Skip site navigation (1) Skip section navigation (2)

Analyzer is clueless

From: David Brown <time(at)bigpond(dot)net(dot)au>
To: pgsql-performance(at)postgresql(dot)org
Subject: Analyzer is clueless
Date: 2004-11-18 02:08:33
Message-ID: (view raw or flat)
Lists: pgsql-performance
I'm doing some performance profiling with a simple two-table query:

SELECT L."ProductID", sum(L."Amount")
FROM "drinv" H
JOIN "drinvln" L ON L."OrderNo" = H."OrderNo"
("OrderDate" between '2003-01-01' AND '2003-04-30')
GROUP BY L."ProductID"

drinv and drinvln have about 100,000 and 3,500,000 rows respectively. Actual data size in the large table is 500-600MB. OrderNo is indexed in both tables, as is OrderDate.

The environment is PGSQL 8 on Win2k with 512MB RAM (results are similar to 7.3 from Mammoth). I've tried tweaking various conf parameters, but apart from using up memory, nothing seems to have had a tangible effect - the Analyzer doesn't seem to take resources into account like some of the doco suggests.

The date selection represents about 5% of the range. Here's the plan summaries:

Three months (2003-01-01 to 2003-03-30) = 1 second

HashAggregate  (cost=119365.53..119368.74 rows=642 width=26)
  ->  Nested Loop  (cost=0.00..118791.66 rows=114774 width=26)
        ->  Index Scan using "drinv_OrderDate" on drinv h  (cost=0.00..200.27 rows=3142 width=8)
              Index Cond: (("OrderDate" >= '2003-01-01'::date) AND ("OrderDate" <= '2003-03-30'::date))
        ->  Index Scan using "drinvln_OrderNo" on drinvln l  (cost=0.00..28.73 rows=721 width=34)
              Index Cond: (l."OrderNo" = "outer"."OrderNo")

Four months (2003-01-01 to 2003-04-30) = 60 seconds

HashAggregate  (cost=126110.53..126113.74 rows=642 width=26)
  ->  Hash Join  (cost=277.55..125344.88 rows=153130 width=26)
        Hash Cond: ("outer"."OrderNo" = "inner"."OrderNo")
        ->  Seq Scan on drinvln l  (cost=0.00..106671.35 rows=3372935 width=34)
        ->  Hash  (cost=267.07..267.07 rows=4192 width=8)
              ->  Index Scan using "drinv_OrderDate" on drinv h  (cost=0.00..267.07 rows=4192 width=8)
                    Index Cond: (("OrderDate" >= '2003-01-01'::date) AND ("OrderDate" <= '2003-04-30'::date))

Four months (2003-01-01 to 2003-04-30) with Seq_scan disabled = 75 seconds

HashAggregate  (cost=130565.83..130569.04 rows=642 width=26)
  ->  Merge Join  (cost=519.29..129800.18 rows=153130 width=26)
        Merge Cond: ("outer"."OrderNo" = "inner"."OrderNo")
        ->  Sort  (cost=519.29..529.77 rows=4192 width=8)
              Sort Key: h."OrderNo"
              ->  Index Scan using "drinv_OrderDate" on drinv h  (cost=0.00..267.07 rows=4192 width=8)
                    Index Cond: (("OrderDate" >= '2003-01-01'::date) AND ("OrderDate" <= '2003-04-30'::date))
        ->  Index Scan using "drinvln_OrderNo" on drinvln l  (cost=0.00..119296.29 rows=3372935 width=34)

Statistics were run on each table before query execution. The random page cost was lowered to 2, but as you can see, the estimated costs are wild anyway.

As a comparison, MS SQL Server took less than 15 seconds, or 4 times faster.

MySQL (InnoDB) took 2 seconds, which is 30 times faster.

The query looks straightforward to me (it might be clearer with a subselect), so what on earth is wrong?


pgsql-performance by date

Next:From: Joshua D. DrakeDate: 2004-11-18 03:30:36
Subject: Re: postgres eating CPU
Previous:From: Tom LaneDate: 2004-11-18 01:57:44
Subject: Re: Analyzer is clueless

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group