tree: 17c0c2dc9f502e5fc06809b6dcb573b61c9ed181 [path history] [tgz]
  1. .gitignore
  3. CMakeLists.txt
  4. Dockerfile
  6. apidoc/
  7. build-support/
  8. cmake_modules/
  9. doc/
  10. examples/
  11. src/
  12. submodules/
  13. thirdparty/
  14. tools/
  15. valgrind.supp

Apache Arrow C++ codebase

This directory contains the code and build system for the Arrow C++ libraries, as well as for the C++ libraries for Apache Parquet.

System setup

Arrow uses CMake as a build configuration system. Currently, it supports in-source and out-of-source builds with the latter one being preferred.

Building Arrow requires:

  • A C++11-enabled compiler. On Linux, gcc 4.8 and higher should be sufficient.
  • CMake
  • Boost

On Ubuntu/Debian you can install the requirements with:

sudo apt-get install cmake \
     libboost-dev \
     libboost-filesystem-dev \

On macOS, you can use Homebrew:

git clone
cd arrow
brew update && brew bundle --file=c_glib/Brewfile

If you are developing on Windows, see the Windows developer guide.

Building Arrow

Simple debug build:

git clone
cd arrow/cpp
mkdir debug
cd debug
cmake ..
make unittest

Simple release build:

git clone
cd arrow/cpp
mkdir release
cd release
cmake .. -DCMAKE_BUILD_TYPE=Release
make unittest

Detailed unit test logs will be placed in the build directory under build/test-logs.

On some Linux distributions, running the test suite might require setting an explicit locale. If you see any locale-related errors, try setting the environment variable (which requires the locales package or equivalent):

export LC_ALL="en_US.UTF-8"

Building and Developing Parquet Libraries

To build the C++ libraries for Apache Parquet, add the flag -DARROW_PARQUET=ON when invoking CMake. The Parquet libraries and unit tests can be built with the parquet make target:

make parquet

Running ctest -L unittest will run all built C++ unit tests, while ctest -L parquet will run only the Parquet unit tests. The unit tests depend on an environment variable PARQUET_TEST_DATA that depends on a git submodule to the repository

git submodule update --init
export PARQUET_TEST_DATA=$ARROW_ROOT/cpp/submodules/parquet-testing/data

Here $ARROW_ROOT is the absolute path to the Arrow codebase.

Statically linking to Arrow on Windows

The Arrow headers on Windows static library builds (enabled by the CMake option ARROW_BUILD_STATIC) use the preprocessor macro ARROW_STATIC to suppress dllimport/dllexport marking of symbols. Projects that statically link against Arrow on Windows additionally need this definition. The Unix builds do not use the macro.

Building/Running benchmarks

Follow the directions for simple build except run cmake with the --ARROW_BUILD_BENCHMARKS parameter set correctly:


and instead of make unittest run either make; ctest to run both unit tests and benchmarks or make runbenchmark to run only the benchmark tests.

Benchmark logs will be placed in the build directory under build/benchmark-logs.

Testing with LLVM AddressSanitizer

To use AddressSanitizer (ASAN) to find bad memory accesses or leaks with LLVM, pass -DARROW_USE_ASAN=ON when building. You must use clang to compile with ASAN, and ARROW_USE_ASAN is mutually-exclusive with the valgrind option ARROW_TEST_MEMCHECK.

Building/Running fuzzers

Fuzzers can help finding unhandled exceptions and problems with untrusted input that may lead to crashes, security issues and undefined behavior. They do this by generating random input data and observing the behavior of the executed code. To build the fuzzer code, LLVM is required (GCC-based compilers won't work). You can build them using the following code:


ARROW_FUZZING will enable building of fuzzer executables as well as enable the addition of coverage helpers via ARROW_USE_COVERAGE, so that the fuzzer can observe the program execution.

It is also wise to enable some sanitizers like ARROW_USE_ASAN (see above), which activates the address sanitizer. This way, we ensure that bad memory operations provoked by the fuzzer will be found early. You may also enable other sanitizers as well. Just keep in mind that some of them do not work together and some may result in very long execution times, which will slow down the fuzzing procedure.

