Re: ANALYZE sampling is too good

On Mon, Dec 9, 2013 at 6:03 PM, Josh Berkus <josh@agliodbs.com> wrote:
>
> It's also applicable for the other stats; histogram buckets constructed
> from a 5% sample are more likely to be accurate than those constructed
> from a 0.1% sample.   Same with nullfrac.  The degree of improved
> accuracy, would, of course, require some math to determine.

This "some math" is straightforward basic statistics.  The 95th
percentile confidence interval for a sample consisting of 300 samples
from a population of a 1 million would be 5.66%. A sample consisting
of 1000 samples would have a 95th percentile confidence interval of
+/- 3.1%.

The histogram and nullfact answers the same kind of question as a
political poll, "what fraction of the population falls within this
subset". This is why pollsters don't need to sample 15 million
Americans to have a decent poll result. That's just not how the math
works for these kinds of questions.

n_distinct is an entirely different kettle of fish. It's a different
kind of problem and the error rate there *is* going to be dependent on
the percentage of the total population that you sampled. Moreover from
the papers I read I'm convinced any sample less than 50-80% is nearly
useless so I'm convinced you can't get good results without reading
the whole table.

-- 
greg