
    info->chunks = std::move(chunks);

-    // It is imortant that chunk is not empty, it has to have columns even if they are empty
-    auto aggr_chunk = Chunk(header.getColumns(), 0);
-    aggr_chunk.getChunkInfos().add(std::move(info));
-    chassert(aggr_chunk);
-    return aggr_chunk;
+    chunks.clear();
+
+    return Chunk(header.cloneEmptyColumns(), 0, info);
 }

-Chunk Squashing::squash(std::vector<Chunk> && input_chunks, Chunk::ChunkInfoCollection && infos)
+Chunk Squashing::squash(std::vector<Chunk> & input_chunks)
 {
+    Chunk accumulated_chunk;
    std::vector<IColumn::MutablePtr> mutable_columns = {};
    size_t rows = 0;
    for (const Chunk & chunk : input_chunks)
@ -119,17 +119,35 @@ Chunk Squashing::squash(std::vector<Chunk> && input_chunks, Chunk::ChunkInfoColl
        for (size_t j = 0, size = mutable_columns.size(); j < size; ++j)
        {
            const auto source_column = columns[j];
+
            mutable_columns[j]->insertRangeFrom(*source_column, 0, source_column->size());
        }
    }
+    accumulated_chunk.setColumns(std::move(mutable_columns), rows);
+    return accumulated_chunk;
+}

-    Chunk result;
-    result.setColumns(std::move(mutable_columns), rows);
-    result.setChunkInfos(infos);
-    result.getChunkInfos().append(std::move(input_chunks.back().getChunkInfos()));
+const ChunksToSquash* Squashing::getInfoFromChunk(const Chunk & chunk)
+{
+    const auto& info = chunk.getChunkInfo();
+    const auto * agg_info = typeid_cast<const ChunksToSquash *>(info.get());

-    chassert(result);
-    return result;
+    if (!agg_info)
+        throw Exception(ErrorCodes::LOGICAL_ERROR, "There is no ChunksToSquash in ChunkInfoPtr");
+
+    return agg_info;
+}
+
+void Squashing::expandCurrentSize(size_t rows, size_t bytes)
+{
+    accumulated_size.rows += rows;
+    accumulated_size.bytes += bytes;
+}
+
+void Squashing::changeCurrentSize(size_t rows, size_t bytes)
+{
+    accumulated_size.rows = rows;
+    accumulated_size.bytes = bytes;
 }

 bool Squashing::isEnoughSize(size_t rows, size_t bytes) const
@ -138,28 +156,4 @@ bool Squashing::isEnoughSize(size_t rows, size_t bytes) const
        || (min_block_size_rows && rows >= min_block_size_rows)
        || (min_block_size_bytes && bytes >= min_block_size_bytes);
 }
-
-bool Squashing::isEnoughSize() const
-{
-    return isEnoughSize(accumulated.getRows(), accumulated.getBytes());
-};
-
-bool Squashing::isEnoughSize(const Chunk & chunk) const
-{
-    return isEnoughSize(chunk.getNumRows(), chunk.bytes());
-}
-
-void Squashing::CurrentSize::add(Chunk && chunk)
-{
-    rows += chunk.getNumRows();
-    bytes += chunk.bytes();
-    chunks.push_back(std::move(chunk));
-}
-
-std::vector<Chunk> Squashing::CurrentSize::extract()
-{
-    auto result = std::move(chunks);
-    *this = {};
-    return result;
-}
 }
--- a/src/Interpreters/Squashing.h
+++ b/src/Interpreters/Squashing.h
@ -8,18 +8,9 @@
 namespace DB
 {

-class ChunksToSquash : public ChunkInfoCloneable<ChunksToSquash>
+struct ChunksToSquash : public ChunkInfo
 {
-public:
-    ChunksToSquash() = default;
-    ChunksToSquash(const ChunksToSquash & other)
-    {
-        chunks.reserve(other.chunks.size());
-        for (const auto & chunk: other.chunks)
-           chunks.push_back(chunk.clone());
-    }
-
-    std::vector<Chunk> chunks = {};
+    mutable std::vector<Chunk> chunks = {};
 };

 /** Merging consecutive passed blocks to specified minimum size.
@ -45,35 +36,32 @@ public:
    static Chunk squash(Chunk && input_chunk);
    Chunk flush();

-    void setHeader(Block header_) { header = std::move(header_); }
-    const Block & getHeader() const { return header; }
-
-private:
-    class CurrentSize
+    bool isDataLeft()
+    {
+        return !chunks_to_merge_vec.empty();
+    }
+
+    Block header;
+private:
+    struct CurrentSize
    {
-        std::vector<Chunk> chunks = {};
        size_t rows = 0;
        size_t bytes = 0;
-
-    public:
-        explicit operator bool () const { return !chunks.empty(); }
-        size_t getRows() const { return rows; }
-        size_t getBytes() const { return bytes; }
-        void add(Chunk && chunk);
-        std::vector<Chunk> extract();
    };

-    const size_t min_block_size_rows;
-    const size_t min_block_size_bytes;
-    Block header;
+    std::vector<Chunk> chunks_to_merge_vec = {};
+    size_t min_block_size_rows;
+    size_t min_block_size_bytes;

-    CurrentSize accumulated;
+    CurrentSize accumulated_size;

-    static Chunk squash(std::vector<Chunk> && input_chunks, Chunk::ChunkInfoCollection && infos);
+    static const ChunksToSquash * getInfoFromChunk(const Chunk & chunk);

-    bool isEnoughSize() const;
+    static Chunk squash(std::vector<Chunk> & input_chunks);
+
+    void expandCurrentSize(size_t rows, size_t bytes);
+    void changeCurrentSize(size_t rows, size_t bytes);
    bool isEnoughSize(size_t rows, size_t bytes) const;
-    bool isEnoughSize(const Chunk & chunk) const;

    Chunk convertToChunk(std::vector<Chunk> && chunks) const;
 };
--- a/src/Interpreters/SystemLog.cpp
+++ b/src/Interpreters/SystemLog.cpp
@ -538,13 +538,7 @@ void SystemLog<LogElement>::flushImpl(const std::vector<LogElement> & to_flush,
        insert_context->makeQueryContext();
        addSettingsForQuery(insert_context, IAST::QueryKind::Insert);

-        InterpreterInsertQuery interpreter(
-            query_ptr,
-            insert_context,
-            /* allow_materialized */ false,
-            /* no_squash */ false,
-            /* no_destination */ false,
-            /* async_isnert */ false);
+        InterpreterInsertQuery interpreter(query_ptr, insert_context);
        BlockIO io = interpreter.execute();

        PushingPipelineExecutor executor(io.pipeline);
--- a/src/Interpreters/TreeRewriter.cpp
+++ b/src/Interpreters/TreeRewriter.cpp
@ -1188,7 +1188,7 @@ bool TreeRewriterResult::collectUsedColumns(const ASTPtr & query, bool is_select
        }
    }

-    /// Check for dynamic subcolumns in unknown required columns.
+    /// Check for dynamic subcolums in unknown required columns.
    if (!unknown_required_source_columns.empty())
    {
        for (const NameAndTypePair & pair : source_columns_ordinary)
--- a/src/Parsers/ExpressionElementParsers.cpp
+++ b/src/Parsers/ExpressionElementParsers.cpp
@ -1129,11 +1129,11 @@ inline static bool makeHexOrBinStringLiteral(IParser::Pos & pos, ASTPtr & node,

    if (hex)
    {
-        hexStringDecode(str_begin, str_end, res_pos);
+        hexStringDecode(str_begin, str_end, res_pos, word_size);
    }
    else
    {
-        binStringDecode(str_begin, str_end, res_pos);
+        binStringDecode(str_begin, str_end, res_pos, word_size);
    }

    return makeStringLiteral(pos, node, String(reinterpret_cast<char *>(res.data()), (res_pos - res_begin - 1)));
--- a/src/Processors/Chunk.cpp
+++ b/src/Processors/Chunk.cpp
@ -19,6 +19,14 @@ Chunk::Chunk(DB::Columns columns_, UInt64 num_rows_) : columns(std::move(columns
    checkNumRowsIsConsistent();
 }

