Re: Fixing GIN for empty/null/full-scan cases

On Jan 14, 2011, at 5:54 PM, Tom Lane wrote:

> "David E. Wheeler" <david@kineticode.com> writes:
>> So some questions:
> 
>> * Is something seriously wrong with GiST index creation on integer[] columns?
> 
>> * Why does GIN performance appear to be no better than table scans on integer[] columns?
> 
>> * Why does it take 3-4x longer to create the GIN than the GiST index on tsvector? I thought that GIN was supposed to be faster to update
> 
> Hard to comment on any of this without a concrete example (including
> data) to look at.  Given the bugs we've recently found in the picksplit
> algorithms for other contrib modules, I wouldn't be too surprised if the
> sucky GiST performance traced to a similar bug in intarray.  But I'm not
> excited about devising my own test case.

I could give you access to the box in question if you'd like to poke at it. Send me a public key.

> One other point here is that GIN index build time is quite sensitive to
> maintenance_work_mem --- what did you have that set to?

64MB

Best,

David