Re: [Devel 9.2] Index-only scan in count aggregation

El día 18 de abril de 2012 18:17, Robert Haas <robertmhaas(at)gmail(dot)com> escribió:
> On Wed, Apr 18, 2012 at 12:13 PM, Emanuel Calvo <postgres(dot)arg(at)gmail(dot)com> wrote:
>> Hi guys,
>>
>> I'm one of the nightly sources of 9.2devel. I was trying some simple
>> queries and I realized something:
>>
>> stuff=# explain (analyze true, costs true, buffers true, timing true,
>> verbose true) select count(i) from lot_of_values;
>>                                                                QUERY
>> PLAN
>> -------------------------------------------------------------------------------------------------------------------------------------------
>>  Aggregate  (cost=213496.00..213496.01 rows=1 width=4) (actual
>> time=60400.788..60400.791 rows=1 loops=1)
>>   Output: count(i)
>>   Buffers: shared hit=2400 read=86096
>>   ->  Seq Scan on public.lot_of_values  (cost=0.00..188496.00
>> rows=10000000 width=4) (actual time=0.371..32227.791 rows=10000000
>> loops=1)
>>         Output: i, t1, r1, r2, r3, d1
>>         Buffers: shared hit=2400 read=86096
>>  Total runtime: 60402.460 ms
>> (7 rows)
>>
>> stuff=# set enable_seqscan=off;
>> SET
>> stuff=# explain (analyze true, costs true, buffers true, timing true,
>> verbose true) select count(i) from lot_of_values;
>>
>>          QUERY PLAN
>> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>>  Aggregate  (cost=351292.03..351292.04 rows=1 width=4) (actual
>> time=63278.472..63278.475 rows=1 loops=1)
>>   Output: count(i)
>>   Buffers: shared hit=1 read=110379
>>   ->  Index Only Scan using lot_of_values_pkey on
>> public.lot_of_values  (cost=0.00..326292.03 rows=10000000 width=4)
>> (actual time=42.028..35217.460 rows=10000000 loops=1)
>>         Output: i
>>         Heap Fetches: 10000000
>>         Buffers: shared hit=1 read=110379
>>  Total runtime: 63278.720 ms
>> (8 rows)
>>
>>
>> I know, still development. Just wanna know if there will be an
>> improvement for this in the next patches or the idea is to maintain
>> this behaviour.
>>
>> Cheers and thanks for the amazing work you all had done!
>
> I'm not sure what you're unhappy about.  It seems that the query
> planner picked the fastest plan (a sequential scan) and then when you
> disabled that it picked the second-fastest plan (an index-only scan).
>
> The index-only scan would have a chance of beating the sequential scan
> if the table had been recently vacuumed, but not in the case where
> every row is going to require a heap fetch.
>

Oh, I see now. Honestly, I thought it wasn't necessary to make a heap
fetch. The table
doesn't have any modifications, but with the vacuum the cost changed.

Checking the source code, I saw what you are talking about:

/*
* We can skip the heap fetch if the TID references a
heap page on
* which all tuples are known visible to everybody.
In any case,
* we'll use the index tuple not the heap tuple as the
data source.
*/

Thanks for the information!

--
--
Emanuel Calvo