Mark kirkwood <markir@slingshot.co.nz> writes:
> There is slightly odd behaviour with the frequencies decreasing with
> increasing number of quantiles (same as 7.2 .. same code here ?).
That does seem curious. With the inevitable sampling error, you'd
expect that some values would be sampled at a bit more than their
true frequency, and others at a bit less. The oversampled ones would
be the ones to get into the MCV list. But what you've got here is
that even the most-commonly-sampled value showed up at a bit less
than its true frequency. Is this repeatable if you do ANALYZE over
and over? Maybe it was just a statistical fluke.
> I am wondering if this is caused by my example not having any "real" most
> common values (they are all as common as each other).
> I am going to fiddle with my data generation script, skew the
> distribution and see what effect that has.
Someone else reported some results that made it look like a logarithmic
frequency distribution was a difficult case for the stats gatherer:
http://archives.postgresql.org/pgsql-general/2002-03/msg01300.php
So please be sure to try that.
regards, tom lane