migrate-from-model-serving

Migrating a Databricks Model Serving agent to Databricks Apps requires downloading model artifacts, extracting code, transforming class-based ResponsesAgent entry points into @invoke and @stream decorated functions, and deploying via Databricks Asset Bundles with correct serving and UC resource permissions.

The best way to preserve streaming behavior is converting the class-based predict_stream method into an @stream decorated function following MLflow GenAI Server conventions, keeping synchronous execution for minimal changes or switching to asynchronous for higher concurrency.

Migrating to Databricks Apps supports stateful memory by configuring Lakebase integrations, preserving short-term memory patterns like thread_id and checkpointer alongside long-term memory patterns like user_id and store.

Converting a class-based MLflow ResponsesAgent endpoint into decorated functions requires extracting original model code and resources, then transforming predict and predict_stream methods into @invoke and @stream decorated functions compatible with MLflow GenAI Server conventions.

You can switch your Databricks agent from synchronous to asynchronous execution during migration, keeping code synchronous for minimal changes or converting I/O operations to async execution to achieve higher concurrency in the new Databricks Apps environment.

Configuring Databricks Asset Bundles for a migrated agent requires producing the expected directory layout and updating bundle configuration with resources extracted from the original model, validating authentication, and applying correct Unity Catalog resource permissions for serving.