View Traces

After logging your traces, you can view them in the MLflow UI, under the "Traces" page in your experiment. The page is for browsing the list of traces with rich search and filtering capabilities, as well as inspecting the details of each trace.

Trace Table

The trace table you see first when you open the "Traces" page includes high-level information about the traces, such as the trace ID, the inputs / outputs of the root span, and more. The table displays the following columns.

Not Seeing All Columns?

Trace table does not display all fields by default. Some columns are hidden by default and can be enabled via the column selector.

Column	Description	Data Source
Trace ID	Unique identifier	Trace trace_id field
Request	Input to the root span	Derived from the root span's mlflow.spanInputs attribute, or customized via the request_preview parameter of the mlflow.update_current_trace() API
Response	Output of the root span	Derived from the root span's mlflow.spanOutputs attribute, or customized via the response_preview parameter of the mlflow.update_current_trace() API
Session	Conversation/session grouper	mlflow.trace.session metadata. See Track Users and Sessions for how to set this.
User	Application user who triggered the trace	mlflow.trace.user metadata. See Track Users and Sessions for how to set this.
Trace name	Name of the trace	Derived from the root span's name, or customized through the mlflow.traceName tag
Tokens	Aggregated input/output/total tokens	mlflow.trace.tokenUsage metadata. See Token Usage and Cost Tracking for more details.
Execution time	Total trace duration	Computed at runtime.
Request time	When the trace was created	Computed at runtime.
State	OK, ERROR, or IN_PROGRESS	Derived from the root span's execution status.
Run name	Associated MLflow run	If the trace was logged within an MLflow run context, the run name is automatically displayed here.
Prompt	Linked prompt versions	If a prompt in MLflow Prompt Registry is loaded during the trace scope, the prompt name and version are displayed here. See Linking Prompts with Traces for more details.
Source	The entry point or script that generated the trace. If it is from a Git repository, the field links to the repository URL with the commit hash and file path.	Derived from Git metadata at runtime if available.
Tags	User-defined key-value pairs	Trace tags (excluding mlflow.* system keys) if set.
Assessments	Feedback scores logged to the trace, either by a human or an automated evaluation.	See Collecting User Feedback and Evaluating Traces for how to log assessments.
Expectations	Ground truth values annotated to the trace.	See Ground Truth Expectations for how to log expectations.

Traces ingested via OTLP

Traces ingested through the OTLP endpoint go through automatic attribute translation to derive these columns from framework-specific OpenTelemetry attributes. MLflow supports translation for OpenTelemetry GenAI Semantic Conventions, and various frameworks including OpenLLMetry, OpenInference, Langfuse, and more. Please see the Attribute Mapping Reference for details on how each framework's attributes are mapped to these columns.

Browsing Single Trace

By clicking on a trace ID tr-... or the request link in the trace table, you can open the detailed trace view. The detailed trace view contains rich information about your AI agent or LLM application execution, such as:

• The span tree that represents the execution flow of the trace.
• Latency of each span and the overall timeline.
• Input and output of each span.
• Tool definitions and calls.
• Exception details if the execution failed.
• Token counts and cost breakdown.
• Feedback scores and ground truth annotated on the trace.

Image and Audio Content

MLflow renders images and audio inline in the trace viewer. For supported formats, framework examples, and how to attach multimodal content to traces, see Image and Audio (Multimodal) Content in Traces.

Performing Actions

From this page, you can also perform a few actions to manage your traces.

Searching Traces

Using the search bar in the UI, you can easily filter your traces based on name, tags, or other metadata. Check out the Searching Traces for details about the query string format.

Filtering Traces

The filter dropdown allows you to filter the traces table by various criteria, such as state, name, user, session, tags, feedback values, and more. See Filtering Traces for details.

Deleting Traces

The UI supports bulk deletion of traces. Simply select the traces you want to delete by checking the checkboxes, and then pressing the "Delete" button. See Delete Traces for how to delete traces programmatically.

tip

Shift+click to select all traces in a range quickly.

Editing Tags

You can also edit key-value tags on your traces via the UI. See Trace Tags for more details about trace tags.

Exporting Traces to Datasets

You can export the traces to Evaluation Datasets by clicking the "Add to evaluation dataset" action from the actions dropdown.

Jupyter Notebook integration

note

The MLflow Tracing Jupyter integration is available in MLflow 2.20 and above

You can also view the trace UI directly within Jupyter notebooks, allowing you to debug your applications without having to tab out of your development environment.

This feature requires using an MLflow Tracking Server, as this is where the UI assets are fetched from. To get started, simply ensure that the MLflow Tracking URI is set to your tracking server (e.g. mlflow.set_tracking_uri("http://localhost:5000")).

By default, the trace UI will automatically be displayed for the following events:

1. When the cell code generates a trace (e.g. via automatic tracing, or by running a manually traced function)
2. When mlflow.search_traces() is called
3. When a mlflow.entities.Trace() object is displayed (e.g. via IPython's display function, or when it is the last value returned in a cell)

To disable the display, simply call mlflow.tracing.disable_notebook_display(), and rerun the cell containing the UI. To enable it again, call mlflow.tracing.enable_notebook_display().