retrieving-mlflow-traces

Retrieving MLflow Traces

Bound Unity Catalog Searches by Time

When searching traces stored in Unity Catalog, include trace.timestamp_ms >= <start_time_ms> in the filter string, with an upper bound when known. This also applies when searching by experiment ID, run ID, session, or other filters: a small result limit does not prevent a scan across the full trace history. In Python, pass the time condition through filter_string; mlflow.search_traces() has no start_time argument.

Use the time window requested by the user or established by the run or incident. For setup verification, record the time immediately before the instrumented run. Do not silently restrict historical investigations to the last hour. If the relevant period cannot be determined, ask for it before starting a broad UC search. The examples below demonstrate individual filters; combine them with the relevant time bounds for UC.

Keep Setup Verification Brief

For verifying newly added instrumentation, follow the verification workflow in instrumenting-with-mlflow-tracing. Enforce a 60-second wall-clock deadline for readback, cancelling the verification command at the deadline even if a call is still retrying internally: make one fetch by a known trace ID, or one time-bounded search with --max-results 1 (Python: max_results=1). Retry at most once after a concrete, quick fix, within the same budget. Stop once one trace and its expected spans are confirmed.

If readback is slow or fails, report what was verified, what remains unverified, and the observed blocker and required user action (such as configuring a missing API key, logging in, obtaining permission, or resolving warehouse availability/capacity). Do not start parallel searches or fall back to direct SQL, raw REST, alternate authentication, or repeated agent runs just to complete setup verification. A dedicated retrieval or debugging request can require broader investigation; preserve the user's scope.

Single Fetch vs Search

Choose the right approach based on what you have:

You have...	Use	Command
Trace ID	Single fetch	mlflow traces get --trace-id <id>
Session/user/filters	Search	mlflow traces search --experiment-id <id> --filter-string "..."

Single fetch - Use when you have a specific trace ID (e.g., from UI, logs, or API response):

mlflow traces get --trace-id tr-69f72a3772570019f2f91b75b8b5ded9

Search - Use when you need to find traces by criteria (session, user, status, time range, etc.):

mlflow traces search --experiment-id 1 --filter-string "metadata.\`mlflow.trace.session\` = 'session_abc'"

---

Trace Data Structure

• TraceInfo: trace_id, status (OK/ERROR), timestamp_ms, execution_time_ms, tags, metadata, assessments (human feedback, evaluation results)
• Spans: Tree of operations with name, type, attributes, start_time, end_time

Workflow

1. Check CLI usage (required): mlflow traces search --help
2. Build filter query using syntax below
3. Execute search with appropriate flags
4. Retrieve details for specific traces if needed

Prerequisite: Check CLI Usage

mlflow traces search --help

Always run this first to get accurate flags for the installed MLflow version.

Searching Traces

The mlflow traces search command is used to search for traces in an MLflow experiment.

By Run ID

Filter traces associated with a specific MLflow run:

# All traces for a run
mlflow traces search --run-id <run_id>

# Failed traces for a run
mlflow traces search --run-id <run_id> --filter-string "trace.status = 'ERROR'"

# Can combine with experiment-id
mlflow traces search --experiment-id 1 --run-id <run_id>

By Session or User (Common for Debugging)

When debugging an issue from the MLflow UI, filter by session or user ID to get all related traces:

# All traces for a specific session (use backticks for special characters in key)
mlflow traces search --experiment-id 1 --filter-string "metadata.\`mlflow.trace.session\` = 'session_abc123'"

# All traces for a specific user
mlflow traces search --experiment-id 1 --filter-string "metadata.\`mlflow.trace.user\` = 'user_456'"

# Failed traces in a session (for root cause analysis)
mlflow traces search --experiment-id 1 --filter-string "metadata.\`mlflow.trace.session\` = 'session_abc123' AND trace.status = 'ERROR'"

# Session traces ordered by time (to see sequence of events)
mlflow traces search --experiment-id 1 --filter-string "metadata.\`mlflow.trace.session\` = 'session_abc123'" --order-by "timestamp_ms ASC"

By Status

mlflow traces search --experiment-id 1 --filter-string "trace.status = 'ERROR'"
mlflow traces search --experiment-id 1 --filter-string "trace.status = 'OK'"

By Time Range

# Timestamps are in milliseconds since epoch
# Get current time in ms: $(date +%s)000
# Last hour: $(( $(date +%s)000 - 3600000 ))

mlflow traces search --experiment-id 1 --filter-string "trace.timestamp_ms > $(( $(date +%s)000 - 3600000 ))"

By Execution Time (Slow Traces)

# Traces slower than 1 second
mlflow traces search --experiment-id 1 --filter-string "trace.execution_time_ms > 1000"

By Tags and Metadata

# By tag
mlflow traces search --experiment-id 1 --filter-string "tag.environment = 'production'"

# By metadata
mlflow traces search --experiment-id 1 --filter-string "metadata.user_id = 'user_123'"

# Escape special characters in key names with backticks
mlflow traces search --experiment-id 1 --filter-string "tag.\`model-name\` = 'gpt-4'"
mlflow traces search --experiment-id 1 --filter-string "metadata.\`user.id\` = 'abc'"

By Assessment/Feedback

mlflow traces search --experiment-id 1 --filter-string "feedback.rating = 'positive'"

Full Text Search

mlflow traces search --experiment-id 1 --filter-string "trace.text LIKE '%error%'"

Pagination

Control result count and iterate through pages:

# Limit results per page
mlflow traces search --experiment-id 1 --max-results 50

# Output includes "Next page token: <token>" if more results exist
# Use --page-token to fetch next page
mlflow traces search --experiment-id 1 --max-results 50 --page-token "eyJvZmZzZXQiOiA1MH0="

Output Options

# Output format (table or json)
mlflow traces search --experiment-id 1 --output json

# Include span details in output
mlflow traces search --experiment-id 1 --include-spans

# Order results
mlflow traces search --experiment-id 1 --order-by "timestamp_ms DESC"

Retrieving Single Trace

When you need to retrieve details about a specific trace, use the mlflow traces get command.

mlflow traces get --trace-id <trace_id>

Filter Syntax

For detailed syntax, fetch from documentation:

WebFetch(
  url: "https://mlflow.org/docs/latest/genai/tracing/search-traces.md",
  prompt: "Extract the filter syntax table showing supported fields, operators, and examples."
)

Common filters:

• trace.status: OK, ERROR, IN_PROGRESS
• trace.execution_time_ms, trace.timestamp_ms: numeric comparison
• metadata.\mlflow.trace.session`, metadata.`mlflow.trace.user``: session/user filtering
• tag.<key>, metadata.<key>: exact match or pattern
• span.name, span.type: exact match or pattern
• feedback.<name>, expectation.<name>: assessments

Pattern operators: LIKE, ILIKE (case-insensitive), RLIKE (regex)

Python API

For mlflow.search_traces(), see: https://mlflow.org/docs/latest/genai/tracing/search-traces.md