Csaba Nagy wrote:
> On Wed, 2007-07-04 at 00:53, Christopher Browne wrote:
>   
>> One at a time took about 15 minutes
>> 100 at a time took 3:15
>> 1000 at a time took longer than 100 at a time (curious, that!)
>> all in one shot took 43 seconds.
>>     
>
> Check the plans for the 100 vs. 1000 cases: I'm pretty sure 100 goes for
> bitmap index scan and 1000 goes for sequential scan... and 10 * 100
> bitmap index scans are probably somewhat faster than 1 sequential scan
> on your table/box. I guess 1000 is close to the limit between the
> performance turnover between the index scan and sequential scan on your
> table/box/setup, but the sequential scan is slightly underestimated by
> the planner.
>
> BTW, the bitmap index scan case should theoretically be the fastest, so
> aiming for the highest chunk size where the planner still chooses bitmap
> index scan (or downright forcing it to do so if possible) would give the
> best performance.
>   
Interestingly, the query plans for 100 and 1000 are much the same:

-  Bitmap Heap Scan on foo atop
    100 (or 1000) ORed Bitmap Index Scans.

I was fully expecting to discover some differing query plan; it was 
quite a surprise to discover both:

a) That the plans were essentially identical, and
b) That it's more expensive to delete 1K tuples at a shot than 100.