MongoDB Backend for MLflow

An MLflow plugin that stores tracking data (experiments, runs, traces, and logged models) and model registry data (registered models, versions, aliases, tags, and prompts) in MongoDB. The tracking store and registry store are configured independently, so either one can use MongoDB while the other uses another MLflow backend.

Licensed under the Apache License 2.0.

Installation

python -m pip install mlflow-mongodb

Supported versions

Component	Minimum	Tested up to
Python	3.11	3.14
MLflow	3.5	3.17
PyMongo	4.0	4.19
MongoDB	8.0	8.0

MLflow 4.x is not supported. Every pull request and release is tested against both the minimum and the latest of these combinations.

Running MLflow

The following configuration stores both tracking data and registry data in MongoDB:

export TRACKING_STORE_URI="mongodb://localhost:27017/mlflow"
export REGISTRY_STORE_URI="mongodb://localhost:27017/mlflow"
export MLFLOW_BIND_HOST="127.0.0.1"
export MLFLOW_BIND_PORT="5000"

mlflow server \
  --backend-store-uri "${TRACKING_STORE_URI}" \
  --registry-store-uri "${REGISTRY_STORE_URI}" \
  --host "${MLFLOW_BIND_HOST}" \
  --port "${MLFLOW_BIND_PORT}"

Open http://127.0.0.1:5000 to use the MLflow UI. Clients can connect to the server with:

export MLFLOW_TRACKING_URI="http://127.0.0.1:5000"

The MongoDB URI must include the database name; mlflow is the database name in the example above. Authentication, replica-set, TLS, and other connection options can be provided using standard MongoDB URI syntax. Both mongodb:// and mongodb+srv:// schemes are supported, so a MongoDB Atlas connection string works unchanged. For example:

mongodb://username:password@mongo.example.com:27017/mlflow?authSource=admin
mongodb+srv://username:password@cluster0.example.mongodb.net/mlflow

When using the MongoDB tracking backend, MLflow 3.5 clients upload full trace payloads to MLflow's existing artifact storage. Configure a reachable experiment artifact location, such as S3 or an MLflow artifact proxy. The MongoDB store returns the trace artifact URI; it does not implement its own artifact provider or upload the payload itself.

On MLflow 3.5 servers, trace retrieval uses artifacts while spans are also persisted in MongoDB. On MLflow 3.6 and newer servers, modern clients use MongoDB-backed trace retrieval. A 3.5 client cannot read a trace written only to MongoDB by a modern client unless an artifact payload also exists. Cross-version reads of that kind are not guaranteed.

Aggregated trace metric queries require MLflow 3.9 or newer; available metrics and dimensions depend on the installed MLflow version. Older versions preserve native MLflow spans without newer OTel translations, and derive cost summaries only when MLflow provides its cost aggregator. These limitations do not disable run metric logging.

Documentation

New to MLflow? Start with the MLflow Tracking Quickstart. The following guides provide additional context for using this plugin:

• Model Registry Tutorial — register and manage model versions.
• Backend Stores — understand where MLflow stores tracking data.
• Tracking Server Configuration — configure tracking and registry stores independently.
• MongoDB Connection Strings — configure authentication, TLS, replica sets, and other URI options.

Reporting issues

Bug reports and feature requests are tracked in GitHub Issues. When reporting a bug, include the MLflow, PyMongo, MongoDB, and mlflow-mongodb versions, the server command you ran, and a minimal reproduction. Do not include connection strings or credentials.

Security vulnerabilities must not be filed as public issues. Report them through MongoDB's vulnerability reporting process instead.

Versioning

This project follows Semantic Versioning (MAJOR.MINOR.PATCH), same as MLflow itself. It's currently on major version 0, meaning the API isn't yet stable and may change between releases. main always carries a .devN version (e.g. 0.1.0.dev0) to signal it's unreleased and ahead of the latest published version.

Verifying releases

Releases are published to PyPI through PyPI's trusted publishing, so no long-lived upload credentials exist. Each GitHub release also carries the wheel and source distribution, a detached GPG signature (.sig) for each, and a SHA256SUMS file.

The signatures are made with MongoDB's Garasign service using this key:

Key	ai-builder-experience <packaging@mongodb.com>
Fingerprint	6E33 09C7 3676 2845 73EC 0E52 5994 2DBC 12A6 CD89
Public key	https://pgp.mongodb.com/ai-builder-experience.pub

To verify a release, download the distribution, its .sig, and SHA256SUMS from the release page, import the key, and check the checksum and signature:

curl -sSLO https://pgp.mongodb.com/ai-builder-experience.pub
gpg --import ai-builder-experience.pub
sha256sum -c SHA256SUMS --ignore-missing
gpg --verify mlflow_mongodb-<version>-py3-none-any.whl.sig mlflow_mongodb-<version>-py3-none-any.whl

Expect Good signature from "ai-builder-experience <packaging@mongodb.com>" and the fingerprint above. Compare the fingerprint against this README, not only against the downloaded key, so a swapped key cannot verify itself.

Contributor setup

For development, clone the repository and install it in editable mode:

git clone https://github.com/mongodb-developer/mlflow-mongodb.git
cd mlflow-mongodb

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .

If you downloaded a source archive instead, extract it and run the venv and pip commands from the extracted repository root. The final . in pip install -e . means “install the project in the current directory,” and -e keeps the installation linked to that source checkout.

Install the development dependencies and Git hooks with:

python -m pip install -e ".[dev]"
pre-commit install --install-hooks

Run all formatting, linting, and repository checks with:

pre-commit run --all-files

Running tests

Install the project in editable mode with its development dependencies first:

python -m pip install -e ".[dev]"

The unit tests need no database:

python -m pytest tests/unit

The functional tests use a real MongoDB 8.0 or newer server; a standalone server is fine, and CI runs them against a single-node replica set. Set MONGODB_URI to a dedicated test database; for example:

export MONGODB_URI="mongodb://localhost:27017/mlflow_functional_test"
python -m pytest tests/functional

Authentication, replica-set, TLS, and other standard MongoDB URI options can be included in MONGODB_URI.

The functional suite deletes documents from its application collections after every test so their indexes can be reused. At the end of the test session, it drops the database selected by MONGODB_URI. Never point MONGODB_URI at a database containing data that must be preserved.