Обсуждение: Avoiding a deadlock

I have a long-running multi-row UPDATE that is deadlocking with a single-row UPDATE:

2013-03-09 11:07:51 CST ERROR:  deadlock detected
2013-03-09 11:07:51 CST DETAIL:  Process 18851 waits for ShareLock on transaction 10307138; blocked by process 24203.
        Process 24203 waits for ShareLock on transaction 10306996; blocked by process 18851.
        Process 18851: UPDATE  taggings tg
                SET     score_tier = COALESCE(x.perc, 0)
                FROM    (SELECT tg2.id,
                                percent_rank() OVER (PARTITION BY tg2.tag_id ORDER BY tg2.score ASC) AS perc
                         FROM   taggings tg2, tags t
                         WHERE  tg2.score IS NOT NULL
                         AND    tg2.tag_id = t.id
                         AND    t.tier >= 2) AS x
                WHERE   tg.id = x.id
                AND     tg.score IS NOT NULL
                ;
        Process 24203: UPDATE "taggings" SET "score" = 2 WHERE "taggings"."id" = 29105523

Note that these two queries are actually updating different columns, albeit apparently in the same row.

Is there anything I can do to avoid a deadlock here? The big query does nothing else in its transaction; the little query's transaction might update several rows from `taggings`, which I guess is the real reason for the deadlock.

I'd be pretty satisfied with approximate values for the big query. As you can see, it is just taking the `score` of each `tagging` and computing the percentage of times it beats other taggings of the same tag. Is there something I can do with transaction isolation levels here? I don't care if the big query operates on slightly-out-of-date values. Since each query updates different columns, I think there should be no issue with them overwriting each other, right?

Thanks,
Paul



--
_________________________________
Pulchritudo splendor veritatis.

Paul Jungwirth wrote:
> I have a long-running multi-row UPDATE that is deadlocking with a single-row UPDATE:
> 
> 2013-03-09 11:07:51 CST ERROR:  deadlock detected
> 2013-03-09 11:07:51 CST DETAIL:  Process 18851 waits for ShareLock on transaction 10307138; blocked by
> process 24203.
>         Process 24203 waits for ShareLock on transaction 10306996; blocked by process 18851.
>         Process 18851: UPDATE  taggings tg
>                 SET     score_tier = COALESCE(x.perc, 0)
>                 FROM    (SELECT tg2.id,
>                                 percent_rank() OVER (PARTITION BY tg2.tag_id ORDER BY tg2.score ASC)
> AS perc
>                          FROM   taggings tg2, tags t
>                          WHERE  tg2.score IS NOT NULL
>                          AND    tg2.tag_id = t.id
>                          AND    t.tier >= 2) AS x
>                 WHERE   tg.id = x.id
>                 AND     tg.score IS NOT NULL
>                 ;
>         Process 24203: UPDATE "taggings" SET "score" = 2 WHERE "taggings"."id" = 29105523
> 
> Note that these two queries are actually updating different columns, albeit apparently in the same
> row.
> 
> Is there anything I can do to avoid a deadlock here? The big query does nothing else in its
> transaction; the little query's transaction might update several rows from `taggings`, which I guess
> is the real reason for the deadlock.
> 
> I'd be pretty satisfied with approximate values for the big query. As you can see, it is just taking
> the `score` of each `tagging` and computing the percentage of times it beats other taggings of the
> same tag. Is there something I can do with transaction isolation levels here? I don't care if the big
> query operates on slightly-out-of-date values. Since each query updates different columns, I think
> there should be no issue with them overwriting each other, right?

The problem is that both updates affect the same rows.
It does not matter if they update different columns, since in any
case a new row version is created (read about PostgreSQL's MVCC
implementation in the documentation).

I can only think of two ways to avoid this deadlock:

1) Each of the "little transactions" modifies no more than one row of the table.

2) All transactions modify table rows in the same order, e.g. ascending "id".
   With the big update you can do that by putting an "ORDER BY tg2.id" into
   the subquery, and with the "little transactions" you'll have to make sure
   that rows are updated in ascending "id" order.

Yours,
Laurenz Albe

On 11 March 2013 13:01, Chris Curvey <chris@chriscurvey.com> wrote:

All the suggestions thus far only reduce the window in which a dead lock can occur.

Alban Hertroys wrote:
> All the suggestions thus far only reduce the window in which a dead lock can occur.

Where do you see a window for deadlocks with my suggestions?

Yours,
Laurenz Albe

I agree this would fix the deadlock. It also seems like the least disruptive way of fixing the problem.

Out of curiosity: any reason the ORDER BY should be in the subquery? It seems like it ought to be in the UPDATE (if that's allowed).

Thanks,
Paul

> Out of curiosity: any reason the ORDER BY should be in the subquery? It seems like it ought to be in the UPDATE (if that's allowed).

Hmm, it's not allowed. :-) It's still surprising that you can guarantee the order of a multi-row UPDATE by ordering a subquery.

Paul

--
_________________________________
Pulchritudo splendor veritatis.

Paul Jungwirth wrote:
>> Out of curiosity: any reason the ORDER BY should be in the subquery? It seems like it ought to be in
> the UPDATE (if that's allowed).
> 
> Hmm, it's not allowed. :-) It's still surprising that you can guarantee the order of a multi-row
> UPDATE by ordering a subquery.

To be honest, I don't think that there is any guarantee for this
to work reliably in all comparable cases, as PostgreSQL does
not guarantee in which order it performs the UPDATEs.

It just happens to work with certain plans (use EXPLAIN
to see wat will happen).

Yours,
Laurenz Albe