MLflow Export Import

This package provides tools to copy MLflow objects (runs, experiments or registered models) from one MLflow tracking server (Databricks workspace) to another.

For more details on MLflow objects (Databricks MLflow) see the Databricks MLflow Object Relationships slide deck.

Architecture

Overview

Why use MLflow Export Import?

• Share and collaborate with other data scientists in the same tracking server (Databricks workspace).
  • For example, clone an experiment from another user.
• Share and collaborate with other data scientists in different tracking servers.
  • For example, clone an experiment from a different user in another tracking server.
• MLOps CI/CD. Migrate runs (or registered models) to another tracking server.
  • Promote a run from the development to the test tracking server.
  • After it passes tests, then promote it to the production tracking server.
• Backup your MLflow objects to external storage so they can be restored if needed.
• Disaster recovery. Save your MLflow objects to external storage so they can be replicated to another tracking server.

MLflow Export Import scenarios

Source tracking server	Destination tracking server	Note
Open source	Open source	common
Open source	Databricks	less common
Databricks	Databricks	common
Databricks	Open source	rare

Two sets of tools

• Open source MLflow Python scripts.
• Databricks notebooks that invoke the Python scripts.

Tools Overview

Python Scripts

There are two sets of Python scripts:

• Individual tools. Use these tools to copy individual MLflow objects between tracking servers. They allow you to specify a different destination object name. For example, if you want to clone the experiment /Mary/Experiments/Iris under a new name, you can specify the target experiment name as /John/Experiments/Iris.

• Collection tools. High-level tools to copy an entire tracking server or a collection of MLflow objects. Full object referential integrity is maintained as well as the original MLflow object names.

Databricks notebooks

Databricks notebooks simply invoke their corresponding Python scripts. Note that only Individual notebooks are currently available.

See README.

Other

• Miscellanous tools

Limitations

General Limitations

• Nested runs are only supported when you import an experiment. For a run, it is still a TODO.
• If the run linked to a registered model version does not exist (has been deleted) the version is not exported since when importing MLflowClient.create_model_version requires a run ID.
• If you need to preserve registered model version nummbers do not use the use-threads option since version numbers will not be exported or imported sequentially.
• Run tags are always exported as a string even if they are an int since the MlflowClienti.get_run() does not return tag type information.
• Importing from a file-based --backend-store-uri implementation is not supported since it does not have the same semantics as a database-based implementation (e.g. primary key constraints are not respected, model registry is not implemented, etc.). This is is not a limitation of mlflow-export-import but rather of the MLflow file-based implementation which is not meant for production. `

Databricks Limitations

Exporting Notebook Revisions

• The notebook revision associated with a run can be exported. It is stored as an artifact in the run's notebooks artifact directory.
• You can save the notebook in the suppported SOURCE, HTML, JUPYTER and DBC formats.
• Examples: notebooks/notebook.dbc or notebooks/notebook.source.

Importing Notebooks

• Partial functionality due to Databricks REST API limitations.
• The Databricks REST API does not support:
  • Importing a notebook with its revision history.
  • Linking an imported run with its associated imported notebook revision.
• The API does allow you to export a notebook revision, but it is simply a notebook with one revision.
• The notebook is exported as a run artifact for convenience.
• When you import a run, the link to its source notebookRevisionID tag will be a dead link and you cannot access the notebook from the MLflow UI. See the dst-notebook-dir option for Import Run.
• As another convenience, the import tools allows you to import the exported notebook into a Databricks workspace directory. For more details, see:
  • README_individual - Import run
  • README_individual - Import experiment
• You must export a notebook in the SOURCE format for the notebook to be imported.

Used ID

• When importing a run or experiment, for open source (OSS) MLflow you can specify a different user owner.
• OSS MLflow - the destination run mlflow.user tag can be the same as the source mlflow.user tag since OSS MLflow allows you to set this tag.
• Databricks MLflow - you cannot set the mlflow.user tag. The mlflow.user will be based upon the personal access token (PAT) of the importing user.

Common options details

notebook-formats - If exporting a Databricks run, the run's notebook revision can be saved in the specified formats (comma-delimited argument). Each format is saved in the notebooks folder of the run's artifact root directory as notebook.{format}. Supported formats are SOURCE, HTML, JUPYTER and DBC. See Databricks Export Format documentation.

use-src-user-id - Set the destination user ID to the source user ID. Source user ID is ignored when importing into Databricks since the user is automatically picked up from your Databricks access token.

export-source-tags - Exports source information under the mlflow_export_import tag prefix. See section below for details.

MLflow Export Import Source Run Tags

For ML governance purposes, original source run information is saved under the mlflow_export_import tag prefix.

For details see README_source_tags.

Setup

Supports python 3.7 or above.

Local setup

First create a virtual environment.

python -m venv mlflow-export-import
source mlflow-export-import/bin/activate

There are several different ways to install the package.

1. Install from PyPI - recommended

pip install mlflow-export-import

2. Install from github directly

pip install git+https:///github.com/mlflow/mlflow-export-import/#egg=mlflow-export-import

3. Install from github clone

git clone https://github.com/mlflow/mlflow-export-import
cd mlflow-export-import
pip install -e .

Databricks setup

There are two different ways to install the package.

1. Install package in notebook

Install notebook-scoped libraries with %pip.

pip install mlflow-export-import

2. Install package as a wheel on cluster

Build the wheel artifact, upload it to DBFS and then install it on your cluster.

git clone https://github.com/mlflow/mlflow-export-import
cd mlflow-export-import
python setup.py bdist_wheel
databricks fs cp dist/mlflow_export_import-1.0.0-py3-none-any.whl {MY_DBFS_PATH}

Databricks MLflow usage

To run the tools externally (from your laptop) against a Databricks tracking server (workspace) set the following environment variables.

export MLFLOW_TRACKING_URI=databricks
export DATABRICKS_HOST=https://mycompany.cloud.databricks.com
export DATABRICKS_TOKEN=MY_TOKEN

For full details see Access the MLflow tracking server from outside Databricks.

Running tools

The main tool scripts can be executed either as a standard Python script or console script.

Python console scripts (such as export-run, import-run, etc.) are provided as a convenience. For a list of scripts see setup.py.

This allows you to use:

export-experiment --help

instead of:

python -u -m mlflow_export_import.experiment.export_experiment --help

Testing

Two types of tests exist: open source and Databricks tests. See tests/README.

Workflow API

• README.md
• The WorkflowApiClient is a Python wrapper around the Databricks REST API to execute job runs in a synchronous polling manner.
• Although a generic tool, in terms of mlflow-export-import, its main use is for testing Databricks notebook jobs.