Skip to content

Contribution Licensing

Since this project is distributed under the terms of the BSL license, contributions that you make are licensed under the same terms. For us to be able to accept your contributions, we will need explicit confirmation from you that you are able and willing to provide them under these terms, and the mechanism we use to do this is the ArcticDB Individual Contributor License Agreement.

Individuals - To participate under these terms, please include the following line as the last line of the commit message for each commit in your contribution. You must use your real name (no pseudonyms, and no anonymous contributions).

Signed-Off By: Random J. Developer <[email protected]>. By including this sign-off line I agree to the terms of the Contributor License Agreement.

Corporations - For corporations who wish to make contributions to ArcticDB, please contact [email protected] and we will arrange for the CLA to be sent to the signing authority within your corporation.

Docker Quickstart

This quickstart builds a release using build dependencies from vcpkg. ArcticDB releases on PyPi use vcpkg dependencies in the manner as described below.

Note the below instructions will build a Linux X86_64 release.

1) Start the ArcticDB build docker image

Run in a Linux terminal:

docker pull
docker run -it

:warning: The below instructions do not have to be run in the provided docker image. They can be run against any Python installation on any Linux distribution as long as the basic build dependencies are available.

If running outside the provided docker image, please change /opt/python/cp39-cp39/bin/python3 in the examples below to an appropriate path for Python.

2) Check out ArcticDB including submodules

git clone
cd ArcticDB
git submodule init && git submodule update

3) Kick off the build

MY_PYTHON=/opt/python/cp39-cp39/bin/python3  # change if outside docker container/building against a different python
$MY_PYTHON -m pip install -U pip setuptools wheel grpcio-tools
# Change the below Python_EXECUTABLE value to build against a different Python version
cmake -DPython_EXECUTABLE=$MY_PYTHON -DTEST=off --preset linux-debug cpp
pushd cpp
cmake --build --preset linux-debug

4) Run ArcticDB

Ensure the below is run in the Git project root:

# PYTHONPATH first part = Python module, second part compiled C++ binary
PYTHONPATH=`pwd`/python:`pwd`/cpp/out/linux-debug-build/arcticdb/ $MY_PYTHON

Now, inside the Python shell:

from arcticdb import Arctic

Rather than setting the PYTHONPATH environment variable, you could install the appropriate paths into your Python environment by running (note that this will invoke the build tooling so will compile any changed files since the last compilation):

$MY_PYTHON -m pip install -ve .

Note that as this will copy the binary to your Python installation this will have to be run after each and every change of a C++ file.

Building using mamba and conda-forge

This section uses build dependencies from conda-forge. It is a pre-requisite for releasing ArcticDB on conda-forge.

⚠️ At the time of writing, installing ArcticDB with this setup under Windows is not possible due to a linkage problems with libprotobuf. See:

  • Install mamba
  • Create the arcticdb environment from its specification (environment_unix.yml):
mamba env create -f environment_unix.yml
  • Activate the arcticdb environment (you will need to do this for every new shell session):
mamba activate arcticdb

Building CMake targets

Several CMake presets are defined for build types, OS's, and build system.

For instance:

  • for debug build on Linux with mamba and conda-forge, use:
cmake -DTEST=off --preset linux-conda-debug cpp
cd cpp

# You might need to use fewer threads than what's possible on your machine
# to not have it swap and freeze (e.g. we use 1 here).
cmake --build --preset linux-conda-debug -j 1
  • for release build on MacOS with mamba and conda-forge, use:

cmake -DTEST=off --preset darwin-conda-release cpp
cd cpp

# You might need to use fewer threads than what's possible your machine
# not to have it swap and freeze (e.g. we use 1 here).
cmake --build --preset linux-conda-debug -j 1
Note: If you don't use presets, you may want to set some useful environment variables:

# To define the number of threads to use
# To enable C++ tests

Building and installing the Python Package

  • Build and install ArcticDB in the arcticdb environment using dependencies installed in this environment. We recommend using the editable installation for development:
# Adapt the CMake preset to your setup.
export ARCTIC_CMAKE_PRESET=linux-conda-debug
python -m pip install --no-build-isolation --no-deps --verbose --editable .
  • Use ArcticDB from Python:
from arcticdb import Arctic


How do I build against different Python versions?

Run cmake (configure, not build) with either:

  1. A different version of Python as the first version of Python on your PATH or...
  2. Point the Python_EXECUTABLE CMake variable to a different Python binary

Note that to build the ArcticDB C++ tests you must have the Python static library available in your installation!

How do I run the Python tests?

See running tests below.

How do I run the C++ tests?

See running tests below.

How do I specify how many cores to build using?

This is determined auto-magically by CMake at build time, but can be manually set by passing in --parallel <num cores> into the build command.

