On 2021-Jan-18, Tom Lane wrote: > Right. If that's the explanation, then adding DISABLE_PAGE_SKIPPING > to the test's VACUUM options should fix it. However, to believe that > theory you have to have some reason to think that some other process > might have the page pinned. What would that be? test1 only has one > small tuple in it, so it doesn't seem credible that autovacuum or > autoanalyze would have fired on it.
I guess the machine would have to be pretty constrained. (It takes almost seven minutes to go through the pg_upgrade test, so it does seems small.) > [ thinks for a bit... ] Does the checkpointer pin pages it's writing > out? I guess it'd have to ... It does, per SyncOneBuffer(), called from BufferSync(), called from CheckPointBuffers(). -- Álvaro Herrera 39°49'30"S 73°17'W