+Chunk::Chunk(Columns columns_, UInt64 num_rows_, ChunkInfoPtr chunk_info_)
+    : columns(std::move(columns_))
+    , num_rows(num_rows_)
+    , chunk_info(std::move(chunk_info_))
+{
+    checkNumRowsIsConsistent();
+}
+
 static Columns unmuteColumns(MutableColumns && mutable_columns)
 {
    Columns columns;
@ -35,11 +43,17 @@ Chunk::Chunk(MutableColumns columns_, UInt64 num_rows_)
    checkNumRowsIsConsistent();
 }

+Chunk::Chunk(MutableColumns columns_, UInt64 num_rows_, ChunkInfoPtr chunk_info_)
+    : columns(unmuteColumns(std::move(columns_)))
+    , num_rows(num_rows_)
+    , chunk_info(std::move(chunk_info_))
+{
+    checkNumRowsIsConsistent();
+}
+
 Chunk Chunk::clone() const
 {
-    auto tmp = Chunk(getColumns(), getNumRows());
-    tmp.setChunkInfos(chunk_infos.clone());
-    return tmp;
+    return Chunk(getColumns(), getNumRows(), chunk_info);
 }

 void Chunk::setColumns(Columns columns_, UInt64 num_rows_)
--- a/src/Processors/Chunk.h
+++ b/src/Processors/Chunk.h
@ -1,9 +1,7 @@
 #pragma once

-#include <Common/CollectionOfDerived.h>
 #include <Columns/IColumn.h>
-
-#include <memory>
+#include <unordered_map>

 namespace DB
 {
@ -11,29 +9,11 @@ namespace DB
 class ChunkInfo
 {
 public:
-    using Ptr = std::shared_ptr<ChunkInfo>;
-
-    ChunkInfo() = default;
-    ChunkInfo(const ChunkInfo&) = default;
-    ChunkInfo(ChunkInfo&&) = default;
-
-    virtual Ptr clone() const = 0;
    virtual ~ChunkInfo() = default;
+    ChunkInfo() = default;
 };

-
-template<class Derived>
-class ChunkInfoCloneable : public ChunkInfo
-{
-public:
-    ChunkInfoCloneable() = default;
-    ChunkInfoCloneable(const ChunkInfoCloneable & other) = default;
-
-    Ptr clone() const override
-    {
-        return std::static_pointer_cast<ChunkInfo>(std::make_shared<Derived>(*static_cast<const Derived*>(this)));
-    }
-};
+using ChunkInfoPtr = std::shared_ptr<const ChunkInfo>;

 /**
 * Chunk is a list of columns with the same length.
@ -52,26 +32,26 @@ public:
 class Chunk
 {
 public:
-    using ChunkInfoCollection = CollectionOfDerivedItems<ChunkInfo>;
-
    Chunk() = default;
    Chunk(const Chunk & other) = delete;
    Chunk(Chunk && other) noexcept
        : columns(std::move(other.columns))
        , num_rows(other.num_rows)
-        , chunk_infos(std::move(other.chunk_infos))
+        , chunk_info(std::move(other.chunk_info))
    {
        other.num_rows = 0;
    }

    Chunk(Columns columns_, UInt64 num_rows_);
+    Chunk(Columns columns_, UInt64 num_rows_, ChunkInfoPtr chunk_info_);
    Chunk(MutableColumns columns_, UInt64 num_rows_);
+    Chunk(MutableColumns columns_, UInt64 num_rows_, ChunkInfoPtr chunk_info_);

    Chunk & operator=(const Chunk & other) = delete;
    Chunk & operator=(Chunk && other) noexcept
    {
        columns = std::move(other.columns);
-        chunk_infos = std::move(other.chunk_infos);
+        chunk_info = std::move(other.chunk_info);
        num_rows = other.num_rows;
        other.num_rows = 0;
        return *this;
@ -82,15 +62,15 @@ public:
    void swap(Chunk & other) noexcept
    {
        columns.swap(other.columns);
+        chunk_info.swap(other.chunk_info);
        std::swap(num_rows, other.num_rows);
-        chunk_infos.swap(other.chunk_infos);
    }

    void clear()
    {
        num_rows = 0;
        columns.clear();
-        chunk_infos.clear();
+        chunk_info.reset();
    }

    const Columns & getColumns() const { return columns; }
@ -101,9 +81,9 @@ public:
    /** Get empty columns with the same types as in block. */
    MutableColumns cloneEmptyColumns() const;

-    ChunkInfoCollection & getChunkInfos() { return chunk_infos; }
-    const ChunkInfoCollection & getChunkInfos() const { return chunk_infos; }
-    void setChunkInfos(ChunkInfoCollection chunk_infos_) { chunk_infos = std::move(chunk_infos_); }
+    const ChunkInfoPtr & getChunkInfo() const { return chunk_info; }
+    bool hasChunkInfo() const { return chunk_info != nullptr; }
+    void setChunkInfo(ChunkInfoPtr chunk_info_) { chunk_info = std::move(chunk_info_); }

    UInt64 getNumRows() const { return num_rows; }
    UInt64 getNumColumns() const { return columns.size(); }
@ -127,7 +107,7 @@ public:
 private:
    Columns columns;
    UInt64 num_rows = 0;
-    ChunkInfoCollection chunk_infos;
+    ChunkInfoPtr chunk_info;

    void checkNumRowsIsConsistent();
 };
@ -137,15 +117,11 @@ using Chunks = std::vector<Chunk>;
 /// AsyncInsert needs two kinds of information:
 /// - offsets of different sub-chunks
 /// - tokens of different sub-chunks, which are assigned by setting `insert_deduplication_token`.
-class AsyncInsertInfo : public ChunkInfoCloneable<AsyncInsertInfo>
+class AsyncInsertInfo : public ChunkInfo
 {
 public:
    AsyncInsertInfo() = default;
-    AsyncInsertInfo(const AsyncInsertInfo & other) = default;
-    AsyncInsertInfo(const std::vector<size_t> & offsets_, const std::vector<String> & tokens_)
-        : offsets(offsets_)
-        , tokens(tokens_)
-    {}
+    explicit AsyncInsertInfo(const std::vector<size_t> & offsets_, const std::vector<String> & tokens_) : offsets(offsets_), tokens(tokens_) {}

    std::vector<size_t> offsets;
    std::vector<String> tokens;
@ -154,11 +130,9 @@ public:
 using AsyncInsertInfoPtr = std::shared_ptr<AsyncInsertInfo>;

 /// Extension to support delayed defaults. AddingDefaultsProcessor uses it to replace missing values with column defaults.
-class ChunkMissingValues : public ChunkInfoCloneable<ChunkMissingValues>
+class ChunkMissingValues : public ChunkInfo
 {
 public:
-    ChunkMissingValues(const ChunkMissingValues & other) = default;
-
    using RowsBitMask = std::vector<bool>; /// a bit per row for a column

    const RowsBitMask & getDefaultsBitmask(size_t column_idx) const;
--- a/src/Processors/Executors/PullingAsyncPipelineExecutor.cpp
+++ b/src/Processors/Executors/PullingAsyncPipelineExecutor.cpp
@ -147,10 +147,13 @@ bool PullingAsyncPipelineExecutor::pull(Block & block, uint64_t milliseconds)

    block = lazy_format->getPort(IOutputFormat::PortKind::Main).getHeader().cloneWithColumns(chunk.detachColumns());

