Since our lock manager works based on namespaces and not UUIDs, there is an "optimistic" protocol which ensures that a UUID is locked exactly as the namespace that it maps to. The protocol is:

do {
  NSSBeforeLock = optimistically resolve NSS to UUID;
  Lock the NSS();
  NSSAfterLock = optimistically resolve NSS to UUID;
while (NSSBeforeLock == NSSAfterLock);

The above protocol guarantees that when one holds an AutoGetCollection object, the NSS <-> UUID mapping reflects the truth and is stable for the duration of the RAII object.

The multi-collection locking path performs the UUID -> NSS resolution before it invokes the CollectionLock constructor (which is what implements the protocol above) and as a result it is possible that the guarantee above is not met. This is easily reproducible if the "fast" path here is skipped.