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

On Mon, 2020-03-16 at 14:34 -0700, Andres Freund wrote:
> > > In particularl, I think it'd make sense to *not* have a lower freezing
> > > horizon for insert vacuums (because it *will* cause problems), but if
> > > the page is dirty anyway, then do the freezing even if freeze_min_age
> > > etc would otherwise prevent us from doing so?
> > 
> > I don't quite see why freezing tuples in insert-only tables will cause
> > problems - are you saying that more WAL will be written compared to
> > freezing with a higher freeze_min_age?
> 
> As far as I understand the patch may trigger additional vacuums e.g. for
> tables that have some heavily updated parts / key ranges, and otherwise
> are largely insert only (as long as there are in total considerably more
> inserts than updates). That's not at all uncommon.
> 
> And for the heavily updated regions the additional vacuums with a 0 min
> age could prove to be costly.  I've not looked at the new code, but it'd
> be particularly bad if the changes were to trigger the
> lazy_check_needs_freeze() check in lazy_scan_heap() - it'd have the
> potential for a lot more contention.

I think I got it.

Here is a version of the patch that does *not* freeze more tuples than
normal, except if a prior tuple on the same page is already eligible for freezing.

lazy_check_needs_freeze() is only called for an aggressive vacuum, which
this isn't.

Does that look sane?

Yours,
Laurenz Albe