Ok, I've submitted https://github.com/OpenNMS/newts/pull/35 to implement this.

The lists are already sorted by value? I'm currently calculating P95/P99 by discarding NaNs, sorting the lists by value, and picking the element at the 95th or 99th percentile in the list. This seemed like a reasonable approach. But maybe I'll just submit code; that will make it easier to work out details.

The collection of samples passed to the aggregation function should already be sorted, in most cases it shouldn't matter though.

Why would the samples need to be sorted for the P95 and P99 though? Shouldn't the percentiles of two lists with the same values be identical?

Great, thanks. I've coded up a P95 and P99 and those seem to be working well in my own fork. Should I keep discussion regarding implementation details on the pull request, or here?

For example, calculating both p95 and p99 requires the incoming sample collection to be sorted. Right now I'm just copying it and sorting it separately in each aggregate. It'd be more effective to sort it outside and offer a guarantee to the aggregate functions that the incoming collection is already sorted, but that means a useless sort if the only aggregates you're using are MIN, MAX, and AVERAGE. I suspect that in most situations that won't matter, but if you have a lot of samples to roll up over, it could get expensive. I guess the simplest approach is to re-sort in each aggregate and not worry about it unless people report issues.

Sounds reasonable. You're welcome to submit a patch