Report what each service spent and consumed as records

[markbackman]

Summary

• Adds ServiceMetricsObserver, which turns each metric a service publishes into a record. Today MetricsLogObserver writes them out for a person reading a terminal; nothing turned them into data.
• on_service_latency → ServiceLatencyRecord: kind (ttfb / ttfa / ttfat), processor, model, timestamp, seconds. A measurement that decomposes keeps its parts — leading_silence_secs behind a time to first audio, thinking_time_secs behind a time to first answer token, and the ttfb_secs both build on.
• on_service_usage → ServiceUsageRecord: kind (stt / llm / tts), processor, model, timestamp, plus audio_seconds, characters, and every token count LLMTokenUsage reports including cache reads, reasoning and audio tokens.
• Nothing is summed. A turn that runs two inferences reports twice, so a consumer groups the records by turn, session or model as it needs — and a session that ends abruptly leaves behind everything up to that point, where a total held in memory goes with the process holding it.
• What a service made someone wait for is here; what it did with its own time is not. Processing time, text aggregation and smart-turn predictions are all deliberately absent, with a test covering it: aggregation already appears as a span in LatencyBreakdown, and the other two describe how work was done rather than what it cost the person waiting.

observer = ServiceMetricsObserver()

@observer.event_handler("on_service_usage")
async def on_service_usage(observer, record):
    logger.info(record.model_dump_json())

The observer reads a time source a test can supply, and remembers a bounded number of frame IDs so a metric relayed along the pipeline is reported once rather than once per hop.

Testing

• uv run pytest tests/test_service_metrics_observer.py — 12 tests, 100% line coverage of the new file: every metric type mapped field by field, the optional LLM token counts, several metrics in one frame, a relayed frame reported once, the two omitted metric types passed over, and the frame memory staying bounded.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/pipecat/observers/service_metrics_observer.py	100.00% <100.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[aconchillo]

elif

[aconchillo]

elif

[aconchillo]

remove

[aconchillo]

sorry, keep return None

[aconchillo]

same in this function

[aconchillo]

LGTM. Approving, just minor comments.