Recovery conflict monitoring

On Thu, Dec 23, 2010 at 13:09, Magnus Hagander <magnus@hagander.net> wrote:

This patch adds counters and views to monitor hot standby generated
recovery conflicts. It extends the pg_stat_database view with one
column with the total number of conflicts, and also creates a new view
pg_stat_database_conflicts that contains a breakdown of exactly what
caused the conflicts.

Documentation still pending, but comments meanwhile is of course appreciated ;)

Heikki pointed out over IM that it's pointless to count stats caused
by recovery conflict with drop database - since we drop the stats
record as soon as it arrives anyway. Here's an updated patch that
removes that, and also adds some documentation.

--
 Magnus Hagander
 Me: http://www.hagander.net/
 Work: http://www.redpill-linpro.com/

On Mon, 2010-12-27 at 14:39 +0100, Magnus Hagander wrote:

On Thu, Dec 23, 2010 at 13:09, Magnus Hagander <magnus@hagander.net> wrote:

This patch adds counters and views to monitor hot standby generated
recovery conflicts. It extends the pg_stat_database view with one
column with the total number of conflicts, and also creates a new view
pg_stat_database_conflicts that contains a breakdown of exactly what
caused the conflicts.

Documentation still pending, but comments meanwhile is of course appreciated ;)

Heikki pointed out over IM that it's pointless to count stats caused
by recovery conflict with drop database - since we drop the stats
record as soon as it arrives anyway. Here's an updated patch that
removes that, and also adds some documentation.

I like the patch, well inspired, code in the right places AFAICS. No
code comments at all.

Couple of thoughts:

* are we safe to issue stats immediately before issuing FATAL? Won't
some of them get lost?

* Not clear what I'd do with database level information, except worry a
lot. Maybe an option to count conflicts per user would be better, since
at least we'd know exactly who was affected by those. Just an idea.

* Would it better to have a log_standby_conflicts that allowed the
opportunity to log the conflicting SQL, duration until cancelation etc?

I'd rather have what you have than nothing at all though... the new
hot_standby_feedback mode should be acting to reduce these, so it would
be useful to have this patch enabled for testing that feature.

--
Simon Riggs http://www.2ndQuadrant.com/books/
PostgreSQL Development, 24x7 Support, Training and Services

On Mon, Jan 3, 2011 at 00:23, Simon Riggs <simon@2ndquadrant.com> wrote:

On Mon, 2010-12-27 at 14:39 +0100, Magnus Hagander wrote:

On Thu, Dec 23, 2010 at 13:09, Magnus Hagander <magnus@hagander.net> wrote:

This patch adds counters and views to monitor hot standby generated
recovery conflicts. It extends the pg_stat_database view with one
column with the total number of conflicts, and also creates a new view
pg_stat_database_conflicts that contains a breakdown of exactly what
caused the conflicts.

Documentation still pending, but comments meanwhile is of course appreciated ;)

Heikki pointed out over IM that it's pointless to count stats caused
by recovery conflict with drop database - since we drop the stats
record as soon as it arrives anyway. Here's an updated patch that
removes that, and also adds some documentation.

I like the patch, well inspired, code in the right places AFAICS. No
code comments at all.

Thanks for reviewing!

Couple of thoughts:

* are we safe to issue stats immediately before issuing FATAL? Won't
some of them get lost?

They shouldn't - not more than other stats messages. Those are often
flushed from on_shmem_exit() which I think runs even later.

* Not clear what I'd do with database level information, except worry a
lot. Maybe an option to count conflicts per user would be better, since
at least we'd know exactly who was affected by those. Just an idea.

Depends on the usage scenario. In a lot of dedicated environments you
really only have one database - but there are many environments where
you do have multiple and it's quite useful to see them separately. And
you can of course very easily sum() them up for a total count, since
it's a view... Better keep the detail than throw it away, even if that
part isn't useful in *all* cases...

Grouping by user would potentially be helpful - I agree. However, that
goes for most pgstat counters ("number of seqscans", "tuples read" etc
are interesting per user as well in some cases). So doing that right
would mean adding per-user tracking all across pgstats in some smart
way - something we don't do now at all. So I see that as a separate
issue.

