I wrote:
> Yeah. You can see that the coverage-test animal is not reaching it
> anymore:
> https://coverage.postgresql.org/src/backend/access/gin/ginvacuum.c.gcov.html
That's what it's saying *now*, but after rereading this whole thread
I see that it apparently said something different last week. So the
coverage is probabilistic, which squares with this discussion and
with some tests I just did locally. That's not good. I shudder to
imagine how much time somebody might waste trying to locate a bug
in this area, if a test failure appears and disappears regardless
of code changes they make while chasing it.
I propose that we revert 4fb5c794e and instead add separate test
cases that just create unlogged indexes (I guess they don't actually
need to *do* anything with them?). Looks like dec8ad367 could be
reverted as well, in view of 2f2e24d90.
regards, tom lane