There is a concurrent API in prometheus that we could use -- it's more stateful but it simplifies the internals of ekg-prometheus-adapter and I think it also improves memory efficiency as we don't have to keep around all those lazy tuples.
We should test out if the switch is viable. A have a local branch which I have ported to the new ekg-prometheus-adapter and things work, but it would interesting to test this out on a big project or to measure memory residency.