-    if (auto agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>())
+    if (auto chunk_info = chunk.getChunkInfo())
    {
-         block.info.bucket_num = agg_info->bucket_num;
-         block.info.is_overflows = agg_info->is_overflows;
+        if (const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(chunk_info.get()))
+        {
+            block.info.bucket_num = agg_info->bucket_num;
+            block.info.is_overflows = agg_info->is_overflows;
+        }
    }

    return true;
--- a/src/Processors/Executors/PullingPipelineExecutor.cpp
+++ b/src/Processors/Executors/PullingPipelineExecutor.cpp
@ -73,10 +73,13 @@ bool PullingPipelineExecutor::pull(Block & block)
    }

    block = pulling_format->getPort(IOutputFormat::PortKind::Main).getHeader().cloneWithColumns(chunk.detachColumns());
-    if (auto agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>())
+    if (auto chunk_info = chunk.getChunkInfo())
    {
-        block.info.bucket_num = agg_info->bucket_num;
-        block.info.is_overflows = agg_info->is_overflows;
+        if (const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(chunk_info.get()))
+        {
+            block.info.bucket_num = agg_info->bucket_num;
+            block.info.is_overflows = agg_info->is_overflows;
+        }
    }

    return true;
--- a/src/Processors/Formats/Impl/ParquetBlockOutputFormat.cpp
+++ b/src/Processors/Formats/Impl/ParquetBlockOutputFormat.cpp
@ -179,9 +179,7 @@ void ParquetBlockOutputFormat::consume(Chunk chunk)
                columns[i]->insertRangeFrom(*concatenated.getColumns()[i], offset, count);

            Chunks piece;
-            piece.emplace_back(std::move(columns), count);
-            piece.back().setChunkInfos(concatenated.getChunkInfos());
-
+            piece.emplace_back(std::move(columns), count, concatenated.getChunkInfo());
            writeRowGroup(std::move(piece));
        }
    }
--- a/src/Processors/IAccumulatingTransform.cpp
+++ b/src/Processors/IAccumulatingTransform.cpp
@ -8,9 +8,8 @@ namespace ErrorCodes
 }

 IAccumulatingTransform::IAccumulatingTransform(Block input_header, Block output_header)
-    : IProcessor({std::move(input_header)}, {std::move(output_header)})
-    , input(inputs.front())
-    , output(outputs.front())
+    : IProcessor({std::move(input_header)}, {std::move(output_header)}),
+    input(inputs.front()), output(outputs.front())
 {
 }

--- a/src/Processors/Merges/Algorithms/FinishAggregatingInOrderAlgorithm.cpp
+++ b/src/Processors/Merges/Algorithms/FinishAggregatingInOrderAlgorithm.cpp
@ -53,11 +53,13 @@ void FinishAggregatingInOrderAlgorithm::consume(Input & input, size_t source_num
    if (!input.chunk.hasRows())
        return;

-    if (input.chunk.getChunkInfos().empty())
+    const auto & info = input.chunk.getChunkInfo();
+    if (!info)
        throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk in FinishAggregatingInOrderAlgorithm");

    Int64 allocated_bytes = 0;
-    if (auto arenas_info = input.chunk.getChunkInfos().get<ChunkInfoWithAllocatedBytes>())
+    /// Will be set by AggregatingInOrderTransform during local aggregation; will be nullptr during merging on initiator.
+    if (const auto * arenas_info = typeid_cast<const ChunkInfoWithAllocatedBytes *>(info.get()))
        allocated_bytes = arenas_info->allocated_bytes;

    states[source_num] = State{input.chunk, description, allocated_bytes};
@ -134,7 +136,7 @@ Chunk FinishAggregatingInOrderAlgorithm::prepareToMerge()
    info->chunk_num = chunk_num++;

    Chunk chunk;
-    chunk.getChunkInfos().add(std::move(info));
+    chunk.setChunkInfo(std::move(info));
    return chunk;
 }

@ -161,7 +163,7 @@ void FinishAggregatingInOrderAlgorithm::addToAggregation()
            chunks.emplace_back(std::move(new_columns), current_rows);
        }

-        chunks.back().getChunkInfos().add(std::make_shared<AggregatedChunkInfo>());
+        chunks.back().setChunkInfo(std::make_shared<AggregatedChunkInfo>());
        states[i].current_row = states[i].to_row;

        /// We assume that sizes in bytes of rows are almost the same.
--- a/src/Processors/Merges/Algorithms/MergeTreePartLevelInfo.h
+++ b/src/Processors/Merges/Algorithms/MergeTreePartLevelInfo.h
@ -6,22 +6,18 @@ namespace DB
 {

 /// To carry part level if chunk is produced by a merge tree source
-class MergeTreePartLevelInfo : public ChunkInfoCloneable<MergeTreePartLevelInfo>
+class MergeTreePartLevelInfo : public ChunkInfo
 {
 public:
    MergeTreePartLevelInfo() = delete;
-    explicit MergeTreePartLevelInfo(ssize_t part_level)
-        : origin_merge_tree_part_level(part_level)
-    { }
-    MergeTreePartLevelInfo(const MergeTreePartLevelInfo & other) = default;
-
+    explicit MergeTreePartLevelInfo(ssize_t part_level) : origin_merge_tree_part_level(part_level) { }
    size_t origin_merge_tree_part_level = 0;
 };

 inline size_t getPartLevelFromChunk(const Chunk & chunk)
 {
-    const auto part_level_info = chunk.getChunkInfos().get<MergeTreePartLevelInfo>();
-    if (part_level_info)
+    const auto & info = chunk.getChunkInfo();
+    if (const auto * part_level_info = typeid_cast<const MergeTreePartLevelInfo *>(info.get()))
        return part_level_info->origin_merge_tree_part_level;
    return 0;
 }
--- a/src/Processors/Merges/Algorithms/ReplacingSortedAlgorithm.cpp
+++ b/src/Processors/Merges/Algorithms/ReplacingSortedAlgorithm.cpp
@ -17,7 +17,7 @@ namespace ErrorCodes

 static IMergingAlgorithm::Status emitChunk(detail::SharedChunkPtr & chunk, bool finished = false)
 {
-    chunk->getChunkInfos().add(std::make_shared<ChunkSelectFinalIndices>(std::move(chunk->replace_final_selection)));
+    chunk->setChunkInfo(std::make_shared<ChunkSelectFinalIndices>(std::move(chunk->replace_final_selection)));
    return IMergingAlgorithm::Status(std::move(*chunk), finished);
 }

--- a/src/Processors/Merges/Algorithms/ReplacingSortedAlgorithm.h
+++ b/src/Processors/Merges/Algorithms/ReplacingSortedAlgorithm.h
@ -3,7 +3,6 @@
 #include <Processors/Merges/Algorithms/MergedData.h>
 #include <Processors/Transforms/ColumnGathererTransform.h>
 #include <Processors/Merges/Algorithms/RowRef.h>
-#include <Processors/Chunk.h>

 namespace Poco
 {
@ -15,13 +14,11 @@ namespace DB

 /** Use in skipping final to keep list of indices of selected row after merging final
  */
-struct ChunkSelectFinalIndices : public ChunkInfoCloneable<ChunkSelectFinalIndices>
+struct ChunkSelectFinalIndices : public ChunkInfo
 {
-    explicit ChunkSelectFinalIndices(MutableColumnPtr select_final_indices_);
-    ChunkSelectFinalIndices(const ChunkSelectFinalIndices & other) = default;
-
    const ColumnPtr column_holder;
    const ColumnUInt64 * select_final_indices = nullptr;
+    explicit ChunkSelectFinalIndices(MutableColumnPtr select_final_indices_);
 };

 /** Merges several sorted inputs into one.
--- a/src/Processors/Merges/IMergingTransform.cpp
+++ b/src/Processors/Merges/IMergingTransform.cpp
@ -157,7 +157,7 @@ IProcessor::Status IMergingTransformBase::prepare()
    bool is_port_full = !output.canPush();

