Feature migration refuses to run if all features requiring upgrade have encountered errors

[gwbrown]

Elasticsearch Version

7.17.0

Installed Plugins

none

Java Version

bundled

OS Version

Doesn't matter

Problem Description

If all features which still require migration have encountered an error, trying to start migration produces no results, only a response claiming no migration is necessary, even if that's not true. This means that transient errors that occur during migration are impossible to retry.

The only known workaround for this is a full-cluster restart, which will clear the migration upgrade result metadata.

Steps to Reproduce

1. Start a cluster using a 6.x version.
2. Upgrade it to 7.17.0.
3. Run system index migration, but ensure that some indices fail (e.g. by triggering [Upgrade] migration system indices fails for tasks: mapper_parsing_exception #83779, which .tasks always will - fortunately that is easy enough to work around by removing the index)
4. Try to re-run migration. Migration will not try to re-run.

Logs (if relevant)

The response will just say no migration is necessary:

{
  "accepted" : false,
  "reason" : "No system indices require migration"
}

[elasticmachine]

Pinging @elastic/es-core-infra (Team:Core/Infra)