Re: pg_stats and range statistics

Hi Tomas,

On 12.07.2021 16:04, Tomas Vondra wrote:
> On 7/12/21 1:10 PM, Egor Rogov wrote:
>> Hi,
>>
>> thanks for the review and corrections.
>>
>> On 11.07.2021 21:54, Soumyadeep Chakraborty wrote:
>>> Hello,
>>>
>>> This should have been added with [1].
>>>
>>> Excerpt from the documentation:
>>> "pg_stats is also designed to present the information in a more readable
>>> format than the underlying catalog — at the cost that its schema must
>>> be extended whenever new slot types are defined for pg_statistic." [2]
>>>
>>> So, I added a reminder in pg_statistic.h.
>> Good point.
>>
>>
>>> Attached is v2 of this patch with some cosmetic changes.
>> I wonder why "TODO: catalog version bump"? This patch doesn't change
>> catalog structure, or I miss something?
>>
> It changes system_views.sql, which is catalog change, as it redefines
> the pg_stats system view (it adds 3 more columns). So it changes what
> you get after initdb, hence catversion has to be bumped.
>
>>> Renamed the columns a
>>> bit and updated the docs to be a bit more descriptive.
>>> (range_length_empty_frac -> empty_range_frac, range_bounds_histogram ->
>>> range_bounds_histograms)
>> I intended to make the same prefix ("range_") for all columns concerned
>> with range types, although I'm fine with the proposed naming.
>>
> Yeah, I'd vote to change empty_range_frac -> range_empty_frac.
>
>>> One question:
>>>
>>> We do have the option of representing the histogram of lower bounds
>>> separately
>>> from the histogram of upper bounds, as two separate view columns.
>>> Don't know if
>>> there is much utility though and there is a fair bit of added
>>> complexity: see
>>> below. Thoughts?
>> I thought about it too, and decided not to transform the underlying data
>> structure. As far as I can see, pg_stats never employed such
>> transformations. For example, STATISTIC_KIND_DECHIST is an array
>> containing the histogram followed by the average in its last element. It
>> is shown in pg_stats.elem_count_histogram as is, although it arguably
>> may be splitted into two fields. All in all, I believe pg_stats's job is
>> to "unpack" stavalues and stanumbers into meaningful fields, and not to
>> try to go deeper than that.
>>
> Not firm opinion, but the pg_stats is meant to be easier to
> read/understand for humans. So far the transformation were simple
> because all the data was fairly simple, but the range stuff may need
> more complex transformation.
>
> For example we do quite a bit more in pg_stats_ext views, because it
> deals with multi-column stats.

In pg_stats_ext, yes, but not in pg_stats (at least until now).

Since no one has expressed a strong desire for a more complex
transformation, should we proceed with the proposed approach (with
further renaming empty_range_frac -> range_empty_frac as you suggested)?
Or should we wait more for someone to weigh in?

>
>
> regards
>