Re: [HACKERS] too low cost of Bitmap index scan

On Tue, Dec 20, 2016 at 2:13 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:
> 2016-12-19 23:28 GMT+01:00 Robert Haas <robertmhaas@gmail.com>:
>> On Sat, Dec 17, 2016 at 3:30 AM, Pavel Stehule <pavel.stehule@gmail.com>
>> wrote:
>> > ->  Bitmap Heap Scan on "Zasilka"  (cost=5097.39..5670.64 rows=1
>> > width=12)
>> > (actual time=62.253..62.400 rows=3 loops=231)
>> ...
>> > When I disable bitmap scan, then the query is 6x time faster
>> ....
>> >    ->  Index Scan using "Zasilka_idx_Dopravce" on "Zasilka"
>> > (cost=0.00..30489.04 rows=1 width=12) (actual time=15.651..17.187 rows=3
>> > loops=231)
>> >         Index Cond: ("Dopravce" = "Dopravce_Ridic_1"."ID")
>> >        Filter: (("StavDatum" > (now() - '10 days'::interval)) AND
>> > (("Stav" =
>> > 34) OR ("Stav" = 29) OR ("Stav" = 180) OR ("Stav" = 213) OR ("Stav" =
>> > 46) OR
>> > (("Produkt" = 33) AND ("Stav" = 179)) OR ((("ZpetnaZasilka" =
>> > '-1'::integer)
>> > OR ("PrimaZasilka" = '-1'::integer)) AND ("Stav" = 40))))
>> >         Rows Removed by Filter: 7596
>>
>> I'm not sure, but my guess would be that the query planner isn't
>> getting a very accurate selectivity estimate for that giant filter
>> condition, and that's why the cost estimate is off.
>
> maybe operator cost is too high?

Hmm, seems like you'd be paying the operator cost either way.  No?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company