    /// Push if has data.
-    if ((state.output_chunk || !state.output_chunk.getChunkInfos().empty()) && !is_port_full)
+    if ((state.output_chunk || state.output_chunk.hasChunkInfo()) && !is_port_full)
        output.push(std::move(state.output_chunk));

    if (!is_initialized)
--- a/src/Processors/Merges/IMergingTransform.h
+++ b/src/Processors/Merges/IMergingTransform.h
@ -129,7 +129,7 @@ public:

        IMergingAlgorithm::Status status = algorithm.merge();

-        if ((status.chunk && status.chunk.hasRows()) || !status.chunk.getChunkInfos().empty())
+        if ((status.chunk && status.chunk.hasRows()) || status.chunk.hasChunkInfo())
        {
            // std::cerr << "Got chunk with " << status.chunk.getNumRows() << " rows" << std::endl;
            state.output_chunk = std::move(status.chunk);
--- a/src/Processors/Sinks/RemoteSink.h
+++ b/src/Processors/Sinks/RemoteSink.h
@ -20,7 +20,7 @@ public:
    }

    String getName() const override { return "RemoteSink"; }
-    void consume (Chunk & chunk) override { write(RemoteInserter::getHeader().cloneWithColumns(chunk.getColumns())); }
+    void consume (Chunk chunk) override { write(RemoteInserter::getHeader().cloneWithColumns(chunk.detachColumns())); }
    void onFinish() override { RemoteInserter::onFinish(); }
 };

--- a/src/Processors/Sinks/SinkToStorage.cpp
+++ b/src/Processors/Sinks/SinkToStorage.cpp
@ -15,8 +15,9 @@ void SinkToStorage::onConsume(Chunk chunk)
      */
    Nested::validateArraySizes(getHeader().cloneWithColumns(chunk.getColumns()));

-    consume(chunk);
-    cur_chunk = std::move(chunk);
+    consume(chunk.clone());
+    if (!lastBlockIsDuplicate())
+        cur_chunk = std::move(chunk);
 }

 SinkToStorage::GenerateResult SinkToStorage::onGenerate()
--- a/src/Processors/Sinks/SinkToStorage.h
+++ b/src/Processors/Sinks/SinkToStorage.h
@ -18,7 +18,8 @@ public:
    void addTableLock(const TableLockHolder & lock) { table_locks.push_back(lock); }

 protected:
-    virtual void consume(Chunk & chunk) = 0;
+    virtual void consume(Chunk chunk) = 0;
+    virtual bool lastBlockIsDuplicate() const { return false; }

 private:
    std::vector<TableLockHolder> table_locks;
@ -37,7 +38,7 @@ class NullSinkToStorage : public SinkToStorage
 public:
    using SinkToStorage::SinkToStorage;
    std::string getName() const override { return "NullSinkToStorage"; }
-    void consume(Chunk &) override {}
+    void consume(Chunk) override {}
 };

 using SinkPtr = std::shared_ptr<SinkToStorage>;
--- a/src/Processors/Sources/BlocksSource.h
+++ b/src/Processors/Sources/BlocksSource.h
@ -43,10 +43,7 @@ protected:
        info->bucket_num = res.info.bucket_num;
        info->is_overflows = res.info.is_overflows;

-        auto chunk = Chunk(res.getColumns(), res.rows());
-        chunk.getChunkInfos().add(std::move(info));
-
-        return chunk;
+        return Chunk(res.getColumns(), res.rows(), std::move(info));
    }

 private:
--- a/src/Processors/Sources/RemoteSource.cpp
+++ b/src/Processors/Sources/RemoteSource.cpp
@ -176,7 +176,7 @@ std::optional<Chunk> RemoteSource::tryGenerate()
        auto info = std::make_shared<AggregatedChunkInfo>();
        info->bucket_num = block.info.bucket_num;
        info->is_overflows = block.info.is_overflows;
-        chunk.getChunkInfos().add(std::move(info));
+        chunk.setChunkInfo(std::move(info));
    }

    return chunk;
--- a/src/Processors/Sources/SourceFromSingleChunk.cpp
+++ b/src/Processors/Sources/SourceFromSingleChunk.cpp
@ -5,9 +5,7 @@
 namespace DB
 {

-SourceFromSingleChunk::SourceFromSingleChunk(Block header, Chunk chunk_) : ISource(std::move(header)), chunk(std::move(chunk_))
-{
-}
+SourceFromSingleChunk::SourceFromSingleChunk(Block header, Chunk chunk_) : ISource(std::move(header)), chunk(std::move(chunk_)) {}

 SourceFromSingleChunk::SourceFromSingleChunk(Block data) : ISource(data.cloneEmpty()), chunk(data.getColumns(), data.rows())
 {
@ -22,7 +20,7 @@ SourceFromSingleChunk::SourceFromSingleChunk(Block data) : ISource(data.cloneEmp
        auto info = std::make_shared<AggregatedChunkInfo>();
        info->bucket_num = data.info.bucket_num;
        info->is_overflows = data.info.is_overflows;
-        chunk.getChunkInfos().add(std::move(info));
+        chunk.setChunkInfo(std::move(info));
    }
 }

--- a/src/Processors/Transforms/AggregatingInOrderTransform.cpp
+++ b/src/Processors/Transforms/AggregatingInOrderTransform.cpp
@ -332,7 +332,7 @@ void AggregatingInOrderTransform::generate()
    variants.aggregates_pool = variants.aggregates_pools.at(0).get();

    /// Pass info about used memory by aggregate functions further.
-    to_push_chunk.getChunkInfos().add(std::make_shared<ChunkInfoWithAllocatedBytes>(cur_block_bytes));
+    to_push_chunk.setChunkInfo(std::make_shared<ChunkInfoWithAllocatedBytes>(cur_block_bytes));

    cur_block_bytes = 0;
    cur_block_size = 0;
@ -351,12 +351,11 @@ FinalizeAggregatedTransform::FinalizeAggregatedTransform(Block header, Aggregati
 void FinalizeAggregatedTransform::transform(Chunk & chunk)
 {
    if (params->final)
-    {
        finalizeChunk(chunk, aggregates_mask);
-    }
-    else if (!chunk.getChunkInfos().get<AggregatedChunkInfo>())
+    else if (!chunk.getChunkInfo())
    {
-        chunk.getChunkInfos().add(std::make_shared<AggregatedChunkInfo>());
+        auto info = std::make_shared<AggregatedChunkInfo>();
+        chunk.setChunkInfo(std::move(info));
    }
 }

--- a/src/Processors/Transforms/AggregatingInOrderTransform.h
+++ b/src/Processors/Transforms/AggregatingInOrderTransform.h
@ -5,7 +5,6 @@
 #include <Processors/ISimpleTransform.h>
 #include <Processors/Transforms/AggregatingTransform.h>
 #include <Processors/Transforms/finalizeChunk.h>
-#include <Processors/Chunk.h>

 namespace DB
 {
@ -13,12 +12,10 @@ namespace DB
 struct InputOrderInfo;
 using InputOrderInfoPtr = std::shared_ptr<const InputOrderInfo>;

-struct ChunkInfoWithAllocatedBytes : public ChunkInfoCloneable<ChunkInfoWithAllocatedBytes>
+struct ChunkInfoWithAllocatedBytes : public ChunkInfo
 {
-    ChunkInfoWithAllocatedBytes(const ChunkInfoWithAllocatedBytes & other) = default;
    explicit ChunkInfoWithAllocatedBytes(Int64 allocated_bytes_)
        : allocated_bytes(allocated_bytes_) {}
-
    Int64 allocated_bytes;
 };

