TiFlash

TiFlash is a columnar storage component of TiDB and TiDB Cloud, the fully-managed service of TiDB. It mainly plays the role of Analytical Processing (AP) in the Hybrid Transactional/Analytical Processing (HTAP) architecture of TiDB.

TiFlash stores data in columnar format and synchronizes data updates in real-time from TiKV by Raft logs with sub-second latency. Reads in TiFlash are guaranteed transactionally consistent with Snapshot Isolation level. TiFlash utilizes Massively Parallel Processing (MPP) computing architecture to accelerate the analytical workloads.

TiFlash repository is based on ClickHouse. We appreciate the excellent work of the ClickHouse team.

Quick Start

Start with TiDB Cloud

Quickly explore TiFlash with a free trial of TiDB Cloud.

See TiDB Cloud Quick Start Guide.

Start with TiDB

See Quick Start with HTAP and Use TiFlash.

Build TiFlash

TiFlash supports building on the following hardware architectures:

• x86-64/amd64
• aarch64

And the following operating systems:

• Linux
• MacOS

1. Checkout Source Code

Assume $WORKSPACE to be the directory under which the TiFlash repo is placed.

cd $WORKSPACE
git clone https://github.com/pingcap/tiflash.git --recursive -j 20

2. Prepare Prerequisites

The following packages are needed for all platforms:

• CMake 3.21.0+

• Rust: Recommended to use rustup to install:

  curl https://sh.rustup.rs -sSf | sh -s -- -y --profile minimal --default-toolchain nightly
  source $HOME/.cargo/env

• Python 3.0+

• Ninja or GNU Make

The following are platform-specific prerequisites. Click to expand details:

Linux specific prerequisites

TiFlash can be built using either LLVM or GCC toolchain on Linux. LLVM toolchain is our official one for releasing.

But for GCC, only GCC 7.x is supported as far, and is not planned to be a long term support. So it may get broken some day, silently.

• LLVM 13.0.0+

  TiFlash compiles using full LLVM toolchain (clang/compiler-rt/libc++/libc++abi) by default. You can use a system-wise toolchain if clang/compiler-rt/libc++/libc++abi can be installed in your environment.

  Click sections below to see detailed instructions:

  Set up LLVM via package managers in Debian/Ubuntu

  # add LLVM repo key
  wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key|sudo apt-key add -
  
  # install LLVM packages, and can find more detailed instructions in https://apt.llvm.org/ when failed
  apt-get install clang-13 lldb-13 lld-13 clang-tools-13 clang-13-doc libclang-common-13-dev libclang-13-dev libclang1-13 clang-format-13 clangd-13 clang-tidy-13 libc++-13-dev libc++abi-13-dev libomp-13-dev llvm-13-dev libfuzzer-13-dev
  
  # install other dependencies
  apt-get install lcov cmake ninja-build libssl-dev zlib1g-dev libcurl4-openssl-dev

  Set up LLVM via package managers in Archlinux

  # install compilers and dependencies
  sudo pacman -S clang libc++ libc++abi compiler-rt openmp lcov cmake ninja curl openssl zlib

• GCC 7.x

  WARNING: This support may not be maintained in the future.

  TiFlash compiles on GCC 7.x (no older, nor newer) only because it hasn't been broken. If you have GCC 7.x, you are probably fine, for now.

MacOS specific prerequisites

• Apple Clang 12.0.0+

• OpenSSL 1.1

  brew install openssl@1.1

3. Build

Assume $BUILD to be the directory under which you want to build TiFlash.

For Ninja:

cd $BUILD
cmake $WORKSPACE/tiflash -GNinja
ninja tiflash

For GNU Make:

cd $BUILD
cmake $WORKSPACE/tiflash
make tiflash -j

NOTE: Option -j (defaults to your system CPU core count, otherwise you can optionally specify a number) is used to control the build parallelism. Higher parallelism consumes more memory. If you encounter compiler OOM or hang, try to lower the parallelism by specifying a reasonable number, e.g., half of your system CPU core count or even smaller, after -j, depending on the available memory in your system.

After building, you can get TiFlash binary under $BUILD/dbms/src/Server/tiflash.

Build Options

TiFlash has several CMake build options to tweak for development purposes. These options SHOULD NOT be changed for production usage, as they may introduce unexpected build errors and unpredictable runtime behaviors.

To tweat options, pass one or multiple -D...=... args when invoking CMake, for example:

cmake $WORKSPACE/tiflash -DCMAKE_BUILD_TYPE=DEBUG

• Build Type:

  • -DCMAKE_BUILD_TYPE=RELWITHDEBINFO: Release build with debug info (default)

  • -DCMAKE_BUILD_TYPE=DEBUG: Debug build

  • -DCMAKE_BUILD_TYPE=RELEASE: Release build

• Build with Unit Tests:

  • -DENABLE_TESTS=OFF: Default

  • -DENABLE_TESTS=ON

• Build with System Libraries:

  Click to expand instructions

  For local development, it is sometimes handy to use pre-installed third-party libraries in the system, rather than to compile them from sources of the bundled (internal) submodules.

