ML Metadata

ML Metadata (MLMD) is a library for recording and retrieving metadata associated with ML developer and data scientist workflows.

NOTE: ML Metadata may be backwards incompatible before version 1.0.

Getting Started

For more background on MLMD and instructions on using it, see the getting started guide

Installing from PyPI

The recommended way to install ML Metadata is to use the PyPI package:

pip install ml-metadata

Then import the relevant packages:

from ml_metadata import metadata_store
from ml_metadata.proto import metadata_store_pb2

Nightly Packages

ML Metadata (MLMD) also hosts nightly packages at https://pypi-nightly.tensorflow.org on Google Cloud. To install the latest nightly package, please use the following command:

pip install --extra-index-url https://pypi-nightly.tensorflow.org/simple ml-metadata

Installing with Docker

This is the recommended way to build ML Metadata under Linux, and is continuously tested at Google.

Please first install docker and docker-compose by following the directions: docker; docker-compose.

Then, run the following at the project root:

DOCKER_SERVICE=manylinux-python${PY_VERSION}
sudo docker compose build ${DOCKER_SERVICE}
sudo docker compose run ${DOCKER_SERVICE}

where PY_VERSION is one of {310, 311, 312, 313}.

A wheel will be produced under dist/, and installed as follows:

pip install dist/*.whl

Installing from source

1. Prerequisites

To compile and use ML Metadata, you need to set up some prerequisites.

Install Bazel

If Bazel is not installed on your system, install it now by following these directions.

Install cmake

If cmake is not installed on your system, install it now by following these directions.

2. Clone ML Metadata repository

git clone https://github.com/google/ml-metadata
cd ml-metadata

Note that these instructions will install the latest master branch of ML Metadata. If you want to install a specific branch (such as a release branch), pass -b <branchname> to the git clone command.

3. Build the pip package

ML Metadata uses Bazel to build the pip package from source:

python setup.py bdist_wheel

You can find the generated .whl file in the dist subdirectory.

4. Install the pip package

pip install dist/*.whl

5.(Optional) Build the grpc server

ML Metadata uses Bazel to build the c++ binary from source:

bazel build -c opt --define grpc_no_ares=true  //ml_metadata/metadata_store:metadata_store_server

Supported platforms

MLMD is built and tested on the following 64-bit operating systems:

macOS 10.14.6 (Mojave) or later.
Ubuntu 20.04 or later.
[DEPRECATED] Windows 10 or later. For a Windows-compatible library, please refer to MLMD 1.14.0 or earlier versions.

Releasing Wheels to PyPI

Setup (Required for both release methods)

Before releasing, you need to set up the PyPI environment and token once:

Step 1: Create PyPI environment

Create a new environment named pypi in the GitHub repository:

Go to https://github.com/google/ml-metadata/settings/environments/new
Name it pypi
Click "Configure environment"

Step 2: Add PYPI_API_TOKEN secret

Add your PyPI token to the pypi environment:

In the pypi environment settings, scroll to "Environment secrets"
Click "Add secret"
Name: PYPI_API_TOKEN (use this exact name)
Value: Your PyPI API token
Click "Add secret"

Step 3: Commit and push your release branch

Ensure your release branch has the correct version set in ml_metadata/version.py, then:

git add ml_metadata/version.py
git commit -m "Prepare release vX.Y.Z"
git push origin your-release-branch

Part 1: Releasing via `workflow_dispatch`

This method allows you to manually trigger a release from any branch without creating a GitHub release.

Steps (after completing setup above):

Navigate to the GitHub Actions page: https://github.com/google/ml-metadata/actions
Find and select the Build ml-metadata with Conda workflow: https://github.com/google/ml-metadata/actions/workflows/conda-build.yml
Click the "Run workflow" dropdown button.
Select your release branch from the dropdown menu.
Click "Run workflow".

The workflow will build wheels for all supported Python versions and automatically upload them to PyPI if the token is configured correctly.

Part 2: Releasing via GitHub Release

This method creates a formal GitHub release with a tag, which automatically triggers the build and upload workflow.