--- a/src/Processors/Transforms/AggregatingTransform.cpp
+++ b/src/Processors/Transforms/AggregatingTransform.cpp
@ -35,7 +35,7 @@ Chunk convertToChunk(const Block & block)

    UInt64 num_rows = block.rows();
    Chunk chunk(block.getColumns(), num_rows);
-    chunk.getChunkInfos().add(std::move(info));
+    chunk.setChunkInfo(std::move(info));

    return chunk;
 }
@ -44,11 +44,15 @@ namespace
 {
    const AggregatedChunkInfo * getInfoFromChunk(const Chunk & chunk)
    {
-        auto agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>();
+        const auto & info = chunk.getChunkInfo();
+        if (!info)
+            throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk.");
+
+        const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get());
        if (!agg_info)
            throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk should have AggregatedChunkInfo.");

-        return agg_info.get();
+        return agg_info;
    }

    /// Reads chunks from file in native format. Provide chunks with aggregation info.
@ -206,7 +210,11 @@ private:

    void process(Chunk && chunk)
    {
-        auto chunks_to_merge = chunk.getChunkInfos().get<ChunksToMerge>();
+        if (!chunk.hasChunkInfo())
+            throw Exception(ErrorCodes::LOGICAL_ERROR, "Expected chunk with chunk info in {}", getName());
+
+        const auto & info = chunk.getChunkInfo();
+        const auto * chunks_to_merge = typeid_cast<const ChunksToMerge *>(info.get());
        if (!chunks_to_merge)
            throw Exception(ErrorCodes::LOGICAL_ERROR, "Expected chunk with ChunksToMerge info in {}", getName());

--- a/src/Processors/Transforms/AggregatingTransform.h
+++ b/src/Processors/Transforms/AggregatingTransform.h
@ -2,7 +2,6 @@
 #include <Compression/CompressedReadBuffer.h>
 #include <IO/ReadBufferFromFile.h>
 #include <Interpreters/Aggregator.h>
-#include <Processors/Chunk.h>
 #include <Processors/IAccumulatingTransform.h>
 #include <Common/Stopwatch.h>
 #include <Common/setThreadName.h>
@ -20,7 +19,7 @@ namespace CurrentMetrics
 namespace DB
 {

-class AggregatedChunkInfo : public ChunkInfoCloneable<AggregatedChunkInfo>
+class AggregatedChunkInfo : public ChunkInfo
 {
 public:
    bool is_overflows = false;
--- a/src/Processors/Transforms/ApplySquashingTransform.h
+++ b/src/Processors/Transforms/ApplySquashingTransform.h
@ -27,12 +27,18 @@ public:
        }

        ExceptionKeepingTransform::work();
+        if (finish_chunk)
+        {
+            data.chunk = std::move(finish_chunk);
+            ready_output = true;
+        }
    }

 protected:
    void onConsume(Chunk chunk) override
    {
-        cur_chunk = Squashing::squash(std::move(chunk));
+        if (auto res_chunk = DB::Squashing::squash(std::move(chunk)))
+            cur_chunk.setColumns(res_chunk.getColumns(), res_chunk.getNumRows());
    }

    GenerateResult onGenerate() override
@ -42,10 +48,16 @@ protected:
        res.is_done = true;
        return res;
    }
+    void onFinish() override
+    {
+        auto chunk = DB::Squashing::squash({});
+        finish_chunk.setColumns(chunk.getColumns(), chunk.getNumRows());
+    }

 private:
    Squashing squashing;
    Chunk cur_chunk;
+    Chunk finish_chunk;
 };

 }
--- a/src/Processors/Transforms/CountingTransform.cpp
+++ b/src/Processors/Transforms/CountingTransform.cpp
@ -1,7 +1,6 @@
-#include <Processors/Transforms/CountingTransform.h>

-#include <IO/Progress.h>
 #include <Interpreters/ProcessList.h>
+#include <Processors/Transforms/CountingTransform.h>
 #include <Common/ProfileEvents.h>
 #include <Common/ThreadStatus.h>

