Horton

RE: Selectively invalidate caches in pgoutput module

Hayato Kuroda (Fujitsu) <kuroda.hayato@fujitsu.com>

View thread View on postgr.es
From: "Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>
To: "Zhijie Hou (Fujitsu)" <houzj.fnst@fujitsu.com>
Cc: "ShlokKumar.Kyal@fujitsu.com" <ShlokKumar.Kyal@fujitsu.com>, 'Amit Kapila' <amit.kapila16@gmail.com>, Masahiko Sawada <sawada.mshk@gmail.com>, "pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>
Date: 2025-03-10T13:12:00Z
Lists: pgsql-hackers

Attachments

Dear Hou,

> Currently, only the leaf partition is invalidated when the published table is
> partitioned. However, I think pgoutput could cache both the partitioned table
> and the leaf partition table as relsync entries.
> 
> For INSERT/UPDATE/DELETE on a partitioned table, only the leaf partition's
> relsync entry is used in pgoutput, but the TRUNCATE references the parent
> table's relsync entry.

I think your analysis is correct. PSA new version. Below part contains my analysis.

In ExecuteTruncate(), if the specified relation has children, all of them are
checked via find_all_inheritors() and listed as target. Also, ExecuteTruncateGuts()
serializes both a parent and children in XLOG_HEAP_TRUNCATE WAL record.
Decoding layer passes relations as-is. These facts mean that output plugins can
store caches on the memory.

Best regards,
Hayato Kuroda
FUJITSU LIMITED

Commits

  1. Avoid invalidating all RelationSyncCache entries on publication rename.

  2. Avoid invalidating all RelationSyncCache entries on publication change.