Build a custom trace view

Note

This feature is in Beta.

A custom trace view is a personalized layout in the Azure Databricks MLflow trace explorer. Instead of the default tabs, it displays the trace fields, metrics, and feedback controls that matter most for your review workflow.

Custom views are AI-assisted: you describe the layout you want in plain language, and Genie generates it for you. No coding is required. Each view is a trace-agnostic template that re-binds automatically to every trace in the experiment, so you build a layout once and reuse it across all of your traces.

Custom views are useful when you want to:

  • Surface a small set of high-signal fields for a focused review, rather than scanning the full span tree.
  • Present the same layout to every reviewer working in an experiment.
  • Add interactive feedback controls so reviewers can score traces directly while they read them.

Create a custom view

  1. Open a trace from the Traces tab. For details on finding and opening traces, see View traces in the Databricks MLflow UI.
  2. Select the Custom view tab in the trace explorer.
  3. In the text box, describe the layout you want in plain language. For example, describe which trace fields to display, which metrics to summarize, and which feedback controls to include.
  4. Click Build with Genie. Genie generates a layout from your description. To refine it, click Edit with Genie and describe the changes you want.
  5. Click Save. MLflow stores saved views with the experiment, and they become available for every trace in that experiment.

To create additional views, open the view switcher and select Create view.

Display components

When you describe a layout, Genie assembles it from a set of display components:

  • Stat card: Displays a single metric or value, such as latency, token count, or a computed score.
  • Assessment board: Displays the assessments logged for a trace, including feedback and evaluation results.
  • Key-value viewer: Displays structured fields from the trace as key-value pairs.
  • Layout primitives: Arrange the components into rows, columns, and sections.

Add feedback controls

You can add interactive controls to a custom view so that reviewers can score and comment on traces directly in the layout. Available controls include:

  • Thumbs up and thumbs down buttons for quick approval or rejection.
  • Radio groups for selecting one option from a set.
  • Text inputs for free-form comments or rationale.

MLflow logs this feedback to the trace as an assessment, so it appears alongside other assessments and is available for evaluation and monitoring.

Manage saved views

Views are shared at the experiment level, so anyone with permission to view the experiment sees the same set of custom views. MLflow stores views as experiment tags.

From the trace explorer, you can:

  • Switch views
  • Rename a view
  • Delete a view

Because views are trace-agnostic templates, switching views changes only the layout. The underlying trace data stays the same, and the selected view re-binds to whichever trace you open next.

Permissions

Access to custom views follows experiment permissions:

  • CAN EDIT or CAN MANAGE: Create, rename, and delete custom views.
  • CAN READ: View and switch between saved views, but not create, rename, or delete them.

Limitations

  • Each experiment can store a maximum of 50 saved views.
  • Each view has a limit of approximately 20,000 UTF-8 bytes. Larger views are compressed to fit within the limit.
  • A view whose stored definition is corrupted appears as a placeholder until you rebuild it.

Additional resources

  • Last updated on