Distributed KV Store for High-Throughput Workloads

I would design a partitioned, replicated in-memory KV store. Data is split across 4,096 partitions using consistent hashing with virtual nodes to ensure even distribution. Each partition has a leader and two followers using Raft consensus for strong consistency. For the read path at 10M RPS, reads are served from any replica's in-memory hash table — since session data tolerates slightly stale reads, this distributes load across all replicas. For strong-consistency reads, clients route to the leader and verify the read index against the committed Raft log. For the write path at 1M WPS, the client hashes the key to find the partition leader, which appends the write to a WAL, batches operations every 1ms, and replicates to followers. After majority acknowledgment, the leader responds and applies the write to its in-memory store. For durability, the WAL is synced to SSD every 100ms. To handle hot keys (e.g., viral user sessions), I would implement key-level caching in a separate hot-key layer that absorbs bursts. Cluster management uses an etcd-backed controller that detects node failures via heartbeats and triggers leader re-election within 2 seconds.