On Fri, Sep 4, 2009 at 9:37 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Fri, Sep 4, 2009 at 4:01 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
>>> Hmm ... reading that over again, it seems like there is a pretty
>>> obvious solution.
>
>> This doesn't seem totally horrible. But, before you go do it, do we
>> have a clearly-defined plan for the rest of the project?
>
> Rest of what project? Removing vacuum full isn't a necessary component
> of that. It would enable doing CLUSTER on pg_class, and it would
> eliminate the crock of REINDEX having to reindex shared indexes
> in-place. It could probably be justified even without any changes in
> our approach to vacuum.
OK, I'm sold.
>> ... only need this if we're absolutely confident that rewriting the table
>> in place is just not an option worth keeping around. It's unclear to
>> me that everyone is convinced of that, and even if they are, it's
>> unclear to me what we plan to implement instead.
>
> I thought we were pretty well agreed that a seqscan variant of
> CLUSTER would be worth doing. Whether we take the next step by
> eliminating vacuum full is a different question, but the shape of
> the substitute seems perfectly clear.
Well, there were some other ideas discussed, but perhaps that's the
only one that had a clear consensus.
...Robert