With kubernetes 1.16 the cadvisor metric labels pod_name and container_name have been dropped (deprecated since 1.14) in favor of pod and container.
We have several grafana dashboards relying on those labels, like:
- https://grafana-rw.wikimedia.org/d/000000473/kubernetes-staging-pods
- Multiple service dashboards for container saturation data
Going forward we will probably have to fix all dashboards to use the new labels but we will have different versions running in production for some time which would mean dashboards only work for the new or the old clusters (and we probably want to compare things between them).
So I suggest we add a rewrite rule to prometheus to duplicate the pod_name and container_name labels into the new ones and patch all dashboards after we have ingested a reasonable amount of history.
Upstream issue: https://github.com/kubernetes/kubernetes/pull/80376