Here's how this should work:

• The migrator should operate on a processed desiredState. The idea is to modify it (the desiredState) to meet migration needs.
• Implementing it as you proposed, will tie our hands for the future. A migrated version might require additional K8s objects (such as ConfigMaps or Secrets). As the migrator executes before the reconciler (and moreover, the reconciler backs off during the migration), those objects will not be created. This may lead to many problems in the future and very messy code. I'd like to avoid it.
• The migrator should obtain a copy of desiredState from the main loop (pass by value), find the StatefulSet object, change its number of replicas to 1 and pass it back to the main loop.
• Once the ActionRunner executes it, the number of replicas should drop to 1 (if your cluster had 10 replicas, 9 of them should get into terminating state) and the StatefulSet rolling upgrade should kill the remaining one during the upgrade.
• The next loop should increase the number of replicas back to 10.

So it seems to me there some bug in the migrator, so that it doesn't return a modified StatefulSet. Would you mind to check it up?

Thanks, @slaskawi. I see now. This was just an idea. I will incorporate your suggestions.