Attached patch adjusts BRIN regression tests to make a non-obvious
dependency on tuple order explicit. Currently, an index-only scan plan
is used by the query that I've adjusted. I'd rather be sure that that
continues.
This was spotted while running the regression tests with the Postgres
default B-Tree fillfactor significantly reduced. There was only one
other failure, but that one was really obvious; in general I wouldn't
bother with "fixing" such things. This, however, seems problematic,
because insertion order matters for the purposes of the test. It's
perhaps implied that the test consistently relies on the stable order
of things within the tenk1 table, but it isn't actually so.
--
Peter Geoghegan