--- a/src/Processors/Transforms/DeduplicationTokenTransforms.cpp
+++ b/src/Processors/Transforms/DeduplicationTokenTransforms.cpp
@ -1,236 +0,0 @@
-#include <Processors/Transforms/DeduplicationTokenTransforms.h>
-
-#include <IO/WriteHelpers.h>
-
-#include <Common/logger_useful.h>
-#include <Common/Exception.h>
-#include <Common/SipHash.h>
-
-
-#include <fmt/core.h>
-
-
-namespace DB
-{
-
-namespace ErrorCodes
-{
-    extern const int LOGICAL_ERROR;
-}
-
-void RestoreChunkInfosTransform::transform(Chunk & chunk)
-{
-    chunk.getChunkInfos().append(chunk_infos.clone());
-}
-
-namespace DeduplicationToken
-{
-
-String TokenInfo::getToken() const
-{
-    if (!isDefined())
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is not defined, stage {}, token {}", stage, debugToken());
-
-    return getTokenImpl();
-}
-
-String TokenInfo::getTokenImpl() const
-{
-    String result;
-    result.reserve(getTotalSize());
-
-    for (const auto & part : parts)
-    {
-        if (!result.empty())
-            result.append(":");
-        result.append(part);
-    }
-
-    return result;
-}
-
-String TokenInfo::debugToken() const
-{
-    return getTokenImpl();
-}
-
-void TokenInfo::addChunkHash(String part)
-{
-    if (stage == UNDEFINED && empty())
-        stage = DEFINE_SOURCE_WITH_HASHES;
-
-    if (stage != DEFINE_SOURCE_WITH_HASHES)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is in wrong stage {}, token {}", stage, debugToken());
-
-    addTokenPart(std::move(part));
-}
-
-void TokenInfo::finishChunkHashes()
-{
-    if (stage == UNDEFINED && empty())
-        stage = DEFINE_SOURCE_WITH_HASHES;
-
-    if (stage != DEFINE_SOURCE_WITH_HASHES)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is in wrong stage {}, token {}", stage, debugToken());
-
-    stage = DEFINED;
-}
-
-void TokenInfo::setUserToken(const String & token)
-{
-    if (stage == UNDEFINED && empty())
-        stage = DEFINE_SOURCE_USER_TOKEN;
-
-    if (stage != DEFINE_SOURCE_USER_TOKEN)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is in wrong stage {}, token {}", stage, debugToken());
-
-    addTokenPart(fmt::format("user-token-{}", token));
-}
-
-void TokenInfo::setSourceWithUserToken(size_t block_number)
-{
-    if (stage != DEFINE_SOURCE_USER_TOKEN)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is in wrong stage {}, token {}", stage, debugToken());
-
-    addTokenPart(fmt::format("source-number-{}", block_number));
-
-    stage = DEFINED;
-}
-
-void TokenInfo::setViewID(const String & id)
-{
-    if (stage == DEFINED)
-        stage = DEFINE_VIEW;
-
-    if (stage != DEFINE_VIEW)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is in wrong stage {}, token {}", stage, debugToken());
-
-    addTokenPart(fmt::format("view-id-{}", id));
-}
-
-void TokenInfo::setViewBlockNumber(size_t block_number)
-{
-    if (stage != DEFINE_VIEW)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "token is in wrong stage {}, token {}", stage, debugToken());
-
-    addTokenPart(fmt::format("view-block-{}", block_number));
-
-    stage = DEFINED;
-}
-
-void TokenInfo::reset()
-{
-    stage = UNDEFINED;
-    parts.clear();
-}
-
-void TokenInfo::addTokenPart(String part)
-{
-    parts.push_back(std::move(part));
-}
-
-size_t TokenInfo::getTotalSize() const
-{
-    if (parts.empty())
-        return 0;
-
-    size_t size = 0;
-    for (const auto & part : parts)
-        size += part.size();
-
-    // we reserve more size here to be able to add delimenter between parts.
-    return size + parts.size() - 1;
-}
-
-#ifdef ABORT_ON_LOGICAL_ERROR
-void CheckTokenTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-
-    if (!token_info)
-        throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk has to have DedupTokenInfo as ChunkInfo, {}", debug);
-
-    LOG_DEBUG(log, "debug: {}, token: {}", debug, token_info->debugToken());
-}
-#endif
-
-String DefineSourceWithChunkHashTransform::getChunkHash(const Chunk & chunk)
-{
-    SipHash hash;
-    for (const auto & colunm : chunk.getColumns())
-        colunm->updateHashFast(hash);
-
-    const auto hash_value = hash.get128();
-    return toString(hash_value.items[0]) + "_" + toString(hash_value.items[1]);
-}
-
-
-void DefineSourceWithChunkHashTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-
-    if (!token_info)
-        throw Exception(
-            ErrorCodes::LOGICAL_ERROR,
-            "TokenInfo is expected for consumed chunk in DefineSourceWithChunkHashesTransform");
-
-    if (token_info->isDefined())
-        return;
-
-    token_info->addChunkHash(getChunkHash(chunk));
-    token_info->finishChunkHashes();
-}
-
-void SetUserTokenTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-    if (!token_info)
-        throw Exception(
-            ErrorCodes::LOGICAL_ERROR,
-            "TokenInfo is expected for consumed chunk in SetUserTokenTransform");
-    token_info->setUserToken(user_token);
-}
-
-void SetSourceBlockNumberTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-    if (!token_info)
-        throw Exception(
-            ErrorCodes::LOGICAL_ERROR,
-            "TokenInfo is expected for consumed chunk in SetSourceBlockNumberTransform");
-    token_info->setSourceWithUserToken(block_number++);
-}
-
-void SetViewIDTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-    if (!token_info)
-        throw Exception(
-            ErrorCodes::LOGICAL_ERROR,
-            "TokenInfo is expected for consumed chunk in SetViewIDTransform");
-    token_info->setViewID(view_id);
-}
-
-void SetViewBlockNumberTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-    if (!token_info)
-        throw Exception(
-            ErrorCodes::LOGICAL_ERROR,
-            "TokenInfo is expected for consumed chunk in SetViewBlockNumberTransform");
-    token_info->setViewBlockNumber(block_number++);
-}
-
-void ResetTokenTransform::transform(Chunk & chunk)
-{
-    auto token_info = chunk.getChunkInfos().get<TokenInfo>();
-    if (!token_info)
-        throw Exception(
-            ErrorCodes::LOGICAL_ERROR,
-            "TokenInfo is expected for consumed chunk in ResetTokenTransform");
-
-    token_info->reset();
-}
-
-}
-}
--- a/src/Processors/Transforms/DeduplicationTokenTransforms.h
+++ b/src/Processors/Transforms/DeduplicationTokenTransforms.h
@ -1,237 +0,0 @@
-#pragma once
-
-#include <Processors/Chunk.h>
-#include <Processors/ISimpleTransform.h>
-
-#include <base/defines.h>
-#include "Common/Logger.h"
-
-
-namespace DB
-{
-    class RestoreChunkInfosTransform : public ISimpleTransform
-    {
-    public:
-        RestoreChunkInfosTransform(Chunk::ChunkInfoCollection chunk_infos_, const Block & header_)
-                : ISimpleTransform(header_, header_, true)
-                , chunk_infos(std::move(chunk_infos_))
-        {}
-
-        String getName() const override { return "RestoreChunkInfosTransform"; }
-
-        void transform(Chunk & chunk) override;
-
-    private:
-        Chunk::ChunkInfoCollection chunk_infos;
-    };
-
-
-namespace DeduplicationToken
-{
-    class TokenInfo : public ChunkInfoCloneable<TokenInfo>
-    {
-    public:
-        TokenInfo() = default;
-        TokenInfo(const TokenInfo & other) = default;
-
-        String getToken() const;
-        String debugToken() const;
-
-        bool empty() const { return parts.empty(); }
-
-        bool isDefined() const { return stage == DEFINED; }
-
-        void addChunkHash(String part);
-        void finishChunkHashes();
-
-        void setUserToken(const String & token);
-        void setSourceWithUserToken(size_t block_number);
-
-        void setViewID(const String & id);
-        void setViewBlockNumber(size_t block_number);
-
-        void reset();
-
-    private:
-        String getTokenImpl() const;
-
-        void addTokenPart(String part);
-        size_t getTotalSize() const;
-
-        /* Token has to be prepared in a particular order.
-        * BuildingStage ensures that token is expanded according the following order.
-        * Firstly token is expanded with information about the source.
-        * It could be done with two ways: add several hash sums from the source chunks or provide user defined deduplication token and its sequentional block number.
-        *
-        * transition // method
-        * UNDEFINED -> DEFINE_SOURCE_WITH_HASHES // addChunkHash
-        * DEFINE_SOURCE_WITH_HASHES -> DEFINE_SOURCE_WITH_HASHES // addChunkHash
-        * DEFINE_SOURCE_WITH_HASHES -> DEFINED // defineSourceWithChankHashes
-        *
-        * transition // method
-        * UNDEFINED -> DEFINE_SOURCE_USER_TOKEN // setUserToken
-        * DEFINE_SOURCE_USER_TOKEN -> DEFINED // defineSourceWithUserToken
-        *
-        * After token is defined, it could be extended with view id and view block number. Actually it has to be expanded with view details if there is one or several views.
-        *
-        * transition // method
-        * DEFINED -> DEFINE_VIEW // setViewID
-        * DEFINE_VIEW -> DEFINED // defineViewID
-        */
-
-        enum BuildingStage
-        {
-            UNDEFINED,
-            DEFINE_SOURCE_WITH_HASHES,
-            DEFINE_SOURCE_USER_TOKEN,
-            DEFINE_VIEW,
-            DEFINED,
-        };
-
-        BuildingStage stage = UNDEFINED;
-        std::vector<String> parts;
-    };
-
-
-#ifdef ABORT_ON_LOGICAL_ERROR
-    /// use that class only with debug builds in CI for introspection
-    class CheckTokenTransform : public ISimpleTransform
-    {
-    public:
-        CheckTokenTransform(String debug_, const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-            , debug(std::move(debug_))
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::CheckTokenTransform"; }
-
-        void transform(Chunk & chunk) override;
-
-    private:
-        String debug;
-        LoggerPtr log = getLogger("CheckInsertDeduplicationTokenTransform");
-    };
-#endif
-
-
-    class AddTokenInfoTransform : public ISimpleTransform
-    {
-    public:
-        explicit AddTokenInfoTransform(const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::AddTokenInfoTransform"; }
-
-        void transform(Chunk & chunk) override
-        {
-            chunk.getChunkInfos().add(std::make_shared<TokenInfo>());
-        }
-    };
-
-
-    class DefineSourceWithChunkHashTransform : public ISimpleTransform
-    {
-    public:
-        explicit DefineSourceWithChunkHashTransform(const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::DefineSourceWithChunkHashesTransform"; }
-
-        // Usually MergeTreeSink/ReplicatedMergeTreeSink calls addChunkHash for the deduplication token with hashes from the parts.
-        // But if there is some table with different engine, we still need to define the source of the data in deduplication token
-        // We use that transform to define the source as a hash of entire block in deduplication token
-        void transform(Chunk & chunk) override;
-
-        static String getChunkHash(const Chunk & chunk);
-    };
-
-    class ResetTokenTransform : public ISimpleTransform
-    {
-    public:
-        explicit ResetTokenTransform(const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::ResetTokenTransform"; }
-
-        void transform(Chunk & chunk) override;
-    };
-
-
-    class SetUserTokenTransform : public ISimpleTransform
-    {
-    public:
-        SetUserTokenTransform(String user_token_, const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-            , user_token(std::move(user_token_))
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::SetUserTokenTransform"; }
-
-        void transform(Chunk & chunk) override;
-
-    private:
-        String user_token;
-    };
-
-
-    class SetSourceBlockNumberTransform : public ISimpleTransform
-    {
-    public:
-        explicit SetSourceBlockNumberTransform(const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::SetSourceBlockNumberTransform"; }
-
-        void transform(Chunk & chunk) override;
-
-    private:
-        size_t block_number = 0;
-    };
-
-
-    class SetViewIDTransform : public ISimpleTransform
-    {
-    public:
-        SetViewIDTransform(String view_id_, const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-            , view_id(std::move(view_id_))
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::SetViewIDTransform"; }
-
-        void transform(Chunk & chunk) override;
-
-    private:
-        String view_id;
-    };
-
-
-    class SetViewBlockNumberTransform : public ISimpleTransform
-    {
-    public:
-        explicit SetViewBlockNumberTransform(const Block & header_)
-            : ISimpleTransform(header_, header_, true)
-        {
-        }
-
-        String getName() const override { return "DeduplicationToken::SetViewBlockNumberTransform"; }
-
-        void transform(Chunk & chunk) override;
-
-    private:
-        size_t block_number = 0;
-    };
-
-}
-}
--- a/src/Processors/Transforms/ExpressionTransform.cpp
+++ b/src/Processors/Transforms/ExpressionTransform.cpp
@ -1,7 +1,5 @@
 #include <Processors/Transforms/ExpressionTransform.h>
 #include <Interpreters/ExpressionActions.h>
-
-
 namespace DB
 {

