On Thu, Mar 20, 2014 at 10:45 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> So you might think that the problem here is that we're assuming
>> uniform density. Let's say there are a million rows in the table, and
>> there are 100 that match our criteria, so the first one is going to
>> happen 1/10,000'th of the way through the table. Thus we set SC =
>> 0.0001 * TC, and that turns out to be an underestimate if the
>> distribution isn't as favorable as we're hoping. However, that is NOT
>> what we are doing. What we are doing is setting SC = 0. I mean, not
>> quite 0, but yeah, effectively 0. Essentially we're assuming that no
>> matter how selective the filter condition may be, we assume that it
>> will match *the very first row*.
>
> I think this is wrong. Yeah, the SC may be 0 or near it, but the time to
> fetch the first tuple is estimated as SC + (TC-SC)/N.
Hmm, you're right, and experimentation confirms that the total cost of
the limit comes out to about TC/selectivity. So scratch that theory.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company