As ever more big data computations start to be in-memory, I/O throughput dominates the running times of many workloads. For distributed storage, the read throughput can be improved using caching, however, the write throughput is limited by both disk and network bandwidth due to data replication for fault-tolerance. This paper proposes a new file system architecture to enable frameworks to both read and write reliably at memory speed, by avoiding synchronous data replication on writes.