Corrupted btree index on HEAD because of covering indexes

Michael Paquier <michael@paquier.xyz>

From: Michael Paquier <michael@paquier.xyz>
To: Postgres hackers <pgsql-hackers@postgresql.org>
Cc: Peter Geoghegan <pg@bowt.ie>, Anastasia Lubennikova <a.lubennikova@postgrespro.ru>, Teodor Sigaev <teodor@sigaev.ru>
Date: 2018-04-19T05:24:36Z
Lists: pgsql-hackers
Hi all,

I was just testing the VACUUM truncation logic, and bumped into what
looks like a corrupted btree index.  Here is a reproducer:
create table aa (a int primary key, b bool);
insert into aa values (generate_series(1,1000000), false);
checkpoint;
update aa set b = false where a > 500000; -- Dirties a set of shared
buffers
delete from aa where a > 750000; -- Delete a set of rows
vacuum aa;
delete from aa where a > 10;
vacuum aa; -- error on btree with right sibling

And here is the actual failure when the second vacuum:
ERROR:  XX000: right sibling 4132 of block 2128 is not next child 5396 of block 412 in index "aa_pkey"
LOCATION:  _bt_mark_page_halfdead, nbtpage.c:1564

This works on REL_10_STABLE, so I am adding an open item.  I have not
investigated the exact problem yet, but bisect is showing me covering
indexes as the culprit (8224de4).

Thanks,
--
Michael

Commits

  1. Add missing and dangling downlink checks to amcheck

  2. Fix wrong validation of top-parent pointer during page deletion in Btree.

  3. Add integrity-checking functions to pg_visibility.

  4. Make the handling of interrupted B-tree page splits more robust.