Kairos is a small Kubernetes controller that exists solely to restart pods based on a cron pattern annotation applied to the controlling resource (Deployment, DaemonSet or StatefulSet). It exists because of the long and storied tradition of restarting services on a regular basis because it's easier than fixing memory leaks, and who wants to wait around for pods to get OOMKilled?
To use, add the annotation kairos.erhudy.com/cron-pattern to your apps/v1 resource. Kairos accepts either 5- or 6-element patterns (with seconds), but if you really need to specify things down to the second, what are you even doing?
Kairos works in a similar manner to Reloader by adding or updating the annotation kairos.erhudy.com/cron-last-restarted-at inside the contained PodTemplateSpec, which will cause Kubernetes to generate a new ReplicaSet and turn all the pods. Kairos itself does not do anything with the pods directly. The kairos.erhudy.com/cron-last-restarted-at annotation is in RFC 3339 format and may be inspected to determine the last time the pod was restarted via Kairos's machinations.
Bear in mind that as with all pod cycles in Kubernetes, the restarts will not happen instantly, so ensure that you do not set a cron pattern so aggressive that you end up in unending ReplicaSet churn.
Kairos also accepts multiple cron patterns in a single annotation, separated by semicolons. Multiple independent restart jobs will be registered for that resource. Kairos does not check whether any of the specified cron patterns overlap or conflict.
Kairos starts up with the scheduler running on local time by default, determined through whatever mechanism Go uses to figure out the local timezone. To override that and specify a particular timezone, indicate it with the -timezone flag.
Timezones for particular jobs may be set by prefixing each cron pattern with TZ= or CRON_TZ=, e.g. TZ=America/New_York 5 12 * * *.
Use the Kustomize directory at deploy:
➜ kairos git:(main) kubectl apply -k deploy/install
serviceaccount/kairos created
clusterrole.rbac.authorization.k8s.io/kairos created
clusterrolebinding.rbac.authorization.k8s.io/kairos created
deployment.apps/kairos created
- proper release versioning and not just
:latest - more comprehensive test suite
- de-duplicate various code paths through unhealthy
reflectwitchcraft - add a feature to allow backfilling restarts if the controller was down or not running during a time when a cron pattern matched the current time, and the last time the pods were restarted can be determined by inspecting the
kairos.erhudy.com/cron-last-restarted-atannotation - Prometheus metrics
- better logging