Instrumenting With Mlflow Tracing

MLflow Tracing Instrumentation Guide

Language-Specific Guides

Based on the user's project, load the appropriate guide:

• Python projects: Read references/python.md
• TypeScript/JavaScript projects: Read references/typescript.md

If unclear, check for package.json (TypeScript) or requirements.txt/pyproject.toml (Python) in the project.

What to Trace

Trace these operations (high debugging/observability value):

Skip tracing these (too granular, adds noise):

• Simple data transformations (dict/list manipulation)
• String formatting, parsing, validation
• Configuration loading, environment setup
• Logging or metric emission
• Pure utility functions (math, sorting, filtering)

Rule of thumb: Trace operations that are important for debugging and identifying issues in your application.

Verification

After instrumenting the code, always verify that tracing is working.

Planning to evaluate your agent? Tracing must be working before you run agent-evaluation. Complete verification below first.

1. Run the instrumented code — execute the application or agent so that at least one traced operation fires
2. Confirm traces are logged — use mlflow.search_traces() or MlflowClient().search_traces() to check that traces appear in the experiment. If the trace is not found, try mlflow.flush_trace_async_logging() to flush the background queue.

import mlflow

mlflow.flush_trace_async_logging()
traces = mlflow.search_traces(experiment_ids=["<experiment_id>"])
print(f"Found {len(traces)} trace(s)")
assert len(traces) > 0, "No traces were logged — check tracking URI and experiment settings"

3. Verify spans were captured — confirm the trace contains the expected spans, not just an empty shell:

trace = traces.iloc[0]
spans = mlflow.get_trace(trace.trace_id).data.spans
print(f"Trace has {len(spans)} span(s)")
for span in spans:
    print(f"  - {span.name} ({span.span_type})")

4. Report the result — tell the user how many traces and spans were found and confirm tracing is working

If no traces appear

Check these in order:

• Verification ran before traces were exported — trace logging is asynchronous by default, so an in-process search_traces() right after the run can return zero before the background queue flushes (up to a few seconds later). Call mlflow.flush_trace_async_logging() before searching, as shown above.
• Tracking URI not set — is mlflow.set_tracking_uri(...) called before the agent run? Without this, traces go to a local ./mlruns directory instead of the configured server.
• Autolog warnings — did mlflow.autolog() or framework-specific mlflow.<framework>.autolog() raise any warnings during setup? Check stderr for patching failures.
• Wrong experiment ID — verify the experiment ID passed to search_traces() matches the experiment active when the code ran (mlflow.get_experiment_by_name(...) to confirm).
• Network/auth issues — can the process reach the tracking server? Check for connection errors or 401/403 responses in logs.

For automated validation, use agent-evaluation/scripts/validate_tracing_runtime.py.

Feedback Collection

Log user feedback on traces for evaluation, debugging, and fine-tuning. Essential for identifying quality issues in production.

See references/feedback-collection.md for:

• Recording user ratings and comments with mlflow.log_feedback()
• Capturing trace IDs to return to clients
• LLM-as-judge automated evaluation

Reference Documentation

Production Deployment

See references/production.md for:

• Environment variable configuration
• Async logging for low-latency applications
• Sampling configuration (MLFLOW_TRACE_SAMPLING_RATIO)
• Lightweight SDK (mlflow-tracing)
• Docker/Kubernetes deployment

Advanced Patterns

See references/advanced-patterns.md for:

• Async function tracing
• Multi-threading with context propagation
• PII redaction with span processors

Distributed Tracing

See references/distributed-tracing.md for:

• Propagating trace context across services
• Client/server header APIs

Databricks (Unity Catalog storage)

See references/databricks.md for storing traces in Unity Catalog Delta tables by binding an experiment to a UnityCatalog trace location (catalog, schema, table prefix).

Next: debug from the traces you just captured

Tracing is now in place. When you move on to debug or improve the agent's behavior, read the spans first. Do not fall back to reading source code and output files alone. The trace shows what each step actually received, produced, and decided, which is the evidence source that pins down where behavior went wrong.

Load the fix-agent-issue skill for this. It grounds the diagnosis in the trace, what the agent did, what it should have done, and why, before any code change, and codifies the fix as a regression test so it sticks. Reach for it as soon as you start asking why the agent produced a given output, not only when someone explicitly reports a bug.