I wrote: > So that's the same hardware, and identical PG source tree, and different > results. This seems to leave only two theories standing:
I forgot theory 3: it's intermittent. Apparently the probability has dropped a lot since 2018, but behold: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=flaviventris&dt=2021-09-28%2014%3A20%3A41 (with successful runs just before and after this one, on the same animal) Note that the delta is not exactly like the previous result, either. So there's more than one symptom, but in any case it seems like we have an issue in WAL replay. I wonder whether it's bloom's fault or a core bug. regards, tom lane
