This covers one possible deadlock, but there are so many more ways to do this. More common (IME) is a true logic deadlock, particularly if you are using strict serialization (say, your DB stores a ledger of financial transactions).

It’s actually really hard to test that your DB is using transactions correctly, since the error cases tend to be races which don’t show up under low load.

We built a prototype tool in Python where you can monkeypatch the transaction context manager, and pause the primary thread’s execution just before committing, so that you can then do evil stuff like running a competing thread of the same or other DB operations to try to break invariants. But even this won’t catch everything; there is a combinatorial explosion of test points and you can’t compare them all in a large app.

I haven’t seen any folks writing about this, another approach would be to wire up some him thing smarter like Jepsen to direct the anomaly search, Kyle said some folks have reported doing this but nobody has published. I’m interested to know if anyone has had success with this sort of transaction/correctness testing.