Hi
On Fri, 2019-05-03 at 15:56 -0700, Paul Jungwirth wrote:
> Hello,
>
> I wrote an extension to add a range_agg function with similar
> behavior
> to existing *_agg functions, and I'm wondering if folks would like
> to
> have it in core? Here is the repo:
> https://github.com/pjungwir/range_agg
This seems like a very useful extension, thank you.
For getting into core though, it should be a more complete set of
related operations. The patch is implicitly introducing the concept of
a "multirange" (in this case, an array of ranges), but it's not making
the concept whole.
What else should return a multirange? This patch implements the union-
agg of ranges, but we also might want an intersection-agg of ranges
(that is, the set of ranges that are subranges of every input). Given
that there are other options here, the name "range_agg" is too generic
to mean union-agg specifically.
What can we do with a multirange? A lot of range operators still make
sense, like "contains" or "overlaps"; but "adjacent" doesn't quite
work. What about new operations like inverting a multirange to get the
gaps?
Do we want to continue with the array-of-ranges implementation of a
multirange, or do we want a first-class multirange concept that might
eliminate the boilerplate around defining all of these operations?
If we have a more complete set of operators here, the flags for
handling overlapping ranges and gaps will be unnecessary.
I think so scope of this patch is correct. Introduction of set of ranges data type based on a array or not, should be different topic.
The question is naming - should be this agg function named "range_agg", and multi range agg "multirange_agg"? Personally, I have not a problem with range_agg, and I have not a problem so it is based on union operation. It is true so only result of union can be implemented as range simply. When I though about multi range result, then there are really large set of possible algorithms how to do some operations over two, three values. So personally, I am satisfied with scope of simple range_agg functions, because I see a benefits, and I don't think so this implementation block any more complex designs in the future. There is really big questions how to implement multi range, and now I think so special data type will be better than possible unordered arrays.
Regards
Pavel