Re: Experimental ARC implementation

Bruce Momjian wrote:

> Greg Stark wrote:
>>
>> Bruce Momjian <pgman(at)candle(dot)pha(dot)pa(dot)us> writes:
>>
>> > Have you considered having the background writer check the pages it is
>> > about to write to see if they can be added to the FSM, thereby reducing
>> > the need for vacuum? Seems we would need to add a statistics parameter
>> > so pg_autovacuum would know how many tuples the background write added
>> > to the freespace map, so it doesn't vacuum a table that doesn't need it.
>>
>> This would suffer from the previously mentioned problem of having to pull in
>> index pages and dirty them when it's trying to flush and clean pages.
>
> I am confused why the index would be involved in this.

Looks Greg mixed up background vacuuming with background fsming, the
problem is a vacuum-some-page-at-hand problem, not related to FSM.

>
>> Conceivably it could just count up the dead tuples and provide that
>> information to something like pg_autovacuum so it knows when it's time to run
>> a vacuum. I don't see that as all that much of a win over the current
>> heuristics. At best it means a big batch update will trigger a vacuum sooner
>> so you don't have to manually run vacuum to avoid overflowing the fsm.
>
> Yea, probably. Another idea would be for tuple reuse to favor pages
> already in the cache so it doesn't have to read in a page from disk.

It has little chance to remember how many dead tuples where added
between this flush and it's corresponding read, or more precise the last
time this buffer got flushed. And we certainly don't want to modify the
on-disk block structure just for that.

Jan

--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck(at)Yahoo(dot)com #