When the primary owner crashes during prepare, the prepare command is retried on the new primary owner. But because the transaction topology id stays the same, the new primary owner does not check for backup locks owned by other transactions from the previous topology.

That makes it possible for the retried prepare command to lock a key that was already locked by another transaction on the primary owner.

This wasn't a problem before the bug 1212795 (ISPN-4546) fix, because a primary owner crash would have forced the transaction to roll back.