Databricks Managed MLFlow with Different Unity Catalog for Multi-tenant Production Tracing

jeffreym9 — Thu, 20 Mar 2025 23:55:20 GMT

Is the Databricks Managed MLFlow only trace LLM traffic through Serving Endpoint? Does it support manual tracing in my LLM application with decorator @mlflow.trace ?

Also, How can Databricks Managed MLFlow support multi-tenant cases where traces need to write to tenant specific unity catalog?

Re: Databricks Managed MLFlow with Different Unity Catalog for Multi-tenant Production Tracing

stbjelcevic — Mon, 10 Nov 2025 20:03:31 GMT

Hi @jeffreym9 ,

Managed MLflow Tracing captures traces in notebooks, local IDEs, jobs, and Serving Endpoints—via autologging or manual instrumentation, so it’s not limited to Serving Endpoints-only. Set your tracking URI to databricks and traces are written to the managed tracking service; when deploying agents, tracing is automatic, and you can also copy traces to Delta tables for analysis.
Yes, manual tracing with @mlflow.trace is supported: Use the decorator to create spans for any function (for example, agent, retriever, parser) and combine it with autologging for LLM SDKs.
Multi‑tenant writing to tenant‑specific Unity Catalog: In production, enable MLflow “Production monitoring” to sync traces to a Delta table and select the UC schema per endpoint or app; combine this with UC isolation patterns (catalog‑per‑tenant or schema‑per‑tenant) to route each tenant’s traces to their governed storage.

Sources:

topic Databricks Managed MLFlow with Different Unity Catalog for Multi-tenant Production Tracing in Administration & Architecture

Databricks Managed MLFlow with Different Unity Catalog for Multi-tenant Production Tracing

Re: Databricks Managed MLFlow with Different Unity Catalog for Multi-tenant Production Tracing