Hi,
On 2022-04-07 11:02:41 -0700, Andres Freund wrote:
> I've now reproduced this, albeit not reliably yet. Looking.
Caused by me misremembering when deduplication happens - somehow recalled that
deduplication didn't happen when payloads. So the statement that was supposed
to guarantee needing more than one page:
SELECT pg_notify('stats_test_use', repeat('0', current_setting('block_size')::int / 2)) FROM generate_series(1, 3);
didn't actually guarantee that. It just failed to fail by chance.
When regression tests and isolation test run in sequence against the same
freshly started cluster, the offsets when starting out are just right to not
need another page in the first test.
I'll change it to use distinct payloads..
Greetings,
Andres Freund