Proving a race condition with Playwright

In bbSlate, students work through pages. Two APIs run close together: one submits the page, and one records the last page the student visited, so they can pick up where they left off.

Sometimes the saved state was wrong, and it was not repeatable by hand. Clicking through the flow worked almost every time. A bug that only shows up under timing is easy to wave away as "can't reproduce".

Both calls touch the same student's state, close together in time. If they arrive in an unlucky order, or overlap, one can overwrite the other. That was a theory; it needed proof.

I wrote a Playwright script that fires the two API calls concurrently, many times, and checks the stored state after each round. A person clicking can't control timing. A script can.

The script reproduced the wrong state reliably. The two APIs were racing on the same data, and the result depended on which one finished last.

With the race reproducible on demand, I fixed it and used the same script as the check: it failed before the fix and passed after. In the same area I also fixed grader submission failures, an HTTP 200 returned for expired tokens, and missing token-expiry logs.

An intermittent "can't reproduce" became a proven, fixed bug with a test that keeps it fixed.

Turn "it happens sometimes" into a test that fails every time. When a bug depends on timing, make the timing the input of the test.