migrate-from-model-serving

To migrate an MLflow ResponsesAgent to Databricks Apps, download the original serving artifacts, refactor predict() and predict_stream() methods into @invoke and @stream decorated functions, update databricks.yml resources, and deploy using Databricks Asset Bundles.

You can keep synchronous logic with minimal changes during migration to Databricks Apps, or choose a fully async code path for better concurrency depending on your agent's original implementation.

Migrating a stateful agent requires handling checkpointer and store migration patterns using Lakebase, and mapping custom_inputs fields like thread_id or user_id to preserve conversation memory in the new app.

Yes, migration requires updating databricks.yml using resources parsed from the original MLmodel file to configure the deployment bundle before validating, deploying, and running the migrated app.

During migration, you must download the model artifacts, analyze the agent code, and copy the /code directory and associated artifacts from the original MLflow model bundle into the new Databricks Apps implementation.

Yes, valid Databricks CLI authentication is required, along with specifying your CLI profile, target app name, and async migration preference to successfully deploy and run the migrated app.