Now you can start one of the fuzzer, e.g.:


This will try to find a malformed input that crashes the payload and will show the stack trace as well as the input data. After a problem was found this way, it should be reported and fixed. Usually, the fuzzing process cannot be continued until the fix is applied, since the fuzzer usually converts to the problem again.

If you build fuzzers with ASAN, you need to set the ASAN_SYMBOLIZER_PATH environment variable to the absolute path of llvm-symbolizer, which is a tool that ships with LLVM.

export ASAN_SYMBOLIZER_PATH=$(type -p llvm-symbolizer)

Note that some fuzzer builds currently reject paths with a version qualifier (like llvm-sanitizer-5.0). To overcome this, set an appropriate symlink (here, when using LLVM 5.0):

ln -sf /usr/bin/llvm-sanitizer-5.0 /usr/bin/llvm-sanitizer

There are some problems that may occur during the compilation process:

  • libfuzzer was not distributed with your LLVM: ld: file not found: .../libLLVMFuzzer.a
  • your LLVM is too old: clang: error: unsupported argument 'fuzzer' to option 'fsanitize='

Third-party dependencies and configuration

Arrow depends on a number of third-party libraries. We support these in a few ways:

  • Building dependencies from source by downloading archives from the internet
  • Building dependencies from source using from local archives (to allow offline builds)
  • Building with locally-installed libraries

See thirdparty/ for details about these options and how to configure your build toolchain.

Building Python integration library (optional)

The optional arrow_python shared library can be built by passing -DARROW_PYTHON=on to CMake. This must be installed or in your library load path to be able to build pyarrow, the Arrow Python bindings.

The Python library must be built against the same Python version for which you are building pyarrow, e.g. Python 2.7 or Python 3.6. NumPy must also be installed.

Building GPU extension library (optional)

The optional arrow_gpu shared library can be built by passing -DARROW_GPU=on. This requires a CUDA installation to build, and to use many of the functions you must have a functioning GPU. Currently only CUDA functionality is supported, though if there is demand we can also add OpenCL interfaces in this library as needed.

The CUDA toolchain used to build the library can be customized by using the $CUDA_HOME environment variable.

This library is still in Alpha stages, and subject to API changes without deprecation warnings.

Building Apache ORC integration (optional)

The optional arrow reader for the Apache ORC format (found in the arrow::adapters::orc namespace) can be built by passing -DARROW_ORC=on. This is currently not supported on windows. Note that this functionality is still in Alpha stages, and subject to API changes without deprecation warnings.

API documentation

To generate the (html) API documentation, run the following command in the apidoc directory:

doxygen Doxyfile

This requires Doxygen to be installed.


This project follows Google's C++ Style Guide with minor exceptions. We do not encourage anonymous namespaces and we relax the line length restriction to 90 characters.

Memory Pools

We provide a default memory pool with arrow::default_memory_pool(). As a matter of convenience, some of the array builder classes have constructors which use the default pool without explicitly passing it. You can disable these constructors in your application (so that you are accounting properly for all memory allocations) by defining ARROW_NO_DEFAULT_MEMORY_POOL.

Error Handling and Exceptions

For error handling, we use arrow::Status values instead of throwing C++ exceptions. Since the Arrow C++ libraries are intended to be useful as a component in larger C++ projects, using Status objects can help with good code hygiene by making explicit when a function is expected to be able to fail.

For expressing invariants and “cannot fail” errors, we use DCHECK macros defined in arrow/util/logging.h. These checks are disabled in release builds and are intended to catch internal development errors, particularly when refactoring. These macros are not to be included in any public header files.

Since we do not use exceptions, we avoid doing expensive work in object constructors. Objects that are expensive to construct may often have private constructors, with public static factory methods that return Status.

There are a number of object constructors, like arrow::Schema and arrow::RecordBatch where larger STL container objects like std::vector may be created. While it is possible for std::bad_alloc to be thrown in these constructors, the circumstances where they would are somewhat esoteric, and it is likely that an application would have encountered other more serious problems prior to having std::bad_alloc thrown in a constructor.

