ctrlqueue

Overview ¶

Package ctrlqueue builds controller-runtime workqueues that keep no metrics, for components that do not export controller-runtime's workqueue_* series.

Why (issue #1131): every controller-runtime priority queue built with a name starts an updateUnfinishedWorkLoop goroutine that wakes every 500 ms for the life of the process to refresh two Prometheus gauges. A node agent runs one queue per controller (gamma, l4route ×3, capture, endpointpolicy, node taint, ...), so that is ~10 timer wakeups a second on an otherwise idle process, and on a node where the Go runtime's threads land on deep-idle CPUs each one is expensive. Nothing scrapes the agent's workqueue_* series (the chart exposes :8080 but no scrape job, dashboard or alert reads it; the agent's own metrics go out over OTLP), so the loops are pure overhead.

There is no manager-wide switch for this. The queue's metrics provider is not a manager option, and a no-op provider alone does not help: the priority queue starts the loop whenever its name is non-empty, whatever the provider. Only a queue with an empty name gets the no-metrics implementation and no loop, so NewQueue builds exactly that, and also passes NoopMetricsProvider so a future controller-runtime that keys the loop off the provider stays quiet too.