--- a/src/Processors/Transforms/JoiningTransform.cpp
+++ b/src/Processors/Transforms/JoiningTransform.cpp
@ -365,9 +365,10 @@ IProcessor::Status DelayedJoinedBlocksWorkerTransform::prepare()
            return Status::Finished;
        }

-        task = data.chunk.getChunkInfos().get<DelayedBlocksTask>();
-        if (!task)
+        if (!data.chunk.hasChunkInfo())
            throw Exception(ErrorCodes::LOGICAL_ERROR, "DelayedJoinedBlocksWorkerTransform must have chunk info");
+
+        task = std::dynamic_pointer_cast<const DelayedBlocksTask>(data.chunk.getChunkInfo());
    }
    else
    {
@ -478,7 +479,7 @@ IProcessor::Status DelayedJoinedBlocksTransform::prepare()
            if (output.isFinished())
                continue;
            Chunk chunk;
-            chunk.getChunkInfos().add(std::make_shared<DelayedBlocksTask>());
+            chunk.setChunkInfo(std::make_shared<DelayedBlocksTask>());
            output.push(std::move(chunk));
            output.finish();
        }
@ -495,7 +496,7 @@ IProcessor::Status DelayedJoinedBlocksTransform::prepare()
        {
            Chunk chunk;
            auto task = std::make_shared<DelayedBlocksTask>(delayed_blocks, left_delayed_stream_finished_counter);
-            chunk.getChunkInfos().add(std::move(task));
+            chunk.setChunkInfo(task);
            output.push(std::move(chunk));
        }
        delayed_blocks = nullptr;
--- a/src/Processors/Transforms/JoiningTransform.h
+++ b/src/Processors/Transforms/JoiningTransform.h
@ -1,7 +1,6 @@
 #pragma once
 #include <Processors/IProcessor.h>
-#include <Processors/Chunk.h>
-#include <memory>
+

 namespace DB
 {
@ -112,12 +111,11 @@ private:
 };


-class DelayedBlocksTask : public ChunkInfoCloneable<DelayedBlocksTask>
+class DelayedBlocksTask : public ChunkInfo
 {
 public:

    DelayedBlocksTask() = default;
-    DelayedBlocksTask(const DelayedBlocksTask & other) = default;
    explicit DelayedBlocksTask(IBlocksStreamPtr delayed_blocks_, JoiningTransform::FinishCounterPtr left_delayed_stream_finish_counter_)
        : delayed_blocks(std::move(delayed_blocks_))
        , left_delayed_stream_finish_counter(left_delayed_stream_finish_counter_)
--- a/src/Processors/Transforms/MaterializingTransform.cpp
+++ b/src/Processors/Transforms/MaterializingTransform.cpp
@ -1,7 +1,6 @@
 #include <Processors/Transforms/MaterializingTransform.h>
 #include <Columns/ColumnSparse.h>

-
 namespace DB
 {

--- a/src/Processors/Transforms/MemoryBoundMerging.h
+++ b/src/Processors/Transforms/MemoryBoundMerging.h
@ -150,7 +150,11 @@ private:
        if (!chunk.hasRows())
            return;

-        const auto & agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>();
+        const auto & info = chunk.getChunkInfo();
+        if (!info)
+            throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk in SortingAggregatedForMemoryBoundMergingTransform.");
+
+        const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get());
        if (!agg_info)
            throw Exception(
                ErrorCodes::LOGICAL_ERROR, "Chunk should have AggregatedChunkInfo in SortingAggregatedForMemoryBoundMergingTransform.");
--- a/src/Processors/Transforms/MergingAggregatedMemoryEfficientTransform.cpp
+++ b/src/Processors/Transforms/MergingAggregatedMemoryEfficientTransform.cpp
@ -30,10 +30,10 @@ void GroupingAggregatedTransform::pushData(Chunks chunks, Int32 bucket, bool is_
    auto info = std::make_shared<ChunksToMerge>();
    info->bucket_num = bucket;
    info->is_overflows = is_overflows;
-    info->chunks = std::make_shared<Chunks>(std::move(chunks));
+    info->chunks = std::make_unique<Chunks>(std::move(chunks));

    Chunk chunk;
-    chunk.getChunkInfos().add(std::move(info));
+    chunk.setChunkInfo(std::move(info));
    output.push(std::move(chunk));
 }

@ -255,10 +255,11 @@ void GroupingAggregatedTransform::addChunk(Chunk chunk, size_t input)
    if (!chunk.hasRows())
        return;

-    if (chunk.getChunkInfos().empty())
+    const auto & info = chunk.getChunkInfo();
+    if (!info)
        throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk in GroupingAggregatedTransform.");

-    if (auto agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>())
+    if (const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get()))
    {
        Int32 bucket = agg_info->bucket_num;
        bool is_overflows = agg_info->is_overflows;
@ -274,7 +275,7 @@ void GroupingAggregatedTransform::addChunk(Chunk chunk, size_t input)
            last_bucket_number[input] = bucket;
        }
    }
-    else if (chunk.getChunkInfos().get<ChunkInfoWithAllocatedBytes>())
+    else if (typeid_cast<const ChunkInfoWithAllocatedBytes *>(info.get()))
    {
        single_level_chunks.emplace_back(std::move(chunk));
    }
@ -303,11 +304,7 @@ void GroupingAggregatedTransform::work()
            Int32 bucket = cur_block.info.bucket_num;
            auto chunk_info = std::make_shared<AggregatedChunkInfo>();
            chunk_info->bucket_num = bucket;
-
-            auto chunk = Chunk(cur_block.getColumns(), cur_block.rows());
-            chunk.getChunkInfos().add(std::move(chunk_info));
-
-            chunks_map[bucket].emplace_back(std::move(chunk));
+            chunks_map[bucket].emplace_back(Chunk(cur_block.getColumns(), cur_block.rows(), std::move(chunk_info)));
        }
    }
 }
@ -322,7 +319,9 @@ MergingAggregatedBucketTransform::MergingAggregatedBucketTransform(

 void MergingAggregatedBucketTransform::transform(Chunk & chunk)
 {
-    auto chunks_to_merge = chunk.getChunkInfos().get<ChunksToMerge>();
+    const auto & info = chunk.getChunkInfo();
+    const auto * chunks_to_merge = typeid_cast<const ChunksToMerge *>(info.get());
+
    if (!chunks_to_merge)
        throw Exception(ErrorCodes::LOGICAL_ERROR, "MergingAggregatedSimpleTransform chunk must have ChunkInfo with type ChunksToMerge.");

