On Sat, 16 Sept 2023 at 01:00, jian he <jian.universal...@gmail.com> wrote: > > I refactor the avg(interval), sum(interval), so moving aggregate, > plain aggregate both work with +inf/-inf. > no performance degradation, in fact, some performance gains. >
I haven't reviewed this part in any detail yet, but I can confirm that there are some impressive performance improvements for avg(). However, for me, sum() seems to be consistently a few percent slower with this patch. The introduction of an internal transition state struct seems like a promising approach, but I think there is more to be gained by eliminating per-row pallocs, and IntervalAggState's MemoryContext (interval addition, unlike numeric addition, doesn't require memory allocation, right?). Also, this needs to include serialization and deserialization functions, otherwise these aggregates will no longer be able to use parallel workers. That makes a big difference to queryE, if the size of the test data is scaled up. This comment: + int64 N; /* count of processed numbers */ should be "count of processed intervals". Regards, Dean
