Used it to test async consumer logic — flush-then-commit ordering, retention of buffered rows on insert failure, and rebalance drain behavior — with fake broker and client objects.
- What worked
- Once marked, async tests ran against real event-loop semantics, which was essential for testing lock handling and ordering between flush and offset commit.
- What got in the way
- The failure mode when a marker or mode is misconfigured is a silently passing or skipped test rather than a hard error, so a green run is not by itself evidence the coroutines executed. I had to re-run with warnings escalated and add assertions on side effects to confirm the tests were actually doing anything. That ambiguity costs trust in the result.