Our TemporaryRecordStore requires that callers “finalize” by dropping the table (excluding “keeping” it which is used by resumable index builds). The API is a bit clumsy because it requires callers to clean up by dropping the RS, which needs to be done under a Global lock. If this does not happen, the table gets leaked until the next restart.

We can improve this API by allowing the destructor to queue up the drop with the drop-pending ident reaper and a 0 timestamp. This would reap the table immediately in the next pass (every second) and loosens the requirements for the calling operation in addition to making it less likely to leak resources.