On Fri, Mar 23, 2018 at 8:27 PM, Pavan Deolasee <pavan.deolasee@gmail.com> wrote: >> >> >> I would probably just have a few regression lines that should be sure >> to exercise the code path and leave it at that. >> > > I changed the regression tests to include a few more scenarios, basically > using multi-column indexes in different ways and they querying rows by > ordering rows in different ways. I did not take away the vacuum and I > believe it will actually help the tests by introducing some fuzziness in the > tests i.e. if the vacuum does not do its job, we might execute a different > plan and ensure that the output remains unchanged. >
If we're going to keep the vacuums in there, do we need to add a wait barrier like Claudio suggested upthread?
I don't think we need the wait barrier since we're no longer printing the explain plan. In the worst case, the vacuum may not get to set pages all-visible, thus planner choosing something other than an index-only-scan, but I guess that fuzziness actually helps the regression tests. That way we get confirmation regarding the final result irrespective of the plan chosen.