On Thu, Mar 29, 2018 at 2:09 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Claudio Freire <klaussfreire@gmail.com> writes:
>> On Wed, Mar 28, 2018 at 6:54 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> After 0001,
>>> there's no reason to assume that vacuum is particularly likely to get
>>> cancelled between having made cleanups and having updated the upper FSM
>>> levels. (Maybe the odds are a bit more for the no-indexes case, but
>>> that doesn't seem like it's common enough to justify a special mechanism
>>> either.)
>
>> Why not?
>
>> Any kind of DDL (even those that don't rewrite the heap) would cancel
>> autovacuum.
>
>> You might think DDL isn't common enough to worry about, but I've seen
>> cases where regular reindex were required to keep index bloat in check
>> (and were cron'd), and those cancel autovacuum all the time.
>
> If you've got a situation where every vacuum gets canceled partway
> through, you've got bloat problems regardless, because the heap tuples are
> never getting removed in the first place; worrying about whether the FSM
> is up to date is pretty pointless. The 0001 patch basically updates the
> FSM as soon as it can after the tuples are actually deleted, so I think
> we've made the window as small as practical, and I don't really see a need
> to do extra work (and add substantial extra complexity) to deal with
> missed cleanup a bit sooner.
>
> People who are dealing with this sort of scenario a lot might be well
> advised to reduce autovacuum's maintenance_work_mem, so that the cleanup
> cycles happen more often. That's kind of costly in terms of the number
> of index scans, but it reduces the amount of cleanup work that can be
> lost to a cancel.
>
> (I'd also argue that a setup such as you describe is very possibly making
> things worse not better. Perhaps the 0001 patch will go some way towards
> making it less necessary to do that.)
Alright, so we just drop 2.