Re: Toast issues with OldestXmin going backwards

From: Andrew Gierth <andrew(at)tao11(dot)riddles(dot)org(dot)uk>
To: Andres Freund <andres(at)anarazel(dot)de>
Cc: Pavan Deolasee <pavan(dot)deolasee(at)gmail(dot)com>, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com>, pgsql-hackers <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: Toast issues with OldestXmin going backwards
Date: 2018-04-23 02:09:22
Message-ID: 87o9iabtkt.fsf@news-spur.riddles.org.uk
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

>>>>> "Andrew" == Andrew Gierth <andrew(at)tao11(dot)riddles(dot)org(dot)uk> writes:

Andrew> Since the tuple was vacuumable at that point, it will never
Andrew> again become visible in any snapshot, so the only code that
Andrew> should ever access its toast values is vacuum full/cluster,

and, unfortunately, CREATE INDEX (non-concurrently), including the index
build done by vacfull/cluster itself, which rather messes things up
because the toasted column might be a parameter to a functional index, or it
might itself be an indexed column (within a narrow size range for btree,
or more generally for other index types); this means the index AM or the
indexed function may try and actually detoast and inspect the value.

So right now it's possible (I think) in the worst case to crash the
server if you can set up exactly the right circumstances (re-use of a
prematurely vacuumed toast oid after wraparound, with the new value
matching the old value's raw length but not its compression state).

(My changes so far don't address this case and don't make it either more
or less likely)

--
Andrew (irc:RhodiumToad)

In response to

Browse pgsql-hackers by date

  From Date Subject
Next Message Amit Langote 2018-04-23 02:37:12 using expression syntax for partition bounds (was: Re: Boolean partitions syntax)
Previous Message Amit Langote 2018-04-23 01:04:11 Re: [HACKERS] path toward faster partition pruning