Spotted this in the replication handler test on windows.
Replication especially can hold some resources out a little after core container shutdown is called.