Fix false reports in pg_visibility
authorAlexander Korotkov <akorotkov@postgresql.org>
Thu, 14 Mar 2024 11:08:53 +0000 (13:08 +0200)
committerAlexander Korotkov <akorotkov@postgresql.org>
Thu, 14 Mar 2024 11:12:05 +0000 (13:12 +0200)
commite85662df44ff47acdf5d2d413339445d60a9c30c
tree4bb7568145ead39658fd113921e2ca01775601e3
parentcc6e64afda530576d83e331365d36c758495a7cd
Fix false reports in pg_visibility

Currently, pg_visibility computes its xid horizon using the
GetOldestNonRemovableTransactionId().  The problem is that this horizon can
sometimes go backward.  That can lead to reporting false errors.

In order to fix that, this commit implements a new function
GetStrictOldestNonRemovableTransactionId().  This function computes the xid
horizon, which would be guaranteed to be newer or equal to any xid horizon
computed before.

We have to do the following to achieve this.

1. Ignore processes xmin's, because they consider connection to other databases
   that were ignored before.
2. Ignore KnownAssignedXids, because they are not database-aware. At the same
   time, the primary could compute its horizons database-aware.
3. Ignore walsender xmin, because it could go backward if some replication
   connections don't use replication slots.

As a result, we're using only currently running xids to compute the horizon.
Surely these would significantly sacrifice accuracy.  But we have to do so to
avoid reporting false errors.

Inspired by earlier patch by Daniel Shelepanov and the following discussion
with Robert Haas and Tom Lane.

Discussion: https://postgr.es/m/1649062270.289865713%40f403.i.mail.ru
Reviewed-by: Alexander Lakhin, Dmitry Koval
contrib/pg_visibility/Makefile
contrib/pg_visibility/meson.build
contrib/pg_visibility/pg_visibility.c
contrib/pg_visibility/t/001_concurrent_transaction.pl [new file with mode: 0644]
src/backend/storage/ipc/procarray.c
src/include/storage/standby.h