Extra debugging help

If you use the CMake option -DARROW_EXTRA_ERROR_CONTEXT=ON it will compile the libraries with extra debugging information on error checks inside the RETURN_NOT_OK macro. In unit tests with ASSERT_OK, this will yield error outputs like:

../src/arrow/ipc/ Failure
NotImplemented: ../src/arrow/ipc/ code: writer->WriteRecordBatch(batch)
../src/arrow/ipc/ code: CheckStarted()
../src/arrow/ipc/ code: schema_writer.Write(&dictionaries_)
../src/arrow/ipc/ code: WriteSchema()
../src/arrow/ipc/ code: WriteSchemaMessage(schema_, dictionary_memo_, &schema_fb)
../src/arrow/ipc/ code: SchemaToFlatbuffer(fbb, schema, dictionary_memo, &fb_schema)
../src/arrow/ipc/ code: FieldToFlatbuffer(fbb, *schema.field(i), dictionary_memo, &offset)
../src/arrow/ipc/ code: TypeToFlatbuffer(fbb, *field.type(), &children, &layout, &type_enum, dictionary_memo, &type_offset)
Unable to convert type: decimal(19, 4)

Deprecations and API Changes

We use the compiler definition ARROW_NO_DEPRECATED_API to disable APIs that have been deprecated. It is a good practice to compile third party applications with this flag to proactively catch and account for API changes.

Keeping includes clean with include-what-you-use

We have provided a build-support/iwyu/ convenience script for invoking Google's include-what-you-use tool, also known as IWYU. This includes various suppressions for more informative output. After building IWYU (following instructions in the README), you can run it on all files by running:

CC="clang-4.0" CXX="clang++-4.0" cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ..
../build-support/iwyu/ all

This presumes that include-what-you-use and are in your $PATH. If you compiled IWYU using a different version of clang, then substitute the version number above accordingly.

We have provided a Docker-based IWYU to make it easier to run these checks. This can be run using the docker-compose setup in the dev/ directory

# If you have not built the base image already
docker build -t arrow_integration_xenial_base -f dev/docker_common/Dockerfile.xenial.base .

dev/ iwyu


We require that you follow a certain coding style in the C++ code base. You can check your code abides by that coding style by running:

make lint

You can also fix any formatting errors automatically:

make format

These commands require clang-format-6.0 (and not any other version). You may find the required packages at or use the Debian/Ubuntu APT repositories on On macOS with Homebrew you can get it via brew install llvm@6.

Checking for ABI and API stability

To build ABI compliance reports, you need to install the two tools abi-dumper and abi-compliance-checker.

Build Arrow C++ in Debug mode, alternatively you could use -Og which also builds with the necessary symbols but includes a bit of code optimization. Once the build has finished, you can generate ABI reports using:

abi-dumper -lver 9 debug/ -o ABI-9.dump

The above version number is freely selectable. As we want to compare versions, you should now git checkout the version you want to compare it to and re-run the above command using a different version number. Once both reports are generated, you can build a comparision report using

abi-compliance-checker -l libarrow -d1 ABI-PY-9.dump -d2 ABI-PY-10.dump

The report is then generated in compat_reports/libarrow as a HTML.

Continuous Integration

Pull requests are run through travis-ci for continuous integration. You can avoid build failures by running the following checks before submitting your pull request:

make unittest
make lint
# The next command may change your code.  It is recommended you commit
# before running it.
make format # requires clang-format is installed

We run our CI builds with more compiler warnings enabled for the Clang compiler. Please run CMake with


to avoid failures due to compiler warnings.

Note that the clang-tidy target may take a while to run. You might consider running clang-tidy separately on the files you have added/changed before invoking the make target to reduce iteration time. Also, it might generate warnings that aren‘t valid. To avoid these you can add a line comment // NOLINT. If NOLINT doesn’t suppress the warnings, you add the file in question to the .clang-tidy-ignore file. This will allow make check-clang-tidy to pass in travis-CI (but still surface the potential warnings in make clang-tidy). Ideally, both of these options would be used rarely. Current known uses-cases when they are required:

  • Parameterized tests in google test.