MLflow Cloudsmith Plugin

MLflow Artifact Repository plugin that stores artifacts as Cloudsmith RAW packages.

Key Features

• Seamless MLflow integration (cloudsmith://owner/repo)
• List/Download artifacts within the MLflow UI
• Organized via tags (mlflow, experiment-<id>, run-<id>, path-<artifact_path>)

Installation

pip install mlflow-cloudsmith

Usage with MLflow

import os
import mlflow

os.environ["CLOUDSMITH_API_KEY"] = "<your-api-key>"
os.environ["MLFLOW_ARTIFACT_URI"] = "cloudsmith://<owner>/<repo>"

with mlflow.start_run():
    mlflow.log_param("learning_rate", 0.01)
    mlflow.log_metric("accuracy", 0.95)
    mlflow.log_artifact("model.pkl")

Direct Repository Usage

from plugin.cloudsmith_repository import CloudsmithArtifactRepository

repo = CloudsmithArtifactRepository("cloudsmith://<owner>/<repo>")
repo.log_artifact("model.pkl", "models/production")

for info in repo.list_artifacts("models"):
    print(info.path, info.file_size, info.is_dir)

Example startup:

mlflow server \
  --host 127.0.0.1 \
  --port 5000 \
  --artifacts-destination cloudsmith://<CLOUDSMITH_NAMESPACE>/<CLOUDSMITH_REPO>

URI Format

cloudsmith://<owner>/<repository>[/<path>]

Examples:

• cloudsmith://my-org/ml-artifacts
• cloudsmith://my-org/ml-artifacts/experiments

Configuration

How It Works (Brief)

• Each MLflow artifact is uploaded as a Cloudsmith RAW package with preserved original filename and metadata.
• The plugin builds an in-memory tree of artifact paths and returns only immediate children for UI browsing.
• Packages are tagged for easy filtering: mlflow, experiment-*, run-*, path-* (slashes replaced with dashes).

Artifact Representation & Tagging

Example Run Context

• experiment_id: 123

• run_id: 0123456789abcdef0123456789abcdef

• Files logged:

  • models/model.pkl
  • conda.yaml

Package Details

• Name: mlflow-<base-filename>-<run_id_first8>-<timestamp> e.g., mlflow-model-01234567-1754914964

• Version: <experiment_id>+<run_id> e.g., 123+0123456789abcdef0123456789abcdef

• Filename: Original filename (e.g., model.pkl)

• Description:

  MLflow artifact: <artifact_path> (experiment: <experiment_id>, run: <run_id>)
  

  e.g., MLflow artifact: models/model.pkl (experiment: 123, run: 0123456789abcdef0123456789abcdef)

• Tags:

  • mlflow
  • experiment-123
  • run-0123456789abcdef0123456789abcdef
  • path-models-model.pkl

Notes

• Descriptions hold the authoritative artifact path for listing & downloads.

• path-* tags replace / with - for fallback reconstruction.

• MLflow UI listing:

  • list_artifacts("") → [models/ (dir), conda.yaml]
  • list_artifacts("models") → [models/model.pkl]

Testing

pytest -q

Integration Tests (Opt-in)

export CLOUDSMITH_RUN_INTEGRATION=1
export CLOUDSMITH_API_KEY=""      # required
export CLOUDSMITH_TEST_OWNER=""   # required
export CLOUDSMITH_TEST_REPO=""    # required

pytest -q

Cleanup Script (Delete by Run/Experiment)

scripts/cleanup_orphans.sh deletes Cloudsmith RAW packages for a given run or experiment ID. No MLflow server is contacted. Requires: bash, curl, jq.

Environment Variables / Flags

Examples:

# Dry-run: show packages for a run-id
CLOUDSMITH_API_KEY="" CLOUDSMITH_OWNER=myorg CLOUDSMITH_REPO=myrepo \
    scripts/cleanup_orphans.sh --run-id 0123456789abcdef0123456789abcdef

# Delete for a run-id (confirmation required)
CLOUDSMITH_API_KEY="" CLOUDSMITH_OWNER=myorg CLOUDSMITH_REPO=myrepo \
    scripts/cleanup_orphans.sh --run-id 0123456789abcdef0123456789abcdef --confirm

# Combine experiment-id + run-id
CLOUDSMITH_API_KEY="" CLOUDSMITH_OWNER=myorg CLOUDSMITH_REPO=myrepo \
    scripts/cleanup_orphans.sh --experiment-id 123 --run-id 0123456789abcdef0123456789abcdef --confirm

Contributing

We welcome contributions! Please see CONTRIBUTING for more details.

License

Apache License 2.0 — see LICENSE.