Hi,
On 2022-02-04 14:58:59 -0500, Tom Lane wrote:
> I've pushed the first draft for $SUBJECT at
>
> https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=ab22eea83169c8d0eb15050ce61cbe3d7dae4de6
+Author: Andres Freund <andres@anarazel.de>
+Branch: master [18b87b201] 2022-01-13 18:13:41 -0800
+Branch: REL_14_STABLE [dad1539ae] 2022-01-14 10:56:12 -0800
+-->
+ <para>
+ Fix corruption of HOT chains when a RECENTLY_DEAD tuple changes
+ state to fully DEAD during page pruning (Andres Freund)
+ </para>
Even if that happens, it's still pretty unlikely to cause corruption - so
maybe s/corruption/chance of corruption/?
+ <para>
+ This happens when the last transaction that could <quote>see</quote>
+ the tuple ends while the page is being pruned.
The transaction doesn't need to have ended while the page is vacuumed - the
horizon needs to have been "refined/updated" while the page is pruned so that
a tuple version that was first considered RECENTLY_DEAD is now considered
DEAD. Which can only happen if RecentXmin changed after
vacuum_set_xid_limits(), which only can happen if catalog snapshot
invalidations and other invalidations are processed in vac_open_indexes() and
RecentXmin changed since vacuum_set_xid_limits(). Then a page involving
tuples in a specific "arrangement" need to be encountered.
That's obviously to complicated for the release notes. Trying to make it more
understandable I came up with the following, which still does not seem great:
This can only happen if transactions, some having performed DDL, commit
within a narrow window at the start of VACUUM. If VACUUM then prunes a
page containing several tuple version that started to be removable within
the aforementioned time window, the bug may cause corruption on that page
(but no further pages). A tuple that is pointed to by a redirect item
elsewhere on the page can get removed. [...]
Greetings,
Andres Freund
