MLflow Tracing - GenAI observability
MLflow Tracing gives you end-to-end observability into GenAI applications, including complex agent-based systems. It records inputs, outputs, intermediate steps, and metadata so you can see exactly how your app behaves.
Tracing allows you to:
- Debug and understand your application
- Monitor performance and optimize cost
- Monitor production applications
- Evaluate and improve high-quality application performance
- Verify auditability and compliance
- Integrate tracing with many popular third-party frameworks
- Use natural language with Genie Code to analyze, debug, and explore trace data
Where traces are stored
You have two options to store traces. Databricks recommends storing traces in Unity Catalog for new and production workloads. Regardless of the option you select to store traces, MLflow experiments are the UI entry point for viewing traces:
- Traces stored in Unity Catalog (recommended): Point your experiment at a Unity Catalog trace location so traces land in OpenTelemetry (OTel) Delta tables. There is no storage cap, traces are queryable with SQL like any Delta table, and Unity Catalog governs access.
- Traces stored in an experiment: MLflow stores traces in the experiment itself, but MLflow caps storage at 100,000 traces per experiment and this option might not be compatible with the latest MLflow features.
Capability | Traces stored in Unity Catalog | Traces stored in an experiment |
|---|---|---|
Storage limit | Unlimited | 100,000 traces per experiment |
Query options | MLflow UI and Python SDK, plus SQL, Genie, AI/BI dashboards, and any Spark-based tool | MLflow UI and Python SDK |
Governance | Unity Catalog schema and table permissions | Experiment-level access controls |
OpenTelemetry compatibility | Traces stored in OTel format, compatible with other OTel clients and tools | Not supported |
Additional resources
- 10-minute tracing demo - Get started with tracing in Databricks.
- Instrument your app - Select between automatic and manual tracing approaches
- Genie Code for agent observability and evaluation - Analyze traces, troubleshoot errors, and explore evaluation data using natural language.