[pgpool-general: 1055] Re: using prepared statements when memory_cache_enabled=on

[Tatsuo Ishii]

> On Sun, Sep 16, 2012 at 12:18 AM, Tatsuo Ishii <ishii at postgresql.org> wrote:
>>>> On Fri, Sep 14, 2012 at 3:37 PM, Tatsuo Ishii <ishii at postgresql.org> wrote:
>>>>>> On Fri, Sep 14, 2012 at 3:09 PM, Tatsuo Ishii <ishii at sraoss.co.jp> wrote:
>>>>>>>> On Thu, Sep 13, 2012 at 6:35 PM, Tatsuo Ishii <ishii at postgresql.org> wrote:
>>>>>>>>>> pg_* functions.  Stuff like pg_prepare, pg_exec, pg_query.
>>>>>>>>>
>>>>>>>>> Can you show us pg_prepare example? I'm interested in what statement
>>>>>>>>> names you are using.
>>>>>>>>
>>>>>>>> pg_prepare($db, $preparedStatementKey, "SELECT
>>>>>>>> id,pass,fail,filtercount,info,current_status,last_update,os,arch,branch,gpu,build_type,subtest,osversion
>>>>>>>> FROM cudasmoke WHERE (last_update > $1 AND last_update < $2) AND os=$3
>>>>>>>> AND arch=$4 AND branch=$5 AND build_type= $6 AND subtest=$7 AND
>>>>>>>> osversion=$8 AND gpu=$9 ORDER BY last_update DESC LIMIT 1" )
>>>>>>>
>>>>>>> So what's the content of $preparedStatementKey?
>>>>>>
>>>>>> Its a unique identifier string for the prepared statement, comprised
>>>>>> of the concatenated sha1sum and md5sum of the prepared statement.
>>>>>>
>>>>>> So it ends up looking something like this:
>>>>>> pg_execute($db,a1a3a2f461a0efc2eee2f60f940df6b77eb552f9f5a985165c3ab1665384fa9cce8b1b62,"SELECT
>>>>>> id,pass,fail,filtercount,info,current_status,last_update,os,arch,branch,gpu,build_type,subtest,osversion
>>>>>> FROM cudasmoke WHERE (last_update > $1 AND last_update < $2) AND os=$3
>>>>>> AND arch=$4 AND branch=$5 AND build_type= $6 AND subtest=$7 AND
>>>>>> osversion=$8 AND gpu=$9 ORDER BY last_update DESC LIMIT 1);
>>>>>
>>>>> The statement name is 72 byte long. This is interesting because
>>>>> PostgreSQL (and pgpool) does not suppose that statement name is longer
>>>>> than 64.  Not sure how this affects to the problem though.
>>>>>
>>>>> src/include/pg_config_manual.h:
>>>>> /*
>>>>>  * Maximum length for identifiers (e.g. table names, column names,
>>>>>  * function names).  Names actually are limited to one less byte than this,
>>>>>  * because the length must include a trailing zero byte.
>>>>>  *
>>>>>  * Changing this requires an initdb.
>>>>>  */
>>>>> #define NAMEDATALEN 64
>>>>
>>>> That is interesting, and I was unaware of this limitation.  Anyway,
>>>> even if, for the purposes of experimentation, I only use the sha1sum
>>>> (and drop the md5sum) such that the identifier is reduced in size to
>>>> 40 bytes, I still see the same failure.  However, one thing that I
>>>> noticed is that pgpool doesn't seem to realize that I've attempted to
>>>> use a new prepared statement identifier (now just the sha1sum, without
>>>> md5sum), and still tries to use the previously cached cached one.
>>>>
>>>> I'd expect that if I'm preparing a statement with an unused (new)
>>>> identifier, it shouldn't pull it from the cache, but it appears to do
>>>> so anyway.
>>>
>>> Why do you think so? If SELECT(and with same parameters) is identical,
>>> it should return exactly same results regardress the statement name(of
>>> courese I do not account underlying tables get modified).
>>
>> BTW, we have recently fixed a bug which can cause buffer overrun error:
>> http://git.postgresql.org/gitweb/?p=pgpool2.git;a=commit;h=3964d8204373473e3130a9ba5b260049420dafb6
>>
>> Because it's a memory destruction error, it may or may not be related to your problem. Can you try with pgpool-II 3.2-stable head?
> 
> ok, I've done that now, but the same problem is still present.

I seem to find the cause of problem and created patch. Please take a
look at bugtrack #21.

http://www.pgpool.net/mantisbt/view.php?id=21

The patch is here:
http://www.pgpool.net/mantisbt/file_download.php?file_id=28&type=bug
--
Tatsuo Ishii
SRA OSS, Inc. Japan
English: http://www.sraoss.co.jp/index_en.php
Japanese: http://www.sraoss.co.jp