lakebase-provisioned

Use the Databricks SDK to generate an OAuth credential, then pass it as the password in a psycopg3 connection string with the instance's read_write_dns host and sslmode=require. Tokens expire after one hour, so regenerate them for long-running processes.

Run a background task that refreshes the token every 50 minutes and inject it into new connections via SQLAlchemy's do_connect event listener. This keeps a connection pool working without interruption despite the one-hour token lifetime.

FULL sync replaces the entire target PostgreSQL table on each run and suits small to medium source tables. INCREMENTAL sync only transfers changed rows using a tracking column, which is better for large Delta tables with reliable change timestamps.

Python's socket.getaddrinfo can fail on macOS with long Lakebase hostnames. Resolve the hostname with the dig command and pass the resulting IP as hostaddr to psycopg while keeping the hostname for TLS SNI.

Yes. Add Lakebase as an app resource with the Databricks CLI, configure it via LAKEBASE_INSTANCE_NAME and LAKEBASE_DATABASE_NAME environment variables, and declare it as an MLflow model resource. The databricks-langchain memory package supports Lakebase-backed agent state.