pg_ctlcluster uses stop immediate on it's own

[Decibel]

Binary package hint: postgresql-8.1

When issued a restart command, pg_ctlcluster will perform a stop immediate without being told to do so. An immediate stop should only be used if an administrator specifically requests it.

[Martin Pitt (pitti)]

What do you propose instead? "smart" is not appropriate, since it needs an interminate waiting time. It uses "fast" by default to get a graceful shutdown. Only if that does not work, "immediate" is used as a fallback.

[Decibel (decibel)]

If fast fails to shut the database down that means that something is seriously wrong, and pg_ctlcluster should exit with an error.

Perhaps it would be best to allow the shutdown mode as an option to the stop command.

I also disagree with changing the default PostgreSQL behavior without a very good reason. It would be better if the default stop behavior was the same as it is in PostgreSQL, though I guess that would mean having a different command for system shutdown to use, which probably won't happen...

[Martin Pitt (pitti)]

So I'll have pg_ctlcluster use the default mode (smart) by default, and add a --force option which the init script will use to ensure a best-effort shutdown.

[Decibel (decibel)]

So --force would do an immediate shutdown then?

[Martin Pitt (pitti)]

Hi,

Decibel [2007-11-11 17:51 -0000]:
> So --force would do an immediate shutdown then?

No, in the current implementation it uses 'fast', and only if that
fails, uses 'immediate'.

[Decibel (decibel)]

On Nov 12, 2007, at 2:58 AM, Martin Pitt wrote:
> Decibel [2007-11-11 17:51 -0000]:
>> So --force would do an immediate shutdown then?
>
> No, in the current implementation it uses 'fast', and only if that
> fails, uses 'immediate'.

I don't see what that buys us. If a fast shutdown fails, that means
there's something seriously wrong, and the administrator should be
given the chance to investigate. They can then issue a kill -9.

What if the script printed out an appropriate sudo kill -9 $PID if
fast shutdown failed?
--
Decibel!, aka Jim C. Nasby, Database Architect <email address hidden>
Give your computer some brain candy! www.distributed.net Team #1828

[Martin Pitt (pitti)]

Hi,

Decibel [2007-11-12 15:29 -0000]:
> >> So --force would do an immediate shutdown then?
> >
> > No, in the current implementation it uses 'fast', and only if that
> > fails, uses 'immediate'.
>
> I don't see what that buys us.

Maybe there is a misunderstanding: pg_ctlcluster uses "smart" by
default now, so if you call it directly, or call the init script with
'restart', then "smart" is used. fast/immediate is only used in the
init script, thus usually when you shutdown the machine. In that case
it is preferable *not* to wait for clients, otherwise you stall the
reboot/shutdown.

[Decibel (decibel)]

On Mon, Nov 12, 2007 at 04:03:48PM -0000, Martin Pitt wrote:
> Hi,
>
> Decibel [2007-11-12 15:29 -0000]:
> > >> So --force would do an immediate shutdown then?
> > >
> > > No, in the current implementation it uses 'fast', and only if that
> > > fails, uses 'immediate'.
> >
> > I don't see what that buys us.
>
> Maybe there is a misunderstanding: pg_ctlcluster uses "smart" by
> default now, so if you call it directly, or call the init script with
> 'restart', then "smart" is used. fast/immediate is only used in the
> init script, thus usually when you shutdown the machine. In that case
> it is preferable *not* to wait for clients, otherwise you stall the
> reboot/shutdown.

No, I understand. But unless the init scripts will absolutely hang a
shutdown if they don't exit then there's no reason to fall through to
immediate. immediate == kill -9, which is what init will do anyway as it
dies.

The problem is that using fast is a completely valid use-case outside of
init, and in that case it's a really bad idea to automatically fall
through to immediate. If fast doesn't work we should leave things as-is
so the admin can investigate. If you want to be admin-friendly, have the
script print out an appropriate sudo command that can be cut and pasted.
--
Decibel!, aka Jim C. Nasby, Database Architect <email address hidden>
Give your computer some brain candy! www.distributed.net Team #1828

[Martin Pitt (pitti)]

Fixed in Hardy in this version:

postgresql-common (82) unstable; urgency=low

  * debian/control: Fix 'commmon' typo. (Closes: #449532)
  * t/030_errors.t: Check that an /etc/postgresql/postgresql.conf leftover is
    not regarded as a cluster directory. This reproduces #450565.
  * PgCommon.pm, get_versions()/get_version_clusters(): Filter out '.' and
    '..' from readdir() output. (Closes: #450565)
  * pg_ctlcluster: Use "smart" mode shutdown by default, and add option
    -f/--force to use "fast" -> "immediate" -> kill -9 approach. (LP: #154012)
  * debian/init.d-functions: Use --force when stopping all clusters, since
    default smart mode can potentially block forever and thus fail to shut
    down a machine.
  * debian/README.Devel: Update for current bzr versions and hosting on
    code.launchpad.net.
  * Update Japanese debconf translations (Closes: #450902)

 -- Martin Pitt <email address hidden> Sun, 30 Dec 2007 22:27:08 +0100

[Decibel (decibel)]

Cool, thanks!

On Sun, Feb 03, 2008 at 11:49:33AM -0000, Martin Pitt wrote:
> Fixed in Hardy in this version:
>
> postgresql-common (82) unstable; urgency=low
>
> * debian/control: Fix 'commmon' typo. (Closes: #449532)
> * t/030_errors.t: Check that an /etc/postgresql/postgresql.conf leftover is
> not regarded as a cluster directory. This reproduces #450565.
> * PgCommon.pm, get_versions()/get_version_clusters(): Filter out '.' and
> '..' from readdir() output. (Closes: #450565)
> * pg_ctlcluster: Use "smart" mode shutdown by default, and add option
> -f/--force to use "fast" -> "immediate" -> kill -9 approach. (LP: #154012)
> * debian/init.d-functions: Use --force when stopping all clusters, since
> default smart mode can potentially block forever and thus fail to shut
> down a machine.
> * debian/README.Devel: Update for current bzr versions and hosting on
> code.launchpad.net.
> * Update Japanese debconf translations (Closes: #450902)
>
> -- Martin Pitt <email address hidden> Sun, 30 Dec 2007 22:27:08 +0100
>
>
> ** Changed in: postgresql-common (Ubuntu)
> Status: Triaged => Fix Released
>
> --
> pg_ctlcluster uses stop immediate on it's own
> https://bugs.launchpad.net/bugs/154012
> You received this bug notification because you are a direct subscriber
> of the bug.
>

--
Decibel!, aka Jim C. Nasby, Database Architect <email address hidden>
Give your computer some brain candy! www.distributed.net Team #1828