logical decoding vs. VACUUM FULL / CLUSTER on table with TOAST-ed data

From: Tomas Vondra <tomas(dot)vondra(at)2ndquadrant(dot)com>
To: PostgreSQL Development <pgsql-hackers(at)postgresql(dot)org>
Subject: logical decoding vs. VACUUM FULL / CLUSTER on table with TOAST-ed data
Date: 2018-11-18 21:52:27
Message-ID: 1a17c643-e9af-3dba-486b-fbe31bc1823a@2ndquadrant.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Hi,

It seems we have pretty annoying problem with logical decoding when
performing VACUUM FULL / CLUSTER on a table with toast-ed data.

The trouble is that the rewritten heap is WAL-logged using XLOG/FPI
records, the TOAST data is logged as regular INSERT records. XLOG/FPI is
ignored in logical decoding, and so reorderbuffer never gets those
records. But we do decode the TOAST data, and reorderbuffer stashes them
in toast_hash hash table. Which gets reset only when handling a row from
the main heap, and that never arrives. So we end up stashing all the
TOAST data in memory :-(

So do VACUUM FULL (or CLUSTER) on a sufficiently large table, and you're
likely to break any logical replication connection. And it does not
matter if you replicate this particular table.

Luckily enough, this can leverage some of the pieces introduced by
e9edc1ba which was meant to deal with rewrites of system tables, and in
raw_heap_insert it added this:

/*
* The new relfilenode's relcache entrye doesn't have the necessary
* information to determine whether a relation should emit data for
* logical decoding. Force it to off if necessary.
*/
if (!RelationIsLogicallyLogged(state->rs_old_rel))
options |= HEAP_INSERT_NO_LOGICAL;

As raw_heap_insert is used only for heap rewrites, we can simply remove
the if condition and use the HEAP_INSERT_NO_LOGICAL flag for all TOAST
data logged from here.

This does fix the issue, because we still decode the TOAST changes but
there are no data and so

if (change->data.tp.newtuple != NULL)
{
dlist_delete(&change->node);
ReorderBufferToastAppendChunk(rb, txn, relation,
change);
}

ends up not stashing the change in the hash table. It's imperfect,
because we still decode the changes (and stash them to disk), but ISTM
that can be fixed by tweaking DecodeInsert a bit to just ignore those
changes entirely.

Attached is a PoC patch with these two fixes.

regards

--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Attachment Content-Type Size
rewrite-toast-fix.patch text/x-patch 2.5 KB

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Tom Lane 2018-11-18 22:02:48 Re: Fixing AC_CHECK_DECLS to do the right thing with clang
Previous Message Tom Lane 2018-11-18 21:43:59 Re: [HACKERS] WIP: Aggregation push-down