Re: Performance degradation in 324577f39bc8738ed0ec24c36c5cb2c2f81ec660 - Mailing list pgsql-performance

From Vladimir Kamarzin
Subject Re: Performance degradation in 324577f39bc8738ed0ec24c36c5cb2c2f81ec660
Date
Msg-id 2012271412757623@web3g.yandex.ru
Whole thread Raw
In response to Re: Performance degradation in 324577f39bc8738ed0ec24c36c5cb2c2f81ec660  (Tom Lane <tgl@sss.pgh.pa.us>)
List pgsql-performance
07.10.2014, 19:59, "Tom Lane" <tgl@sss.pgh.pa.us>:
> Vladimir Kamarzin <vvk@vvk.pp.ru> writes:
>>  After upgrade from 9.3.1 to 9.3.5 we expirienced a slight performance degradation of all queries. Query time
increasedto some amount of ms, mostly in range of 100ms. Some actions in our application results in a lot of small
queriesand in such cases performance degradation is very significant - total action performs for a 2-3 times longer
thenbefore (15s -> 40s, etc). 
>>  Using git-bisect we've found a bad revision causes performance drop: it is 324577f39bc8738ed0ec24c36c5cb2c2f81ec660
>
> Hm.  If you're going to do queries that involve update/delete across large
> inheritance trees, that bug fix is unavoidably going to cost you some
> cycles.

Yeah, we're actually noticed significantly increased CPU load while running on 9.3.5.

> I am wondering if you've
> misidentified the commit that made the difference --- especially since you
> claim there's a penalty for "all" queries, which there manifestly couldn't
> be with this particular patch.

No, problem appears exactly on this commit. Actually I don't really sure about "all": we don't see degradation when
performingplain SELECTs manually, 
but comparing logged query time of some SELECTs we see the differences.

Here is example 42ms -> 250ms:
http://pastebin.ca/2855292
http://pastebin.ca/2855290

>  If not, there must be something rather
> unusual about your queries or schema.  Care to provide a self-contained
> test case?

I'm afraid we cannot do this now. If you wish, we can give you ssh access to the test server to investigate the
problem.


pgsql-performance by date:

Previous
From: and
Date:
Subject: Bad optimization/planning on Postgres window-based queries (partition by(, group by?)) - 1000x speedup
Next
From: Andrey Lizenko
Date:
Subject: Re: query plan question, nested loop vs hash join