On 7 March 2016 at 16:44, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Geoff Winkless <pgsqladmin@geoff.dj> writes:
>> But as far as I can see, apart from the absolute extremes, the
>> index-only scan is _always_ going to be quicker than the index+table
>> scan.
>
> Well, that is a different issue: what does the planner think of an
> index-only scan as compared to a regular index scan. I suspect that
> it's pricing the IOS very high because a lot of the table is dirty
> and therefore will have to be visited even in a nominally index-only
> scan. You might check whether the plan choice changes immediately
> after a VACUUM of the table.
I ran VACUUM FULL and VACUUM ANALYZE. It made no difference. I would
have thought that if it were the case then the equality-test queries
would suffer from the same problem anyway, no?
Even being fairly kind and selecting an scdate range that's only 1%
into the set the query takes over 4 times the amount of time taken by
the indexed query - so the "best" range for the index+table method is
utterly tiny - it would be reasonable only when the scdate field is
uniformly distributed, which even in a table without correlation
between the fields is likely to be almost never.
Geoff
