On 10/25/24 10:05, Andrei Lepikhov wrote:
On 10/24/24 22:30, jian he wrote:
hi. I think I found a bug.
PostgreSQL 18devel_debug_build_45188c2ea2 on x86_64-linux, compiled by
gcc-14.1.0, 64-bit
commit at 45188c2ea2.
Ubuntu 22.04.4 LTS
setup:
drop table t cascade;
create table t(a int PRIMARY key);
IN session1:
step "change data type" {begin; alter table t alter column a set data
type int4;}
step "s1" {commit;}
IN session2:
step "psql_another_session" {\d t}
permutation "change data type" "psql_another_session" "s1"
ERROR: cache lookup failed for attribute 1 of relation 34418
Yes, it looks like a bug existing for a long time, at least since PG11
(I didn't trace further down).
It seems that the backend didn't apply invalidation messages before
touching system caches. Backtrace:
After a short discovery, I found the origins:
The pg_get_indexdef has an incoming index oid and gets all the stuff
needed just by looking up sys-caches. But it wants to build a list of
relation column names at a specific moment and opens the heap relation.
After that operation, we already have syscaches updated and the old
index oid replaced with the new one.
It may be have made sense to lock the row of replaced index in pg_class
and pg_index until the transaction, altered it will be commmitted. But,
because ALTER TABLE is not fully MVCC-safe, it may be expected (or
acceptable) behaviour.
--
regards, Andrei Lepikhov
