Hi,
I found that t/100_vacuumdb.pl has a fragile ordering check that fails
if the query plan for vacuumdb's catalogue query changes. I sometimes
see how this test fails when writing an optimisation-related extension.
The test checks that vacuumdb processes "Foo".bar before "Bar".baz:
qr/VACUUM\ \(SKIP_DATABASE_STATS\)\ "Foo".bar
.*VACUUM\ \(SKIP_DATABASE_STATS\)\ "Bar".baz
/sx,
Both tables being tested, "Foo".bar and "Bar".baz, are created empty.
This means pg_class.relpages is 0 for both and the sort order is
completely unstable. The output order depends entirely on which query
plan will be chosen. Any change in the planner that affects the plan for
this query, such as a new join path type or a cost model change, may
flip the order and cause the test to fail.
AFAICS, The fix is quite trivial. Change the test regex (in
100_vacuumdb.pl) to use order-independent lookaheads instead of a
sequential match:
qr/(?=.*VACUUM\ \(SKIP_DATABASE_STATS\)\ "Foo"\.bar)
(?=.*VACUUM\ \(SKIP_DATABASE_STATS\)\ "Bar"\.baz)
/sx,
This makes the test robust regardless of the order in which the server
returns results.
Hence, it doesn’t change anything important. I think it deserves to be
back-patched down to v.16 (like the commit 2143d96dc7b introduced this
test) so other extensions can be stable with check-world tests.
Regards,
Andrei Lepikhov
pgEdge
