NVIDIA / nodewright / 38655426437

84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(operator): add an opt-in pprof endpoint (PPROF_PORT) (#795)

When the operator's memory grows on a large cluster, go_memstats says how
much is allocated and how often the GC runs, but nothing on the manager
says which allocations. controller-runtime already serves net/http/pprof
when PprofBindAddress is set; expose it through PPROF_PORT, empty by
default so existing deployments are unchanged. Set to 127.0.0.1:6060 the
listener stays off the pod network and kubectl port-forward reaches it.

The knob ships with its companions: a "Profiling the operator" section in
docs/operations/resources-at-scale.md (loopback-only binding, port-forward
and go tool pprof, the endpoint is unauthenticated, changing it restarts
the manager), controllerManager.manager.env.pprofPort in the chart (the
env var is only emitted when set, so rendered output is unchanged for
existing installs), and the PPROF_PORT mirror in operator/config.

Signed-off-by: Taylor Daugherty <tdaugherty@nvidia.com>

Coverage Stats

1 of 1 new or added line in 1 file covered. (100.0%)

15 existing lines in 3 files now uncovered.

10040 of 12039 relevant lines covered (83.4%)

7.62 hits per line