databricks / mlops-stacks

This repo provides a customizable stack for starting new ML projects on Databricks that follow production best-practices out of the box.

https://docs.databricks.com/en/dev-tools/bundles/mlops-stacks.html

Apache License 2.0

456 stars 156 forks source link

Models uc update #106

Closed niall-turbitt closed 1 year ago

niall-turbitt commented 1 year ago

Models in UC addition to MLOps Stack

Additional prompts added for project instantiation:
- include_models_in_unity_catalog: Whether to use Models in UC for model registration. Creates resources using DABs. Creates models under the catalog, mapped to the target environment. For example, if the bundle is deployed to prod, the model will be registered to the prod catalog. We assumed there exists a catalog mapped to each of these. Given that test is also a defined target, this assumes the existence of a test catalog if deploying to the test environment
- schema_name: Name of the schema under each catalog to register models to. Assumes the same schema name for each catalog. Defaults to "schema_name"
- unity_catalog_read_user_group: Used to define the grants block assigned to the registered model in UC. Default to "account users"
Notably, changes to accommodate models in UC were made to:
- Runtime - Bumped from DBR 12.2 ML to 13.3 ML
- Training - add model signature, set the registry URI to UC if using models in UC
- Model validation - assign "challenger" alias if all validation checks pass
- Model deployment - update model alias to "champion", remove "challenger" alias
- Batch inference - perform inference against alias, according to environment in which inference is running. e.g. "challenger" when in prod environment, conda env updated to virtualenv

NOTE:

There are currently NO staged doc updates in this PR - to be added
Models in UC is not compatible with MLflow Recipes, and as such this functionality has NOT been added
- Functionality for models in UC has not been added for the Feature Store version. This should be possible, but I did not get the chance to try this out

Testing Example AWS repo with Feature Store and models in UC: https://github.com/databricks/example-aws-github-project-fs-uc

Example AWS repo with no Feature Store and models in UC: https://github.com/databricks/example-aws-github-project-uc

mingyu89 commented 1 year ago

Amazing. This is such a huge change. Great work and many thanks.

niall-turbitt commented 1 year ago

@arpitjasa-db made some changes to get Feature Store working with models in UC

Example AWS repo with Feature Store and models in UC: https://github.com/databricks/example-aws-github-project-fs-uc

Example AWS repo with no Feature Store and models in UC: https://github.com/databricks/example-aws-github-project-uc