Explicitly setting autoflush to False on concurrent write paths, or enforcing a uniform lock ordering, is a cheaper choice than investigating deadlocks after the fact.