Hi!
> I think if we update the stats in copy_heap_data after copying the
> data, then we don't see such problem. Attached patch should fix the
> issue. You can try this patch to see if it fixes the issue for you.
I'm afraid I'm not able to reproduce the problem which patch should fix.
What I did (today's master, without patch):
autovacuum off
pgbench -i -s 100
select relname, relpages, reltuples from pg_class where relname =
'pgbench_accounts';
relname | relpages | reltuples
------------------+----------+-----------
pgbench_accounts | 163935 | 1e+07
vacuum full pgbench_accounts;
# select relname, relpages, reltuples from pg_class where relname =
'pgbench_accounts';
relname | relpages | reltuples
------------------+----------+-----------
pgbench_accounts | 163935 | 1e+07
I've tried to add hash index to that table and print notice about number of
pages and tuples immediately after estimate_rel_size() in hashbuild(). hash
index got right estimation even I deleted all rows before vacuum full. What am I
doing wrong?
Patch looks good except, seems, updating stats is better to move to
swap_relation_files(), then it will work even for toast tables.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/
