Re: Planning to force reindex of hash indexes

Tom Lane wrote:
> I've found a number of infelicities in the hash index code that can't be
> fixed without an on-disk format change. The biggest one is that the
> hashm_ntuples field in hash meta pages is only uint32, meaning that
> hash index space management will become confused if the number of
> entries exceeds 4G. I'd like to change it to a "double", and clean up
> a couple other uglinesses at the same time.
>
> Ordinarily I'd just force an initdb for such a change, but at this late
> stage of the 7.4 cycle it seems better to avoid requiring initdb,
> especially since many beta testers wouldn't be using hash indexes anyway
> and shouldn't need to reload. What I intend to do instead is increment
> the version number that already exists in the hash metapage, and add
> code to spit out a "please reindex this index" error if the version
> number isn't right. A REINDEX command would be sufficient to
> reconstruct the index in the new format.
>
> Any objections?

Good plan.

--
Bruce Momjian | http://candle.pha.pa.us
pgman(at)candle(dot)pha(dot)pa(dot)us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073