Re: Berserk Autovacuum (let's save next Mandrill)

On Thu, 2020-03-05 at 19:40 +1300, David Rowley wrote:
> I think we need to move forward with doing something to cope with
> INSERT-only tables not being auto-vacuumed.
> 
> I think the patch you have is something along the lines to what I'd
> have imagined we should do. However, there are a few things that I'd
> do a different way.

Thanks for the review, and that's good news.

> 1. I'd go for 2 new GUCs and reloptions.
> autovacuum_vacuum_insert_threshold (you're currently calling this
> autovacuum_vacuum_insert_limit. I don't see why the word "limit" is
> relevant here). The other GUC I think should be named
> autovacuum_vacuum_insert_scale_factor and these should work exactly
> the same way as autovacuum_vacuum_threshold and
> autovacuum_vacuum_scale_factor, but be applied in a similar way to the
> vacuum settings, but only be applied after we've checked to ensure the
> table is not otherwise eligible to be vacuumed.

Yes, "threshold" is better than "limit" I have renamed the GUC and
the reloption.

I disagree about the scale_factor (and have not added it to the
updated version of the patch).  If we have a scale_factor, then the
time between successive autovacuum runs would increase as the table
gets bigger, which defeats the purpose of reducing the impact of each
autovacuum run.

Since autovacuum skips pages where it has nothing to do, we can expect
that runs on a large table won't be much more expensive than runs on a
smaller table, right?

> 3. The name "insert_only" does not seem the best for the new boolean
> variable that you're using in various places.  That name seems to be
> too closely related to our current intended use case. Maybe
> skip_index_cleanup is more to the point.

I originally called the variable "skip_indexes", but when I decided
that such vacuum runs also aggressively freeze the table, I thought
that the name was misleading and renamed it.

I won't put up a fight about this, though.

> 4. Are you sure you mean "Maximum" here?  Isn't it the minimum? At
> least it will be once you add both options. Otherwise, I think Maximum
> is not the correct word. Perhaps "The threshold"
> 
> + {"autovacuum_vacuum_insert_limit", PGC_SIGHUP, AUTOVACUUM,
> + gettext_noop("Maximum number of tuple inserts prior to vacuum."),
> + NULL
> + },

I had actually been debating whether to use "maximum" or "minimum".
I realize now that this strange uncertainty stems from the fact that
there is (yet) only a single parameter to govern this.

The updated patch desctibes the GUC as
"Number of tuple inserts prior to vacuum."

> 5. I think the new field in this struct should be named vacuum_insert_threshold
> 
> @@ -252,6 +252,7 @@ typedef struct AutoVacOpts
>  {
>   bool enabled;
>   int vacuum_threshold;
> + int vacuum_ins_limit;

I agree as above, renamed.

> 6. Are you sure you meant to default this to 50?
> 
> index e58e4788a8..9d96d58ed2 100644
> --- a/src/backend/utils/misc/postgresql.conf.sample
> +++ b/src/backend/utils/misc/postgresql.conf.sample
> @@ -598,6 +598,8 @@
>  #autovacuum_naptime = 1min # time between autovacuum runs
>  #autovacuum_vacuum_threshold = 50 # min number of row updates before
>   # vacuum
> +#autovacuum_vacuum_insert_limit = 50 # max number of row inserts before
> + # vacuum
> 
> Seems excessive given there's no scale factor in the current patch.

That was a mistake.
I chose 10000000 as the actual default value, but forgot to put the
same value into "postgresql.conf".

> 7. I know you know.... missing docs... would be good to get those.

The updated version of the patch has documentation.

I just wanted to get a feeling if my patch would be killed cold before
I went to the effort of writing documentation.

> 8. Should we care when setting the insert counter back to 0 if
> auto-vacuum has skipped pages?

Since this is only an approximate value anyway, I decided not to care.
I don't know if that is acceptable.

> 9. You should add a new column to the pg_stat_all_tables view to allow
> visibility of the insert since the last vacuum.  The column should be
> named n_ins_since_vacuum. This seems like the best combination of
> n_mod_since_analyze and n_tup_ins.

Done.

> 10. I'm slightly worried about the case where we don't quite trigger a
> normal vacuum but trigger a vacuum due to INSERTs then skip cleaning
> up the indexes but proceed to leave dead index entries causing indexes
> to become bloated.  It does not seem impossible that given the right
> balance of INSERTs and UPDATE/DELETEs that this could happen every
> time and the indexes would just become larger and larger.

I understand.

This might particularly be a problem with larger tables, where
a normal autovacuum is rare because of the scale_factor.

Perhaps we can take care of the problem by *not* skipping index
cleanup if "changes_since_analyze" is substantially greater than 0.

What do you think?

> 11. We probably do also need to debate if we want this on or off by
> default.   I'd have leaned towards enabling by default if I'd not
> personally witnessed the fact that people rarely* increase auto-vacuum
> to run faster than the standard cost settings. I've seen hundreds of
> servers over the years with all workers busy for days on something
> they'll never finish quickly enough.  We increased those settings 10x
> in PG12, so there will be fewer people around suffering from that now,
> but even after having reduced the vacuum_cost_delay x10 over the PG11
> settings, it's by no means fast enough for everyone.  I've mixed
> feelings about giving auto-vacuum more work to do for those people, so
> perhaps the best option is to keep this off by default so as not to
> affect the people who don't tune auto-vacuum.  They'll just suffer the
> pain all at once when they hit max freeze age instead of more
> gradually with the additional load on the workers.   At least adding
> this feature gives the people who do tune auto-vacuum some ability to
> handle read-only tables in some sane way.
> 
> An alternative way of doing it would be to set the threshold to some
> number of million tuples and set the scale_factor to 0.2 so that it
> only has an effect on larger tables, of which generally people only
> have a smallish number of.

Yes, I think that disabling this by default defeats the purpose.

Knowledgeable people can avoid the problem today by manually scheduling
VACUUM runs on insert-only tables, and the functionality proposed here
is specifically to improve the lives of people who don't know enough
to tune autovacuum.

My original idea was to set the threshold to 10 million and have no scale
factor.


Updated patch attached.

Yours,
Laurenz Albe