@ -331,10 +330,11 @@ void MergingAggregatedBucketTransform::transform(Chunk & chunk)
    BlocksList blocks_list;
    for (auto & cur_chunk : *chunks_to_merge->chunks)
    {
-        if (cur_chunk.getChunkInfos().empty())
+        const auto & cur_info = cur_chunk.getChunkInfo();
+        if (!cur_info)
            throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk in MergingAggregatedBucketTransform.");

-        if (auto agg_info = cur_chunk.getChunkInfos().get<AggregatedChunkInfo>())
+        if (const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(cur_info.get()))
        {
            Block block = header.cloneWithColumns(cur_chunk.detachColumns());
            block.info.is_overflows = agg_info->is_overflows;
@ -342,7 +342,7 @@ void MergingAggregatedBucketTransform::transform(Chunk & chunk)

            blocks_list.emplace_back(std::move(block));
        }
-        else if (cur_chunk.getChunkInfos().get<ChunkInfoWithAllocatedBytes>())
+        else if (typeid_cast<const ChunkInfoWithAllocatedBytes *>(cur_info.get()))
        {
            Block block = header.cloneWithColumns(cur_chunk.detachColumns());
            block.info.is_overflows = false;
@ -361,7 +361,7 @@ void MergingAggregatedBucketTransform::transform(Chunk & chunk)
    res_info->is_overflows = chunks_to_merge->is_overflows;
    res_info->bucket_num = chunks_to_merge->bucket_num;
    res_info->chunk_num = chunks_to_merge->chunk_num;
-    chunk.getChunkInfos().add(std::move(res_info));
+    chunk.setChunkInfo(std::move(res_info));

    auto block = params->aggregator.mergeBlocks(blocks_list, params->final, is_cancelled);

@ -405,7 +405,11 @@ bool SortingAggregatedTransform::tryPushChunk()

 void SortingAggregatedTransform::addChunk(Chunk chunk, size_t from_input)
 {
-    auto agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>();
+    const auto & info = chunk.getChunkInfo();
+    if (!info)
+        throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk in SortingAggregatedTransform.");
+
+    const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get());
    if (!agg_info)
        throw Exception(ErrorCodes::LOGICAL_ERROR,
            "Chunk should have AggregatedChunkInfo in SortingAggregatedTransform.");
--- a/src/Processors/Transforms/MergingAggregatedMemoryEfficientTransform.h
+++ b/src/Processors/Transforms/MergingAggregatedMemoryEfficientTransform.h
@ -3,7 +3,6 @@
 #include <Core/SortDescription.h>
 #include <Common/HashTable/HashSet.h>
 #include <Interpreters/Aggregator.h>
-#include <Processors/Chunk.h>
 #include <Processors/IProcessor.h>
 #include <Processors/ISimpleTransform.h>
 #include <Processors/ResizeProcessor.h>
@ -143,9 +142,9 @@ private:
    void addChunk(Chunk chunk, size_t from_input);
 };

-struct ChunksToMerge : public ChunkInfoCloneable<ChunksToMerge>
+struct ChunksToMerge : public ChunkInfo
 {
-    std::shared_ptr<Chunks> chunks;
+    std::unique_ptr<Chunks> chunks;
    Int32 bucket_num = -1;
    bool is_overflows = false;
    UInt64 chunk_num = 0; // chunk number in order of generation, used during memory bound merging to restore chunks order
--- a/src/Processors/Transforms/MergingAggregatedTransform.cpp
+++ b/src/Processors/Transforms/MergingAggregatedTransform.cpp
@ -32,10 +32,11 @@ void MergingAggregatedTransform::consume(Chunk chunk)
    total_input_rows += input_rows;
    ++total_input_blocks;

-    if (chunk.getChunkInfos().empty())
+    const auto & info = chunk.getChunkInfo();
+    if (!info)
        throw Exception(ErrorCodes::LOGICAL_ERROR, "Chunk info was not set for chunk in MergingAggregatedTransform.");

-    if (auto agg_info = chunk.getChunkInfos().get<AggregatedChunkInfo>())
+    if (const auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get()))
    {
        /** If the remote servers used a two-level aggregation method,
          * then blocks will contain information about the number of the bucket.
@ -48,7 +49,7 @@ void MergingAggregatedTransform::consume(Chunk chunk)

        bucket_to_blocks[agg_info->bucket_num].emplace_back(std::move(block));
    }
-    else if (chunk.getChunkInfos().get<ChunkInfoWithAllocatedBytes>())
+    else if (typeid_cast<const ChunkInfoWithAllocatedBytes *>(info.get()))
    {
        auto block = getInputPort().getHeader().cloneWithColumns(chunk.getColumns());
        block.info.is_overflows = false;
@ -88,8 +89,7 @@ Chunk MergingAggregatedTransform::generate()

    UInt64 num_rows = block.rows();
    Chunk chunk(block.getColumns(), num_rows);
-
-    chunk.getChunkInfos().add(std::move(info));
+    chunk.setChunkInfo(std::move(info));

    return chunk;
 }
--- a/src/Processors/Transforms/PlanSquashingTransform.cpp
+++ b/src/Processors/Transforms/PlanSquashingTransform.cpp
@ -10,20 +10,20 @@ namespace ErrorCodes
 }

 PlanSquashingTransform::PlanSquashingTransform(
-    Block header_, size_t min_block_size_rows, size_t min_block_size_bytes)
-    : IInflatingTransform(header_, header_)
-    , squashing(header_, min_block_size_rows, min_block_size_bytes)
+    const Block & header, size_t min_block_size_rows, size_t min_block_size_bytes)
+    : IInflatingTransform(header, header), squashing(header, min_block_size_rows, min_block_size_bytes)
 {
 }

 void PlanSquashingTransform::consume(Chunk chunk)
 {
-    squashed_chunk = squashing.add(std::move(chunk));
+    if (Chunk current_chunk = squashing.add(std::move(chunk)); current_chunk.hasChunkInfo())
+        squashed_chunk.swap(current_chunk);
 }

 Chunk PlanSquashingTransform::generate()
 {
-    if (!squashed_chunk)
+    if (!squashed_chunk.hasChunkInfo())
        throw Exception(ErrorCodes::LOGICAL_ERROR, "Can't generate chunk in SimpleSquashingChunksTransform");

    Chunk result_chunk;
@ -33,11 +33,12 @@ Chunk PlanSquashingTransform::generate()

 bool PlanSquashingTransform::canGenerate()
 {
-    return bool(squashed_chunk);
+    return squashed_chunk.hasChunkInfo();
 }

 Chunk PlanSquashingTransform::getRemaining()
 {
-    return squashing.flush();
+    Chunk current_chunk = squashing.flush();
+    return current_chunk;
 }
 }
--- a/src/Processors/Transforms/PlanSquashingTransform.h
+++ b/src/Processors/Transforms/PlanSquashingTransform.h
@ -10,7 +10,7 @@ class PlanSquashingTransform : public IInflatingTransform
 {
 public:
    PlanSquashingTransform(
-        Block header_, size_t min_block_size_rows, size_t min_block_size_bytes);
+        const Block & header, size_t min_block_size_rows, size_t min_block_size_bytes);

    String getName() const override { return "PlanSquashingTransform"; }

@ -23,6 +23,7 @@ protected:
 private:
    Squashing squashing;
    Chunk squashed_chunk;
+    Chunk finish_chunk;
 };
 }

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