gc-controller

Pod garbage collector controller. This controller cleans evicted/failed pods and can keep a configurable number of them. Unlike the vanilla gc collector this controller is workload aware and collects evicted pods by namespaces and can keep a number of evicted pods for each governing workload resource. Despite this the vanilla one is configured by default to collect only if there are more than 12500 evicted pods --terminated-pod-gc-threshold. See https://kubernetes.io/docs/reference/command-line-tools-reference/kube-controller-manager/. That said this flag is usually not not even configurable on hosted kubernetes platforms since the control plane can't be modified.

Installation

Helm

Please see chart/gc-controller for the helm chart docs.

Kustomize

Alternatively you may get the bundled manifests in each release to deploy it using kustomize or use them directly.

Configure controller

You may change some settings using command line args. Note: by default the garbace collection keeps 2 (--keep=2) evicted pods by workload but deletes (--max-age=168h) any evicted pod older than 1 week.

--concurrent int                            The number of concurrent Pod reconciles. (default 4)
--enable-leader-election                    Enable leader election for controller manager. Enabling this will ensure there is only one active controller manager.
--graceful-shutdown-timeout duration        The duration given to the reconciler to finish before forcibly stopping. (default 10m0s)
--health-addr string                        The address the health endpoint binds to. (default ":9557")
--insecure-kubeconfig-exec                  Allow use of the user.exec section in kubeconfigs provided for remote apply.
--insecure-kubeconfig-tls                   Allow that kubeconfigs provided for remote apply can disable TLS verification.
--keep int                                  The number of pods to keep for each workload. (default 2)
--kube-api-burst int                        The maximum burst queries-per-second of requests sent to the Kubernetes API. (default 300)
--kube-api-qps float32                      The maximum queries-per-second of requests sent to the Kubernetes API. (default 50)
--leader-election-lease-duration duration   Interval at which non-leader candidates will wait to force acquire leadership (duration string). (default 35s)
--leader-election-release-on-cancel         Defines if the leader should step down voluntarily on controller manager shutdown. (default true)
--leader-election-renew-deadline duration   Duration that the leading controller manager will retry refreshing leadership before giving up (duration string). (default 30s)
--leader-election-retry-period duration     Duration the LeaderElector clients should wait between tries of actions (duration string). (default 5s)
--log-encoding string                       Log encoding format. Can be 'json' or 'console'. (default "json")
--log-level string                          Log verbosity level. Can be one of 'trace', 'debug', 'info', 'error'. (default "info")
--max-age duration                          The number of pods to keep for each workload. (default 168h0m0s)
--max-retry-delay duration                  The maximum amount of time for which an object being reconciled will have to wait before a retry. (default 15m0s)
--metrics-addr string                       The address the metric endpoint binds to. (default ":9556")
--min-retry-delay duration                  The minimum amount of time for which an object being reconciled will have to wait before a retry. (default 750ms)
--watch-all-namespaces                      Watch for resources in all namespaces, if set to false it will only watch the runtime namespace. (default true)
--watch-label-selector string               Watch for resources with matching labels e.g. 'sharding.fluxcd.io/shard=shard1'.