Reader algorithm suboptimal (quadratic)

[olorin]

We're putting fixing this on hold due to other priorities, but it should be documented.

Currently, a batch read will enumerate the buckets it hits O(n) times in the number of points read/read each point O(n^2) times, due to naive filtering out of points which aren't the one being currently read. This can be fixed by only reading each bucket once per batch (either memoize it, or rethink the intermediate datastructure used to accumulate points in a request).