Re: Question on session timeout

On Tue, Oct 1, 2024 at 4:10 AM Adrian Klaver <adrian(dot)klaver(at)aklaver(dot)com>
wrote:

> On 9/30/24 13:01, sud wrote:
> > Hello,
> > We are frequently seeing the total DB connection reaching ~2000+ whereas
>
> Where are you getting the ~2000 count from?
>
> > the total number of active sessions in pg_stat_activity staying <100 at
> > any point in time. And when we see the sessions from DB side they are
> > showing most of the sessions with state as 'idle' having
> > backend_start/xact_start showing date ~10days older. We do use
> > application level connection pooling, and we have ~120 sets as both the
>
> What do you mean by ~120 sets, in particular what is a set?
>
> > "max idle" and "max active" connection count and "maxage" as 7 days, so
> > does this suggest any issue at connection pool setup?
>
> Using what pooler?
>
> >
> > We do see keep alive queries in the DB (select 1), not sure if that is
> > making this scenario. When checking the
>
> How often do to keep alive queries run?
>
> > "idle_in_transaction_session_timeout" it is set as 24hours and
> > "idle_session_timeout" set as "0". So my question is , should we set the
> > parameter to a lesser value in DB cluster level like ~5minutes or so, so
> > as not to keep the idle sessions lying so long in the database and what
>
> '"idle_in_transaction_session_timeout" it is set as 24hours' is a foot
> gun as explained here:
>
> https://www.postgresql.org/docs/current/runtime-config-client.html
>
> idle_in_transaction_session_timeout (integer)
>
> [...]
>
> "This option can be used to ensure that idle sessions do not hold locks
> for an unreasonable amount of time. Even when no significant locks are
> held, an open transaction prevents vacuuming away recently-dead tuples
> that may be visible only to this transaction; so remaining idle for a
> long time can contribute to table bloat. See Section 24.1 for more
> details."
>
>
> With '"idle_session_timeout" set as "0"' a session without an open
> transaction is not going to timeout.
>
>
>

*Where are you getting the ~2000 count from?*
Seeing this in the "performance insights" dashboard and also its matching
when I query the count of sessions from pg_stat_activity.

*What do you mean by ~120 sets, in particular what is a set?*These are the
values set as mentioned in the properties file which the application team
uses for connection pooling.

*Using what pooler?*I need to check on this as Its Java application(jdbc
driver for connecting to DB), so I thought it must be using standard
connection pooling. Will double check.

*How often do to keep alive queries run?*Need to check. But I am not sure,
in general , if these "keep alive" queries are used for keeping a
transaction alive or a session alive?

As you described, a long time open transaction with a session state as
"idle" will be threatening as that will cause locking and "transaction ID
wrap around" issues to surface whereas having "idle sessions" of a closed
transaction may not cause any issue as they will do no harm. Does it mean
we can have any number of idle sessions or we should also have some non
zero "timeout" setup for the "ide_session_timeout" parameter too
(maybe ~1hr or so)?

Is it correct to assume the session in pg_stat_activity with very old
XACT_START are the one which are part of long running open transaction(i.e.
driven by idle_in_transaction_session_timeout) whereas the ones with older
BACKEND_START or QUERY_START are the one are just the idle session(driven
by idle_session_timeout) but not tied to any open transaction?

Few observations:-

I do see, "MaximumUsedTransactionIDs" staying consistently ~200M for a long
time then coming down. And its matching to "autovacuum_freeze_max_age"
which is set as 200M. Hope it's expected. We have max_connections set as
5000.

"Database connection" touching ~2000 then coming down till 200. And we see
uneven spikes in those, it seems to be matching with the pattern , when we
have some errors occurring during the insert queries which are submitted by
the Java application to insert the data into the tables.