Re: 9.2.1 & index-only scans : abnormal heap fetches after VACUUM FULL

On Tue, Dec 3, 2013 at 11:25 AM, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> On Thu, Nov 28, 2013 at 05:38:05PM -0500, Bruce Momjian wrote:
>> > I wonder if we ought to mark each page as all-visible in
>> > raw_heap_insert() when we first initialize it, and then clear the flag
>> > when we come across a tuple that isn't all-visible. We could try to
>> > set hint bits on the tuple before placing it on the page, too, though
>> > I'm not sure of the details.
>>
>> I went with the per-page approach because I wanted to re-use the vacuum
>> lazy function. Is there some other code that does this already? I am
>> trying to avoid yet-another set of routines that would need to be
>> maintained or could be buggy. This hit bit setting is tricky.
>>
>> And thanks much for the review!
>
> So, should I put this in the next commit fest?

+1.

> I still have an unknown
> about the buffer number to use here:
>
> ! /* XXX use 0 or real offset? */
> ! ItemPointerSet(&(tuple.t_self), BufferIsValid(buf) ?
> ! BufferGetBlockNumber(buf) : 0, offnum);
>
> Is everyone else OK with this approach? Updated patch attached.

I started looking at this a little more the other day but got bogged
down in other things before I got through it all. I think we're going
to want to revise this so that it doesn't go through the functions
that current assume that they're always deal with a shared_buffer, but
I haven't figured out exactly what the most elegant way of doing that
is yet.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company