06.02.2024 09:48, Amit Kapila wrote:
cool, is it possible to see whether this patch changes the runtime of this test in any noticeable way?
Yes, unfortunately it does. I've measured duration of 100 tests runs without the patch (with pristine bgwriter and with NO_TEMP_INSTALL): real 6m46,031s real 6m52,406s real 6m51,014s But with the patched test, I've got: real 9m39,872s real 9m40,044s real 9m38,236s (nearly 2 seconds increase per one test run) Under Valgrind, the original test run takes: Files=1, Tests=36, 334 wallclock secs ( 0.02 usr 0.00 sys + 163.14 cusr 7.98 csys = 171.14 CPU) But the patched one: Files=1, Tests=36, 368 wallclock secs ( 0.02 usr 0.00 sys + 182.16 cusr 8.90 csys = 191.08 CPU) (30 seconds increase) Maybe the more CPU-efficient solution would be disabling bgworker, as was proposed in another discussion of tests instability: https://www.postgresql.org/message-id/ZaTxhjnPygOdosJ4%40ip-10-97-1-34.eu-west-3.compute.internal Though I think that devising a way to control bgwriter may take more time than we can afford given the current 031 failure rate on the buildfarm (17 failures for the last 3 days). Best regards, Alexander
