The journaled in-memory implementation (with fine-grained locks to limit contention) is now more stable and the tests are no longer flaky. Race conditions identified and redesigned to fix.
In-memory should now be as reliable and durable as the Postgres implementation (which is still a solid choice!), but much simpler to set up.