Horton

Re: [COMMITTERS] pgsql: Fix freezing of a dead HOT-updated tuple

Peter Geoghegan <pg@bowt.ie>

View thread View on postgr.es
From: Peter Geoghegan <pg@bowt.ie>
To: Alvaro Herrera <alvherre@alvh.no-ip.org>
Cc: "Wood, Dan" <hexpert@amazon.com>, Michael Paquier <michael.paquier@gmail.com>, "Wong, Yi Wen" <yiwong@amazon.com>, PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Date: 2017-10-06T18:07:45Z
Lists: pgsql-hackers 
On Fri, Oct 6, 2017 at 7:59 AM, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:
> By the way, I still wonder if there's any way for a new tuple to get
> inserted in the place where a HOT redirect would be pointing to, and
> have it be marked as Frozen, where the old redirect contains a
> non-invalid Xmax.  I tried to think of a way for that to happen, but
> couldn't think of anything.
>
> What I imagine is a sequence like this:
>
> 1. insert a tuple
> 2. HOT-update a tuple
> 3. prune the page, making lp 1 be a redirect (lp 2 is the new tuple)
> 4. start transaction
> 5. HOT-update the tuple again, creating HOT in lp 3
> 6. abort transaction (leaving aborted update in lp 3)
> 7. somehow remove tuple from lp 3, make slot available
> 8. new transaction comes along, inserts tuple in lp 3
> 9. somebody freezes tuple in lp3 (???)
>
> Then we follow the HOT chain, see that Xmin=2 in lp3 and conclude that
> the tuple is part of the chain because of an xid "match".
>
> Basically from step 7 onwards I don't think this is possible, but maybe
> I'm just blind.

For the record, I also think that this is impossible, in part because
pruning requires a cleanup buffer lock (and because HOT chains cannot
span pages). I wouldn't say that I am 100% confident about this,
though.

BTW, is this comment block that appears above
heap_prepare_freeze_tuple() now obsolete, following 20b65522 (and
maybe much earlier commits)?

 * NB: It is not enough to set hint bits to indicate something is
 * committed/invalid -- they might not be set on a standby, or after crash
 * recovery.  We really need to remove old xids.
 */

We WAL-log setting hint bits during freezing now, iff tuple xmin is
before the Xid cutoff and tuple is a heap-only tuple.

-- 
Peter Geoghegan

Commits

  1. Backport addition of rs_old_rel to rewriteheap's state.

  2. Perform a lot more sanity checks when freezing tuples.

  3. Revert bogus fixes of HOT-freezing bug

  4. Fix traversal of half-frozen update chains

  5. Fix freezing of a dead HOT-updated tuple

  6. During index build, check and elog (not just Assert) for broken HOT chain.

  7. Fix WAL replay of locking an updated tuple

  8. Change the way we mark tuples as frozen.