Обсуждение: Routine maintenance - vacuum, analyse and autovacuum

On Thu, Jan 29, 2004 at 18:59:28 +0900,
  Iain <iain@mst.co.jp> wrote:

I haven't used auto_vacuum, but I seen comments on the lists indicating
that it works OK.

> If I start a VACUUM FULL, is it ok to interrupt it part way through? I mean,
> is a partially completed vacuum full worth anything, and is there any danger
> in routinely killing it?

With 7.4 you shouldn't have to do VACUUM FULL. If you have FSM set high
enough, then normal VACUUM should suffice. The exception would be if
you ran something that may the DB much larger than normal and you wanted
to shrink the space used by the DB. Then you would need to run a VACUUM FULL.

"Iain" <iain@mst.co.jp> writes:
> If we use it, can we stop using nightly vacuums? I'd like to do this if
> possible, since the current version (it's a 7.1 db) of the system runs a
> number of batch processes at night (large deletes and large inserts), and
> finishes with a vacuum analyse. The vacuum takes quite a long time and the
> nightly window is gradually getting tighter and tighter.

Just moving from 7.1 to 7.4 should help, since you can start using plain
vacuum (7.1's vacuum is equivalent to vacuum full IIRC).

> If I start a VACUUM FULL, is it ok to interrupt it part way through? I mean,
> is a partially completed vacuum full worth anything, and is there any danger
> in routinely killing it?

It's crash-safe but you may leave your indexes bigger than before :-(

            regards, tom lane

> Just moving from 7.1 to 7.4 should help, since you can start using plain
> vacuum (7.1's vacuum is equivalent to vacuum full IIRC).

That old DB server has been running contnuously for 2 years they tell me.
Apparently it is also growing despite nightly vacuums and a stable data set
size (old data is purged dai.ly and monthly). Nightly delete/insert batches
and vacuums are taking longer and longer too. I suspect the vacuum isn't
doing what they think - possibly due to the FSM being at the default, and
possibly due to ghost processes holding old record versions open. It has
been stable nonetheless. Anyway, I'm sure we can do much better in 7.4.

> > If I start a VACUUM FULL, is it ok to interrupt it part way through? I
mean,
> > is a partially completed vacuum full worth anything, and is there any
danger
> > in routinely killing it?
>
> It's crash-safe but you may leave your indexes bigger than before :-(

OK, that's the kind of thing I wanted to know. Basically, partial runs of
vacuum full are not likely to be of any benefit. So, unless you intend to
let it run to completion, best to avoid it.

Given that our system experiences quite heavy usage during the day, and has
a nightly window for batch processing I'm wondering if the best way to go
about this is to give auto vacuum a miss and just do a standard vacuum after
the batch processes have completed. As long as the FSM is big enough to
handle the amount of data changed in 1 day, it would be OK, right?

Thanks for the input, (and Bruno too)
Iain

"Iain" <iain@mst.co.jp> writes:
>> Just moving from 7.1 to 7.4 should help, since you can start using plain
>> vacuum (7.1's vacuum is equivalent to vacuum full IIRC).

> That old DB server has been running contnuously for 2 years they tell me.
> Apparently it is also growing despite nightly vacuums and a stable data set
> size (old data is purged dai.ly and monthly). Nightly delete/insert batches
> and vacuums are taking longer and longer too. I suspect the vacuum isn't
> doing what they think - possibly due to the FSM being at the default, and
> possibly due to ghost processes holding old record versions open. It has
> been stable nonetheless. Anyway, I'm sure we can do much better in 7.4.

Kinda sounds like index bloat to me --- have you done any looking to
determine which files are growing?

> OK, that's the kind of thing I wanted to know. Basically, partial runs of
> vacuum full are not likely to be of any benefit. So, unless you intend to
> let it run to completion, best to avoid it.

Right.

> Given that our system experiences quite heavy usage during the day, and has
> a nightly window for batch processing I'm wondering if the best way to go
> about this is to give auto vacuum a miss and just do a standard vacuum after
> the batch processes have completed. As long as the FSM is big enough to
> handle the amount of data changed in 1 day, it would be OK, right?

Yes, though plain VACUUM is not as much of a drag on performance as
VACUUM FULL (== 7.1 VACUUM).  So you might want to experiment and
see if you can get away with it during the day.  Also note there is work
afoot to reduce the VACUUM performance drag even further in 7.5.

            regards, tom lane

> Kinda sounds like index bloat to me --- have you done any looking to
> determine which files are growing?
>

No, I havn't. I don't actually have access to that machine, and there is a
cirtain reluctance to tamper with it. However Ithink that it will need some
attention soon. I'll keepin mind what you said.

regards
iain