https://prow.ci.openshift.org/view/gcs/origin-ci-test/logs/release-openshift-ocp-installer-e2e-aws-serial-4.8/1371784347064995840

: [bz-Image Registry] clusteroperator/image-registry should not change condition/Available expand_less

2 unexpected clusteroperator state transitions during e2e test run 

image-registry was Available=true, but became Available=false at 2021-03-16 12:42:10.168205263 +0000 UTC -- Available: The deployment does not have available replicas
NodeCADaemonAvailable: The daemon set node-ca has available replicas
ImagePrunerAvailable: Pruner CronJob has been created
image-registry was Available=false, but became Available=true at 2021-03-16 12:42:19.916265445 +0000 UTC -- Available: The registry has minimum availability
NodeCADaemonAvailable: The daemon set node-ca has available replicas
ImagePrunerAvailable: Pruner CronJob has been created

Occuring in 1/8 runs

The serial test adds and removes machines gracefully (creates 3 new machines, delete 3 old ones), it's likely the operator is detecting the drain of the image registry and acting too aggressively. Adding and removing worker nodes is normal and the operator should not go unavailable unless all instances are down. If they are, this bug is urgent and needs to be fixed ASAP, because a PDB should prevent that).  A single instance down is not "unavailable" (we run 2 for that reason).  We have alerts that catch when a scrape target is down for longer than normal, that would fire in this case after 15m so the operator does not need to report unavailable if one is down (if it's during an upgrade and we wedge we should be degraded, if it's a normal ingress controller rollout we should also be degraded after some period of time).