Re: pgpool-II: cannot use serializable mode in a hot standby

On Wed, Apr 13, 2016 at 1:30 AM, Alexander Pyhalov <alp(at)rsu(dot)ru> wrote:
> Tatsuo Ishii писал 13.04.2016 02:36:
>>>
>>> On 04/12/2016 16:50, Adrian Klaver wrote:
>>>>
>>>> On 04/12/2016 01:06 AM, Alexander Pyhalov wrote:
>>>
>>> I understand. I mean perhaps pgpool shouldn't forward these statements
>>> to slaves.

It should not forward transactions which are requested to be
SERIALIZABLE to standbys. If you just suppress the SET statement
(or substitute REPEATABLE READ), queries in that transaction can
return incorrect results.

>> Yeah, PostgreSQL used to accept the command on standbys (at least in
>> 9.0). The restriction was added later on.

... in 9.1, for a reason.

>> It woule be nice if you send
>> a bug report to the pgpool-II bug tracker to not forget it.
>>
>> http://pgpool.net/mediawiki/index.php/Bug_tracking_system
>
> Filed http://www.pgpool.net/mantisbt/view.php?id=191

As the entry stands at the moment, the suggestions for fixes will
allow incorrect query results. See this example, and imagine that
the transaction generating the list of receipts for the closed
batch is run on the standby before the transaction adding the last
receipt commits. Or test it.

--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company