Detailed Build Information

Docker Image Construction

The above docker image is built from ManyLinux. Build script is located here.

GitHub output here.

We recommend you use this image for compilation and testing!

Setting up Linux

The codebase and build system can work with any reasonably recent Linux distribution with at least GCC 8 (10+ recommended) and CMake 3.12 (these instructions assume 3.21+).

A development install of Python 3.6+ (with libpython.a or .so and full headers) is also necessary. See pybind11 configuration.

We require a Mongo executable for a couple of Python tests on Linux. You can check whether you have it with mongod --version.

Search the internet for "mongo installation Linux" for instructions for your distro if you do not already have mongod available.

Dependencies by distro

Distro Versions reported to work Packages
Ubuntu 20.04, 22.04 build-essential g++-10 libpcre3-dev libsasl2-dev libsodium-dev libkrb5-dev libcurl4-openssl-dev python3-dev
Centos 7 devtoolset-10-gcc-c++ openssl-devel cyrus-sasl-devel devtoolset-10-libatomic-devel libcurl-devel python3-devel

Setting up Windows

We recommend using Visual Studio 2022 (or later) to install the compiler (MSVC v142 or newer) and tools (Windows SDK, CMake, Python).

The Python that comes with Visual Studio is sufficient for creating release builds, but for debug builds, you will have to separately download from

After building arcticdb_ext, you need to symlink to the .pyd for the Python tests to run against it:

# From the root of the ArcticDB git root checkout, in an administrator powershell session
# Change Python version as appropriate - below is for Python 3.11.
New-Item -Path .\python\arcticdb_ext.cp311-win_amd64.pyd -ItemType SymbolicLink -Value .\cpp\out\windows-cl-debug-build\arcticdb\arcticdb_ext.cp311-win_amd64.pyd

Running Python tests

With python pointing to a Python interpeter with ArcticDB installed/on the PYTHON_PATH:

python -m pip install arcticdb[Testing]
python -m pytest python/tests

Running C++ tests

Configure ArcticDB with TEST=ON (default is OFF):

cmake -DPython_EXECUTABLE=<path to python> -DTEST=ON --preset linux-debug cpp

Or you can set the following environment variable:


Note that <path to python> must point to a Python that is compatible with Development.Embed. This will probably be the result of installing python3-devel from your dependency manager.

Inside the provided docker image, python3-devel resolves to Python 3.6 installed at /usr/bin/python3, so the resulting command will be:

cmake -DPython_EXECUTABLE=/usr/bin/python3 -DTEST=ON --preset linux-debug cpp

Then build and run the tests:

cd cpp/out/linux-debug-build
make -j 1 arcticdb_rapidcheck_tests
make -j 1 test_unit_arcticdb
make test


Our source repo works with CIBuildWheel which runs the compilation and tests against all supported Python versions in isolated environments. Please follow their documentation.


CMake presets

To make it easier to set and share all the environment variables, config and commands, we recommend using the CMake presets feature.

Recent versions of some popular C++ IDEs support reading/importing these presets: * Visual Studio & Code * CLion

And it's equally easy to use on the command line.

We already ship a CMakePresets.json in the cpp directory, which is used by our builds. You can add a CMakeUserPresets.json in the same directory for local overrides. Inheritance is supported.

If you're working on Linux but not using our Docker image, you may want to create a preset with these cacheVariables: * CMAKE_MAKE_PROGRAM - make or ninja should work * CMAKE_C_COMPILER and CMAKE_CXX_COMPILER - If your preferred compiler is not cc and cxx

More examples:

Windows Preset to specify a Python version
  "version": 3,
  "configurePresets": [
      "name": "alt-vcpkg-debug:py3.10",
      "inherits": "windows-cl-debug",
      "cacheVariables": {
        "Python_ROOT_DIR": "C:\\Program Files\\Python310"
      "environment": {
        "PATH": "C:\\Users\\me\\AppData\\Roaming\\Python\\Python310\\Scripts;C:\\Program Files\\Python310;$penv{PATH}",
        "PYTHONPATH": "C:\\Program Files\\Python310\\Lib;C:\\Users\\me\\AppData\\Roaming\\Python\\Python310\\site-packages"
  "buildPresets": [
      "name": "alt-vcpkg-debug:py3.10",
      "configurePreset": "alt-vcpkg-debug:py3.10",
      "inheritConfigureEnvironment": true

vcpkg caching

We use vcpkg to manage the C++ dependencies.

Compiling the dependencies uses a lot of disk space. Once CMake configuration is done, you can remove the cpp\vcpkg\buildtrees folder.

You may also want to configure some caches: * Binary caching * Asset caching

pybind11 configuration

We augmented pybind11's Python discovery with our own PythonUtils to improve diagnostics. Please pay attention to warning messages from PythonUtils.cmake in the CMake output which highlights any configuration issues with Python.

