Andres Freund <andres@anarazel.de> writes:
> I think you don't event need checkpointer to be involved, normal buffer
> replacement would do the trick. We briefly pin the page in BufferAlloc()
> even if the page is clean. Longer when it's dirty, of course.
True, but it seems unlikely that the pages in question here would be
chosen as replacement victims. These are non-parallel tests, so
there's little competitive pressure. I could believe that a background
autovacuum is active, but not that it's dirtied so many pages that
tables the test script just created need to get swapped out.
The checkpointer theory seems good because it requires no assumptions
at all about competing demand for buffers. If the clock sweep gets
to the table page (which we know is recently dirtied) at just the right
time, we'll see a failure.
regards, tom lane
