This is what I also understood after reading this thread. So, my question is why not just review v3 and commit something on those lines even though it would take a bit more time. It is possible that if we decide to go with v5, we can make it happen earlier, but later when we try to get v3, the code committed as part of v5 might not be of any use or if it is useful, then in which cases?
Yes, approach #2 (v3) would be generally better than approach #1 (v5) in
performance. I started with approach #1 because it is much easier.
If we decide to go with approach #2, I think we can now concentrate on v3 patch.
For v3 patch, we have some other idea, which is to perform a normal grouping sets aggregation in the final phase, with 'GroupingSetId' included in the group keys (as described in the previous email). With this idea, we can avoid a lot of hacky codes in current v3 patch.