We compile against the first python on the PATH by default.

To override that, use one of the following CMake variables*:

  • Python_ROOT_DIR - The common path "prefix" for a particular Python install. Usually, the python executable is in the same directory or the bin subdirectory. This directory should also contain the include and lib(rary) subdirectories.
    E.g. /usr for a system-wide install on *nix; /opt/pythonXX for a locally-managed Python install; /home/user/my_virtualenv for a virtual env; C:\Program Files\PythonXX for a Windows Python install

  • Python_EXECUTABLE - The path to the Python executable. (CMake 3.15+) CMake will try to extract the include and library paths by running this program. This differs from the default behaviour of FindPython.

(* Note CMake variables are set with -D on the CMake command line or with the cacheVariables key in CMake*Presets.json. The names are case-sensitive.

(Only) Python_ROOT_DIR can also be set as an environment variable. Setting the others in the environment might have no effect.)

Development Guidelines

ArcticDB has a lot of write-time options to configure how the data is stored on disk. When adding new features, it is important to test not only the simplest case, but also how the new feature interacts with these various options. This section serves as a guide as to what should be considered when adding a new feature.

Backwards compatibility

Data stored on disk

Due to ArcticDB being a purely client-side database, it is possible for data to be written by clients that are of later versions than clients that need to read the same data. If the change is breaking, such that older clients will not be able to read the data, then this should be clearly documented both in the PR and in the online documentation. If possible, a version number should also be added such that future changes in the same area will display a more clear error message.


Any changes to the API (including when exceptions are thrown and the type of the exception) must be weighed up against the change breaking behaviour for existing users. Please make this as clear to reviewers as possible by ensuring API changes are clearly described in the PR description. If the change is breaking, please also ensure that that is appropriately highlighted in the PR description as well. This is particularly true of the NativeVersionStore API, as this has many users inside Man Group.


Symbols are almost completely decoupled from one another in ArcticDB. The major feature for which this is not the case is snapshots. Whenever a change involves deleting any keys from the storage, care must be taken that those keys are not in fact still needed by a snapshot. Similarly, any operation that modifies a snapshot, including deleting it, must ensure that any keys for which this snapshot was the last reference are also deleted. See the tutorial and API documentation for more details.

Batch methods

Almost all of the major reading and writing methods have batch equivalents. If a feature is added for the non-batch version, it should almost always work with the batch version as well. We should also strive for consistency of behaviour across the batch methods in circumstances that could be encountered by any of them. e.g. read_batch and read_metadata_batch should have the same behaviour if the specified version does not exist.

Dynamic schema

Static schema is much simpler than dynamic schema, and so features added and only tested with static schema may not "just work" with dynamic schema. In particular, the following cases should be considered:

  • A column that is not present in an initial call to write is present in a subsequent call to append.
  • A column that is present in an initial call to write is not present in a subsequent call to append.
  • A numeric column with a "small" type (e.g. uint8) is appended to with a numeric column of a larger type (e.g. int16), and vice versa.


Most tests are written with small amounts of data to keep the runtime down as low as possible. However, many bugs are only exposed when data being written is column-sliced and/or row-sliced. Using test fixtures configured to slice into small data segments (e.g. 2x2) can help to catch these issues early.

Data sub-selection

Many use cases of ArcticDB involve writing huge symbols with hundreds of thousands of columns or billions of rows. In these cases, methods for reading data will almost never be called without some parameters to cut down the amount of data returned to the user. The most commonly used are:

  • The stand-alone methods head and tail*
  • columns - select only a subset of the columns
  • date_range and row_range - select a contiguous range of rows
  • query_builder - see section on processing pipeline

Processing pipeline

Reading data using only columns, date_range, and row_range arguments is heavily optimised to avoid copying data in memory unnecessarily. Any read-like method provided with the query_builder optional argument takes quite a different code path, as we do not know the shape of the output data a priori. While it is not necessary to exhaustively test all new features against every possible query_builder argument, it is generally worth having a test using a simple filter that excludes some of the data that would otherwise be returned to the user.

Sparse columns

The majority of data written into ArcticDB uses dense columns, meaning that within a given row-slice, every row has an associated value in every column (even if that value is NaN). The concept of sparse columns is implemented as well, where columns may have values missing for none, some, or all rows within a row-slice. Any read-like method should be able to handle both types of stored column.


When data that cannot be normalized to a supported data type in ArcticDB, it can still be stored using write_pickle and similar. There are many operations that cannot be performed on pickled data, such as append, update, date_range search, and many more. It is important that if a user attempts an operation that is not supported with pickled data, that they receive a helpful error message.