Re: Performance regression between 8.3 and 8.4 on heavy text indexing

Hi Gaël,

On Fri, Aug 21, 2009 at 3:37 PM, Gaël Le Mignot<gael@pilotsystems.net> wrote:
> With 8.3 ::
>
>  Limit  (cost=752.67..752.67 rows=1 width=24)
>  (11 rows)
>
> With 8.4 ::
>  (8 rows)

Could you provide us the EXPLAIN *ANALYZE* output of both plans?

From what I can see, one of the difference is that the estimates of
the number of rows are / 3 for this part of the query:
8.3 ->  Bitmap Index Scan on keywords_index  (cost=0.00..48.97 rows=574 width=0)
8.4 ->  Bitmap Index Scan on keywords_index  (cost=0.00..14.03 rows=192 width=0)

It might be interesting to see if 8.4 is right or not.

Before 8.4, the selectivity for full text search was a constant (as
you can see it in your 8.3 plan: the number of rows are equal in both
bitmap index scan). 8.4 is smarter which might lead to other plans.

--
Guillaume