+++ This bug was initially created as a clone of Bug #1925276 +++

Creating this bug for tracking as original bug is labeled as private.

We've seen this issue before on another platform.  I'm not entirely sure that the cache is the problem in this particular case, though it may be a contributing factor.  I think what is happening is that when we patch the object (in this case, the status object for the 'phase') that is queuing up another reconcile.  This reconcile contains otherwise stale data about the machine.  This compounds with the fact that AWS's API is eventually consistent; since the AWS API isn't up to date yet, we search by tags and get no instance.  If the machine-object wasn't stale, we'd look up the instance-id directly and requeue with an error if we didn't find the instance in question: https://github.com/openshift/cluster-api-provider-aws/blob/master/pkg/actuators/machine/reconciler.go#L242

This generally only happens when you scale up a single machineset by one.  If there are 2 or more machines being created at once, that seems to be enough time for the cache to catch back up.  The work around until a patch is shipped is to keep an eye on pending/unapproved CSRs.  Since a machine can only have one associated node, the extra instance will not be able to automatically join the cluster.  Unfortunately, that instance will need to be deleted via the cloud provider directly (eg, ec2 web console or CLI).  As you observed, if you delete the machine that is associated with such an instance, both instances will be cleaned up.

--- Additional comment from Michael Gugino on 2021-02-05 03:15:59 JST ---



--- Additional comment from Joel Speed on 2021-02-10 02:28:48 JST ---

This will now need to target 4.8 as we have past code freeze

--- Additional comment from Selim Jahangir on 2021-02-16 08:07:19 JST ---

Hi
Will the fix in this BZ be backported to ocp4.6 and ocp4.7?

Regards
Selim