A couple of months ago, I was working on something that exactly keeps track of things like market data, where you have a sliding window of datapoints that you want to aggregate in possibly multiple ways.

The code was uploaded to Github; it’s not yet on Hexpm because it is still a little bare-bones (although there are tests!).
Actually, the tests might be the best explanation as to how it works .

When I researching into this topic a while back I thought to use ETS ordered sets. Basically use the ordered set with a counter like {n, payload} and monotonically increase n. Then you can use :ets.next/1 for access. But of course then you need a sweeper to delete the head of the set on some kind of regular basis. Or manage n predefined slots in an ordered set and manage the read and write “pointers” manually.

I also tried the list-based approach you describe above but it does involve a lot of list copying on pretty much each update. I never finished up my experiments so I have no useful data to offer.

Can you amortize the cost of popping the old items from the list by allowing the list to grow up to 2x the desired size, then calling Enum.take to keep the required prefix?

I would test with https://hex.pm/packages/epocxy which already has a ring buffer implemented in ETS.

You may be interested in riak-ts, an open source NoSQL database optimized for time-series and written in Erlang.

http://docs.basho.com/riak/ts/1.5.2/

I’ve not used it myself yet but it looks interesting for logging time series data.