Skip to content

metrics.md

Latest commit

 

History

History
49 lines (38 loc) · 2.99 KB

metrics.md

File metadata and controls

49 lines (38 loc) · 2.99 KB
Raw

Metrics

The Inference Payload Processor (IPP) exposes Prometheus metrics for build information, request processing, per-plugin latency, and model selection. They are served on a dedicated endpoint (default port 9090, path /metrics).

All metric names carry the ipp_ prefix (the ipp Prometheus subsystem). Every metric is published at the ALPHA stability level, reflected as an [ALPHA] annotation at the start of each metric's help text.

Metrics Reference

Name Type Labels Description
ipp_info Gauge commit, build_ref Build information of the running IPP.
ipp_success_total Counter (none) Requests processed successfully.
ipp_body_field_not_found_total Counter field Times a field was not found in a request body.
ipp_body_field_empty_total Counter field Times a field was found in a request body but was empty.
ipp_plugin_duration_seconds Histogram extension_point, plugin_type, plugin_name Per-plugin processing latency, by pipeline extension point.
ipp_model_selector_e2e_duration_seconds Histogram (none) End-to-end model selection latency.
ipp_model_selector_attempt_total Counter status Model-selection attempts, by status (success / failure).
ipp_request_ttft_seconds Histogram model Time to first token, per model.

Notes:

  • ipp_info is emitted once at startup with a constant value of 1; the build identity lives in its labels. Join on it to attribute other series to a build.
  • ipp_body_field_empty_total distinguishes a present-but-empty field from a missing one (ipp_body_field_not_found_total).
  • status="failure" on ipp_model_selector_attempt_total covers cases such as filtering leaving zero candidate models.
  • Histogram buckets: ipp_plugin_duration_seconds and ipp_model_selector_e2e_duration_seconds span 0.0001s–0.1s (sub-second plugin work); ipp_request_ttft_seconds spans 0.05s–30s (end-to-end serving latency).

Scraping

Metrics are served by the controller-runtime metrics server on --metrics-port (default 9090) at /metrics. Access is governed by --metrics-endpoint-auth, enabled by default: when on, scrapers must present Kubernetes credentials authorized to read the endpoint; set it to false to serve without authentication. See Configuration for how these flags are wired through the Helm chart.

References

  • Configuration — Metrics serving and CLI flags.
  • Architecture — ext-proc integration, the processing pipeline, and model selection.
  • metrics.go — Authoritative metric definitions in source.