On Mon, Nov 9, 2015 at 2:42 AM, Simon Riggs <si...@2ndquadrant.com> wrote:
> On 9 November 2015 at 10:08, <kawami...@tkl.iis.u-tokyo.ac.jp> wrote:
>> We guessed the cause of this error would be in the cost model of Postgres,
>> and investigated the source code of optimizer, and we found the cause of
>> this problem. It was in the index cost estimation process. On scanning inner
>> table, if loop count is greater than 1, its I/O cost is counted as random
>> access. In the case of Query2, in one loop (i.e. one inner table scan) ,
>> much data is read sequentially with clustered index, so it seems to be wrong
>> that optimizer thinks its I/O workload is random access.
> We don't have a clustered index in Postgres. We do store correlation stats
> for the index, which is used in some places to reduce cost.
> Could you look some more at this and see if a model change that uses the
> correlation can improve this?
That is already happening. min_IO_cost is set on the assumption of
perfect correlation, max_IO_cost is set on the assumption of no
correlation, and then later in the code it interpolates between these
two based on the observed correlation:
run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
But the min_IO_cost is very pessimistic in this particular case. So I
think that their proposed change is already exactly what you are
asking them to try.
Sent via pgsql-hackers mailing list (email@example.com)
To make changes to your subscription: