Consider the following interleaving:

1. 'Shard1' is the db-primary shard for database 'dbA'. There's an unsharded collection 'dbA.coll'.
2. User runs movePrimary(dbA, to: 'Shard2') -> On 'shard1', the MovePrimary coordinator commits the operation and releases the critical section. But hangs before cleaning the "stale db data".
3. User runs dropDatabase -> On 'shard2', which is now the db-primary shard.
4. User now recreates 'dbA' on 'shard1' and does some writes on 'dbA.coll'.
5. Now the MovePrimary coordinator of point (2) proceeds with cleaning "stale data". It will drop 'dbA.coll', so the writes that happened on point (4) are lost!