* Would it better to have a log_standby_conflicts that allowed the
opportunity to log the conflicting SQL, duration until cancelation etc?

Logging is useful to figure out why you have a certain scenario, yes.
But absolutely not as a *replacement* for the statistics counters, but
as an addition. Just like we have (the now incorrectly named)
pg_stat_bgwriter view *and* log_checkpoints... Different usecases for
the same basic information.

I'd rather have what you have than nothing at all though... the new
hot_standby_feedback mode should be acting to reduce these, so it would
be useful to have this patch enabled for testing that feature.

It will help reduce it, but not take it away, right? Plus, it's an
optional feature...

--
 Magnus Hagander
 Me: http://www.hagander.net/
 Work: http://www.redpill-linpro.com/

On Mon, 2011-01-03 at 10:03 +0100, Magnus Hagander wrote:

I like the patch, well inspired, code in the right places AFAICS. No
code comments at all.

Thanks for reviewing!

All good here. Test and commit please.

--
Simon Riggs http://www.2ndQuadrant.com/books/
PostgreSQL Development, 24x7 Support, Training and Services

Couple of doc suggestions:

--- doc/src/sgml/high-availability.sgml
+     The number of query cancels and the reason for them can be viewed 
using
+     the <structname>pg_stat_database_conflicts</> system view on the slave
+     server.

For compleness sake, this should also mention the per-database summary,
even though I'm not sure how valuable that view is. Also, "on a standby
server" instead of "on the slave server" here. "slave" is mentioned
once as a synonym in high-availability.sgml once, but that's it, and
there can be more than one standby you want to pull these stats from.

*** doc/src/sgml/monitoring.sgml
! number of rows returned, fetched, inserted, updated and deleted, and
! total number of queries cancelled due to conflict with recovery.

This would be clearer if it said you're talking about standby recovery
here, and possibly that this info is only available on the standby. I
could see someone reading this and thinking it's possible for general
database crash recovery to produce cancelled queries, instead of the way
connections are actually blocked until that's done.

! <entry><structname>pg_stat_database_conflicts</>
! <entry>One row per database, showing database OID, database name and
! the number of queries that have been cancelled in this database
due to
! dropped tablespaces, lock timeouts, old snapshots, pinned
buffers and
! deadlocks.

A clarification that you're talking about standby query cancellation
here might be helpful too. I don't think that's necessary for all of
the detailed pg_stat_get_* functions that regular users are less likely
to care about, just these higher level ones.

--
Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD
PostgreSQL Training, Services and Support www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

On Mon, Jan 3, 2011 at 11:35, Greg Smith <greg@2ndquadrant.com> wrote:

Couple of doc suggestions:

--- doc/src/sgml/high-availability.sgml
+     The number of query cancels and the reason for them can be viewed
using
+     the <structname>pg_stat_database_conflicts</> system view on the slave
+     server.

For compleness sake, this should also mention the per-database summary, even
though I'm not sure how valuable that view is.  Also, "on a standby server"
instead of "on the slave server" here.  "slave" is mentioned once as a
synonym in high-availability.sgml once, but that's it, and there can be more
than one standby you want to pull these stats from.

Good point, changed and added.

*** doc/src/sgml/monitoring.sgml
!       number of rows returned, fetched, inserted, updated and deleted, and
!       total number of queries cancelled due to conflict with recovery.

This would be clearer if it said you're talking about standby recovery here,
and possibly that this info is only available on the standby.  I could see
someone reading this and thinking it's possible for general database crash
recovery to produce cancelled queries, instead of the way connections are
actually blocked until that's done.

!       <entry><structname>pg_stat_database_conflicts</>
!       <entry>One row per database, showing database OID, database name and
!       the number of queries that have been cancelled in this database due
to
!       dropped tablespaces, lock timeouts, old snapshots, pinned buffers
and
!       deadlocks.

A clarification that you're talking about standby query cancellation here
might be helpful too.  I don't think that's necessary for all of the
detailed pg_stat_get_* functions that regular users are less likely to care
about, just these higher level ones.

Yeah, those both make sense - I've updated the docs and am running tests ATM.

--
 Magnus Hagander
 Me: http://www.hagander.net/
 Work: http://www.redpill-linpro.com/