Dne 20.1.2011 09:10, Heikki Linnakangas napsal(a):
> It seems that the suggested multi-column selectivity estimator would be
> more sensitive to ndistinct of the individual columns. Is that correct?
> How is it biased? If we routinely under-estimate ndistinct of individual
> columns, for example, does the bias accumulate or cancel itself in the
> multi-column estimate?
>
> I'd like to see some testing of the suggested selectivity estimator with
> the ndistinct estimates we have. Who knows, maybe it works fine in
> practice.
The estimator for two columns and query 'A=a AND B=b' is about
0.5 * (dist(A)/dist(A,B) * Prob(A=a) + dist(B)/dist(A,B) * Prob(B=b))
so it's quite simple. It's not that sensitive to errors or ndistinct
estimates for individual columns, but the problem is in the multi-column
ndistinct estimates. It's very likely that with dependent colunms (e.g.
with the ZIP codes / cities) the distribution is so pathological that
the sampling-based estimate will be very off.
I guess this was a way too short analysis, but if you can provide more
details of the expected tests etc. I'll be happy to provide that.
regards
Tomas