I wrote:
> Specifically, I notice that in the last couple of failures involving
> tests on "mcv_lists", we are considering stats on a 5000-row table.
> With the standard setting default_statistics_target = 100, ANALYZE
> will take a 3000-row random sample, meaning that its results are
> *inherently* not 100% reproducible.
No, wait, scratch that --- I misremembered the multiplier. Actually
ANALYZE will try to acquire a 30000-row sample, so its sample *should*
always include the entire table. Autovacuum or not, the stats ought
to come out exactly the same every time in this test case. But they
are not doing so, at least on some machines. We need to figure out
why that is.
Anyway, I remain suspicious that the instability is the fault
of something in the statistics code, not of autovacuum.
regards, tom lane