What's the problem?

The real issue is that we are clearing the filtering metadata without actually holding the recoverable critical section. This mostly happens in DDL coordinators: it could happen that the phase that releases the critical section is retried but the critical section was already released. In this case, we shouldn't clear the filtering metadata (e.g. here) because it was done before.
Note that this is not just a problem of that command but potentially it affects other DDL operations. Any DDL that has a phase that 1) clears the filtering metadata and 2) releases the critical section might be affected as well. 

Solutions
A potential solution would be to clear the filtering metadata on primaries and secondaries when releasing the critical section. Right now we are only doing on secondaries.