MLflow Metrics
Run scripts/fetch_metrics.py to query metrics from an MLflow tracking server.
Examples
Token usage summary:
python scripts/fetch_metrics.py -s http://localhost:5000 -x 1 -m total_tokens -a SUM,AVG
Output: AVG: 223.91 SUM: 7613
Hourly token trend (last 24h):
python scripts/fetch_metrics.py -s http://localhost:5000 -x 1 -m total_tokens -a SUM \
-t 3600 --start-time="-24h" --end-time=now
Output: Time-bucketed token sums per hour
Latency percentiles by trace:
python scripts/fetch_metrics.py -s http://localhost:5000 -x 1 -m latency -a AVG,P95 -d trace_name
Error rate by status:
python scripts/fetch_metrics.py -s http://localhost:5000 -x 1 -m trace_count -a COUNT -d trace_status
Quality scores by evaluator (assessments):
python scripts/fetch_metrics.py -s http://localhost:5000 -x 1 -v ASSESSMENTS \
-m assessment_value -a AVG,P50 -d assessment_name
Output: Average and median scores for each evaluator (e.g., correctness, relevance)
Assessment count by name:
python scripts/fetch_metrics.py -s http://localhost:5000 -x 1 -v ASSESSMENTS \
-m assessment_count -a COUNT -d assessment_name
JSON output: Add -o json to any command.
Arguments
| Arg | Required | Description |
|---|---|---|
-s, --server | Yes | MLflow server URL |
-x, --experiment-ids | Yes | Experiment IDs (comma-separated) |
-m, --metric | Yes | trace_count, latency, input_tokens, output_tokens, total_tokens |
-a, --aggregations | Yes | COUNT, SUM, AVG, MIN, MAX, P50, P95, P99 |
-d, --dimensions | No | Group by: trace_name, trace_status |
-t, --time-interval | No | Bucket size in seconds (3600=hourly, 86400=daily) |
--start-time | No | -24h, -7d, now, ISO 8601, or epoch ms |
--end-time | No | Same formats as start-time |
-o, --output | No | table (default) or json |
For SPANS metrics (span_count, latency), add -v SPANS.
For ASSESSMENTS metrics, add -v ASSESSMENTS.
See references/api_reference.md for filter syntax and full API details.