Re: Statistics and selectivity estimation for ranges

On Mon, Aug 27, 2012 at 5:00 PM, Heikki Linnakangas <
heikki(dot)linnakangas(at)enterprisedb(dot)com> wrote:

> On 24.08.2012 18:51, Heikki Linnakangas wrote:
>
>> On 20.08.2012 00:31, Alexander Korotkov wrote:
>>
>>> New version of patch.
>>> * Collect new stakind STATISTIC_KIND_BOUNDS_**HISTOGRAM, which is lower
>>> and
>>> upper bounds histograms combined into single ranges array, instead
>>> of STATISTIC_KIND_HISTOGRAM.
>>>
>>
>> One worry I have about that format for the histogram is that you
>> deserialize all the values in the histogram, before you do the binary
>> searches. That seems expensive if stats target is very high. I guess you
>> could deserialize them lazily to alleviate that, though.
>>
>> * Selectivity estimations for>,>=,<,<=,<<,>>,&<,&> using this
>>> histogram.
>>>
>>
>> Thanks!
>>
>> I'm going to do the same for this that I did for the sp-gist patch, and
>> punt on the more complicated parts for now, and review them separately.
>> Attached is a heavily edited version that doesn't include the length
>> histogram, and consequently doesn't do anything smart for the &< and &>
>> operators. && is estimated using the bounds histograms. There's now a
>> separate stakind for the empty range fraction, since it's not included
>> in the length-histogram.
>>
>> I tested this on a dataset containing birth and death dates of persons
>> that have a wikipedia page, obtained from the dbpedia.org project. I can
>> send a copy if someone wants it. The estimates seem pretty accurate.
>>
>> Please take a look, to see if I messed up something.
>>
>
> Committed this with some further changes.

Thanks! Sorry for I didn't provide a feedback for previous message.
Commited patch looks nice for me. I'm going to provide additional patch
with length-histogram and more selectivity estimates.

------
With best regards,
Alexander Korotkov.