UPDATE on NOT JOIN - Mailing list pgsql-performance

From Gabriel Biberian
Subject UPDATE on NOT JOIN
Date
Msg-id 4F3BFA8D.1050200@beemotechnologie.com
Whole thread Raw
Responses Re: UPDATE on NOT JOIN  (Marti Raudsepp <marti@juffo.org>)
List pgsql-performance
Hello,

I'm working on a system that detects changes in someone's filesystem.
We first scan the entire filesystem which is dumped into the table
'filesystem' containing the full_path of every file and it's
corresponding md5 hash.
We then do subsequent scans of the filesystem which are dumped into a
temporary table and we would like to find out which files are not
present anymore.  To do so, we update the field 'dead' in the
'filesystem' table for every row that does not exist in the 'temporary'
table.

\d filesystem;
               Table « public.filesystem »
   Colonne   |          Type          |    Modificateurs
------------+------------------------+----------------------
  hash       | character(32)          |
  full_name  | text                   |
  dead       | integer                | default 0
Index :
     « filesystem_unique » UNIQUE, btree (hash)

\d temporary;
               Table « public.filesystem »
   Colonne   |          Type          |    Modificateurs
------------+------------------------+----------------------
  hash       | character(32)          |
  full_name  | text                   |
Index :
     « temporary_unique » UNIQUE, btree (hash)

Currently, i use the following query to update the filesystem table with
the missing files :
UPDATE filesystem SET dead=some_value WHERE dead=0 AND (SELECT 1 FROM
temporary AS t WHERE t.hash=filesystem.hash LIMIT 1) IS NULL

This works correctly for regular filesystems.  But when the 'filesystem'
table contains more than a few million rows, the update query can take days.

Here's an explain of the query :
=# UPDATE filesystem SET dead=some_value WHERE dead=0 AND (SELECT 1 FROM
temporary AS t WHERE t.hash=filesystem.hash LIMIT 1) IS NULL
                                                       QUERY PLAN
----------------------------------------------------------------------------------------------------------------------
  Seq Scan on filesystem  (cost=0.00..97536479.02 rows=25747 width=241)
    Filter: ((dead = 0) AND ((subplan) IS NULL))
    SubPlan
      ->  Limit  (cost=0.00..9.53 rows=1 width=0)
            ->  Index Scan using temporary_hash on temporary t
(cost=0.00..9.53 rows=1 width=0)
                  Index Cond: (hash = $0)
(6 lignes)

Is there a better way to update a table if it doesn't join another table ?

Best Regards,

Gabriel Biberian

pgsql-performance by date:

Previous
From: Andres Freund
Date:
Subject: Re: Fwd: [HACKERS] client performance v.s. server statistics
Next
From: Zhou Han
Date:
Subject: Fwd: [HACKERS] client performance v.s. server statistics