  Options are supplied to control whether to use internal third-party libraries (bundled in TiFlash) or to try using the pre-installed system ones.

  WARNING: It is NOT guaranteed that TiFlash would still build if any of the system libraries are used. Build errors are very likely to happen, almost all the time.

  You can view these options along with their descriptions by running:

  cd $BUILD
  cmake -LH | grep "USE_INTERNAL" -A3

  All of these options are default as ON, as the names tell, using the internal libraries and build from sources.

  There is another option to append extra paths for CMake to find system libraries:

  • PREBUILT_LIBS_ROOT: Default as empty, can be specified with multiple values, seperated by ;

  Specifically, for TiFlash proxy:

  • USE_INTERNAL_TIFLASH_PROXY=TRUE (default) / FALSE

    One may want to use external TiFlash proxy, e.g., if he is developing TiFlash proxy together with TiFlash, assume $TIFLASH_PROXY_REPO to be the path to the external TiFlash proxy repo

    Usually need to be combined with PREBUILT_LIBS_ROOT=$TIFLASH_PROXY_REPO, and $TIFLASH_PROXY_REPO should have the following directory structure:

    • Header files are under directory $TIFLASH_PROXY_REPO/raftstore-proxy/ffi/src

    • Built library is under directory $TIFLASH_PROXY_REPO/target/release

Run Unit Tests

To run unit tests, you need to build with -DCMAKE_BUILD_TYPE=DEBUG:

cd $BUILD
cmake $WORKSPACE/tiflash -GNinja -DCMAKE_BUILD_TYPE=DEBUG
ninja gtests_dbms       # Most TiFlash unit tests
ninja gtests_libdaemon  # Settings related tests
ninja gtests_libcommon

And the unit-test executables are at $BUILD/dbms/gtests_dbms, $BUILD/libs/libdaemon/src/tests/gtests_libdaemon and $BUILD/libs/libcommon/src/tests/gtests_libcommon.

Run Sanitizer Tests

TiFlash supports testing with thread sanitizer and address sanitizer.

To generate unit test executables with sanitizer enabled:

cd $BUILD
cmake $WORKSPACE/tiflash -GNinja -DENABLE_TESTS=ON -DCMAKE_BUILD_TYPE=ASan # or TSan
ninja gtests_dbms
ninja gtests_libdaemon
ninja gtests_libcommon

There are known false positives reported from leak sanitizer (which is included in address sanitizer). To suppress these errors, set the following environment variables before running the executables:

LSAN_OPTIONS=suppressions=$WORKSPACE/tiflash/test/sanitize/asan.suppression

Run Integration Tests

1. Build your own tiflash binary in $BUILD with -DCMAKE_BUILD_TYPE=DEBUG.

cd $BUILD
cmake $WORKSPACE/tiflash -GNinja -DCMAKE_BUILD_TYPE=DEBUG
ninja tiflash

2. Run tidb cluster locally using tiup playgroud or other tools.

tiup playground nightly --tiflash.binpath $BUILD/dbms/src/Server/tiflash

3. Check $WORKSPACE/tests/_env.sh to make the port and build dir right.
4. Run your integration tests using commands like "./run-test.sh fullstack-test2/ddl" under $WORKSPACE/tests dir

Run MicroBenchmark Tests

To run micro benchmark tests, you need to build with -DCMAKE_BUILD_TYPE=RELEASE -DENABLE_TESTS=ON:

cd $BUILD
cmake $WORKSPACE/tiflash -GNinja -DCMAKE_BUILD_TYPE=RELEASE -DENABLE_TESTS=ON
ninja bench_dbms       

And the microbenchmark-test executables are at $BUILD/dbms/bench_dbms, you can run it with ./bench_dbms or ./bench_dbms --benchmark_filter=xxx . More usage please check with ./bench_dbms --help.

Generate LLVM Coverage Report

TBD.

Contributing

Here is the overview of TiFlash architecture The architecture of TiFlash's distributed storage engine and transaction layer.

See TiFlash Development Guide and TiFlash Design documents.

Before submitting a pull request, please use format-diff.py to format source code, otherwise CI build may raise error.

NOTE: It is required to use clang-format 12.0.0+.

cd $WORKSPACE/tiflash
python3 format-diff.py --diff_from `git merge-base ${TARGET_REMOTE_BRANCH} HEAD`

License

TiFlash is under the Apache 2.0 license. See the LICENSE file for details.