MLflow Tracing

MLflow Tracing is a Databricks ai / ml capability, introduced June 2024.

The observability layer for GenAI apps and agents on Databricks: it records the inputs, outputs, latency, token usage, and cost of every intermediate step of a run - LLM calls, retrievers, tools, sub-agents - and shows them in the MLflow trace explorer, with traces kept in an MLflow experiment or, since 2026, as OpenTelemetry rows in Unity Catalog tables.

• The docs page for watching what your agent does opens with a "Coding agent quickstart" - "Using a coding agent? Paste this prompt" - so the first recommended reader of the tracing documentation is the agent you are about to trace. 🔗
• Keep traces in an MLflow experiment and the backend caps you at 100,000 traces per experiment; point the experiment at a Unity Catalog table instead and the same row of the comparison table just says Unlimited. 🔗
• Since August 2026 the trace explorer can be redecorated by Genie - you describe the layout you want in plain language and it generates a reusable view with no coding required, which is an AI building the dashboard you use to check on your AI. 🔗

⚠ Limitations: For traces stored in Unity Catalog: ingestion is initially limited to 200 traces per second per workspace and 100 MB per second per table (a raisable default - ask your account team for more); an experiment can be bound to a Unity Catalog trace location only when it is created; Unity Catalog traces are not supported with Knowledge Assistant or Supervisor Agent; individual traces cannot be deleted (delete rows from the underlying tables with SQL instead); and traces cannot yet be written to a default storage catalog or to storage protected by Private Link. Traces stored in an experiment instead are hard-capped at 100,000 per experiment. 🔗

Category

AI / ML

Introduced

June 2024🔗

Also known as

Tracing, MLflow traces, trace explorer, unified trace table

Verified

2026-09-20