@marcbrooker Interesting blog post! This reminds me of a problem I came up with about a year ago (but never got around to seriously thinking about).

Namely, in a geodistributed system with geodistributed clients, can erasure coding yield any benefits for the latency experienced by those clients in accessing some piece of data? Specifically, here, I'm neglecting the message handling cost entirely and only considering the network latency.

Here's the problem statement I came up with:

Given a graph G = (V, E), latencies for each edge l: E -> R+, an assignment of readers to nodes r: V -> N, d \in N data chunks, and a threshold t <= d, we want to find the optimal placement of these data chunks. That is, we want a function p: V -> N such that \sum_{v \in V} p(v) = d. The latency experienced by a reader at node v is the (maximum) latency between v and a set of nodes whose total weight under p is at least t (minimized across all such sets, of course). Optimal here could mean average (weighted by r) or worst-case etc.

The non-coded case is when t=1.

The question is, keeping G, l, r, and d/t constant (i.e., storing the same amount of data), can increasing t (and changing p accordingly) yield latency benefits under any of our definitions of optimality?