Skip to content


Choose a tag to compare
@aarondav aarondav released this 18 Aug 02:47
· 4520 commits to master since this release

MLflow 0.5.0 offers some major improvements, including Keras and PyTorch first-class support as models, SFTP support as an artifactory, a new scatterplot visualization to compare runs, and a more complete Python SDK for experiment and run management.

Breaking changes:

  • The Tracking API has been split into two pieces, a "basic logging" API and a "tracking service" API. The "basic logging" API deals with logging metrics, parameters, and artifacts to the currently-active active run, and is accessible in mlflow (e.g., mlflow.log_param). The tracking service API allow managing experiments and runs (especially historical runs) and is available in mlflow.tracking. The tracking service API will look analogous to the upcoming R and Java Tracking Service SDKs. Please be aware of the following breaking changes:

    • mlflow.tracking no longer exposes the basic logging API, only mlflow. So, code that was written like from mlflow.tracking import log_param will have to be from mlflow import log_param (note that almost all examples were already doing this).

    • Access to the service API goes through the mlflow.tracking.get_service() function, which relies on the same tracking server set by either the environment variable MLFLOW_TRACKING_URI or by code with mlflow.tracking.set_tracking_uri(). So code that used to look like mlflow.tracking.get_run() will now have to do mlflow.tracking.get_service().get_run(). This does not apply to the basic logging API.

    • mlflow.ActiveRun has been converted into a lightweight wrapper around mlflow.entities.Run to enable the Python with syntax. This means that there are no longer any special methods on the object returned when calling mlflow.start_run(). These can be converted to the service API.

    • The Python entities returned by the tracking service API are now accessible in mlflow.entities directly. Where previously you may have used mlflow.entities.experiment.Experiment, you would now just use mlflow.entities.Experiment. The previous version still exists, but is deprecated and may be hidden in a future version.

  • REST API endpoint /ajax-api/2.0/preview/mlflow/artifacts/get has been moved to $static_prefix/get-artifact. This change is coversioned in the JavaScript, so should not be noticeable unless you were calling the REST API directly (#293, @andremchen)


  • [Models] Keras integration: we now support logging Keras models directly in the log_model API, model format, and serving APIs (#280, @ToonKBC)
  • [Models] PyTorch integration: we now support logging PyTorch models directly in the log_model API, model format, and serving APIs (#264, @vfdev-5)
  • [UI] Scatterplot added to "Compare Runs" view to help compare runs using any two metrics as the axes (#268, @ToonKBC)
  • [Artifacts] SFTP artifactory store added (#260, @ToonKBC)
  • [Sagemaker] Users can specify a custom VPC when deploying SageMaker models (#304, @dbczumar)
  • Pyfunc serialization now includes the Python version, and warns if the major version differs (can be suppressed by using load_pyfunc(suppress_warnings=True)) (#230, @dbczumar)
  • Pyfunc serve/predict will activate conda environment stored in MLModel. This can be disabled by adding --no-conda to mlflow pyfunc serve or mlflow pyfunc predict (#225, @0wu)
  • Python SDK formalized in mlflow.tracking. This includes adding SDK methods for get_run, list_experiments, get_experiment, and set_terminated. (#299, @aarondav)
  • mlflow run can now be run against projects with no conda.yaml specified. By default, an empty conda environment will be created -- previously, it would just fail. You can still pass --no-conda to avoid entering a conda environment altogether (#218, @smurching)

Bug fixes: