Bruce Momjian <pgman@candle.pha.pa.us> writes:
> Also, Tom, should we preload the disbursion buckets with most common
> value from the previous run on the assumption we will get a better
> value?
Hmm, I don't see why that'd make it better. It'd probably bias us
in favor of continuing to report the same value from run to run,
which might have some benefit. The thing is that if no value in
the table is especially common, the existing algorithm will give a
nearly random selection that might not be the most common value or
even very close to it.
I'm not really excited about spending effort on marginal tweaks
of the current method, however. What we really need is more stats
(and more reliable stats) than we have, and that's going to require
some work.
regards, tom lane