kubernetes-sigs / inference-perf / 39681093176

84%

Build Type

push

github

Committed by web-flow

Commit Message

Make the runtime metrics endpoint opt-in (#884)

Part of #489

#740 served the runtime metrics endpoint by default and set `host:
0.0.0.0` in the root `config.yml` and the Helm chart values for pod-IP
scraping. The root `config.yml` is also what the README quickstart, the
Docker image's default command and the deploy/ ConfigMap run, so a
laptop run from the README bound all interfaces.

The endpoint is now off unless asked for. `observability.metrics` is an
optional block with a single `port`. Setting it, or passing
`--observability.metrics.port`, serves `/metrics` on every interface.
Leaving it out serves nothing. The chart and the deploy/ Job turn it on,
the root `config.yml` does not. `enabled` and `host` are gone. Nothing
shipped since #740, so there is no deprecation path.

Changes:
* `ObservabilityConfig.metrics` is `Optional[RuntimeMetricsConfig]`,
default `None`. `RuntimeMetricsConfig` keeps only `port`.
* `main.py` starts the endpoint only when the block is set, always on
`0.0.0.0`. A bind failure is always an error.
* `deploy/inference-perf/values.yaml` sets `observability.metrics.port:
9464`. `deploy/manifests.yaml` passes `--observability.metrics.port
9464` to the Job.
* Drop the `observability` block from `config.yml`.
* Unit tests pin the config shape: no block, an empty block, the flag
alone, and both deploy surfaces turning it on.
* `e2e/tests/test_runtime_metrics_endpoint.py` scrapes /metrics during a
mock run. The YAML block or the flag serves it on loopback and the LAN
address. No block serves nothing.
* Update docs/config.md and regenerate docs/cli_flags.md.

```yaml
observability:
  metrics:
    port: 9464
```

Coverage Stats

3 of 8 new or added lines in 2 files covered. (37.5%)

1 existing line in 1 file now uncovered.

12375 of 14785 relevant lines covered (83.7%)

0.84 hits per line