Re: Different execution plans in PG17 and pgBouncer...

On 5/5/25 13:27, Mladen Marinović wrote:
>
>
> On Mon, May 5, 2025 at 12:07 PM Achilleas Mantzios
> <a(dot)mantzios(at)cloud(dot)gatewaynet(dot)com> wrote:
>
>
> On 5/5/25 11:00, Mladen Marinović wrote:
>>
>>
>> On Mon, May 5, 2025 at 11:24 AM Achilleas Mantzios
>> <a(dot)mantzios(at)cloud(dot)gatewaynet(dot)com> wrote:
>>
>>
>> On 5/5/25 09:52, Mladen Marinović wrote:
>>> Hi,
>>>
>>> We recently migrated our production instances from PG11 to
>>> PG17. While doing so we upgraded our pgBouncer instances
>>> from 1.12 to 1.24. As everything worked on the test servers
>>> we pushed this to production a few weeks ago. We did not
>>> notice any problems until a few days ago (but the problems
>>> were here from the start). The main manifestation of the
>>> problems is a service that runs a fixed query to get a
>>> backlog of unprocessed data (limited to a 1000 rows). When
>>> testing the query using pgAdmin connected directly to the
>>> database we get a result in cca. 20 seconds. The same query
>>> runs for 2 hours when using pgBouncer to connect to the same
>>> database.
>>
>>
>> That's a huge jump, I hope you guys did extensive testing of
>> your app. In which language is your app written? If java,
>> then define prepareThreshold=0 in your jdbc and set
>> max_prepared_statements = 0 in pgbouncer.
>>
>> Mainly python, but the problem was noticed in a java service.
>> Prepare treshold was already set to 0. We changed the
>> max_prepared_statements to 0 from the default (200) but no change
>> was noticed.
>>
>> How about search paths ? any difference on those between the
>> two runs ? Do you set search_path in pgbouncer ? what is
>> "cca." btw ?
>>
>>>
>>> The more interesting part is that when we issue an explain
>>> of the same query we get different plans. We did this a few
>>> seconds apart so there should be no difference in collected
>>> statistics. We ruled out prepared statements, as we
>>> suspected the generic plan might be the problem, but it is
>>> not. Is there any pgBouncer or PG17 parameter that might be
>>> the cause of this?
>>
>>
>> Does this spawn any connections (such as dblink) ? are there
>> limits per user/db pool_size in pgbouncer ?
>>
>> No additional connection nor dbling. Just plain SQL (CTE, SELECT,
>> INSERT, UPDATE, DELETE,...) There are limits, but they are not
>> hit. The query just uses a different plan and runs slower because
>> of that.
>>
>> Pgbouncer, in contrast to its old friend PgPool-II is
>> completely passive, just passes through SQL to the server as
>> fast as possible as it can. But I am sure you know that. Good
>> luck, keep us posted!
>>
>> Yes, that is what puzzles me.
>
> What is the pgbouncer's timeout in the server connections ?
>
> How about "idle in transaction" ? do you get any of those? What's
> the isolation level ?
>
> How about the user ? is this the same user doing pgadmin queries
> VS via the app ?
>
> Can you identify the user under which the problem is manifested and :
>
> ALTER user "unlucky_user" SET log_statement = 'all';
>
> ALTER user "unlucky_user" SET log_min_duration_statement = 0; --
> to help you debug the prepared statements .. just in case , and
> other stuff not printed by log_statement = all.
>
> None of those parameters should affect the fact that when issuing the
> explain select query (the statement is not prepared) from psql
> directly gives a different result than issuing it over the pgbouncer
> connection. The result is repeatable.
>
> We have rolled back pgbouncer to 1.12. and it seems the problem
> persists. This is one of the weirdest things I have ever seen with
> PostgreSQL.

ok, this is something, at least one more extreme thought ruled out. How
about search_path ?  is this the SAME user that is issuing the
statements in pgadmin VS pgbouncer ?

Is there a connect_query inside pgbouncer's conf ?

you have to show all configuration involved and also full logging on the
backend for said user.

>> Regards,
>> Mladen Marinović
>