Design a Key-Value Store

Hard

System design · Storage

Design a durable distributed key-value database in the style of Dynamo or Cassandra (and contrast it with Raft-based stores such as etcd). Clients put, get and delete values by key. Scale: 100TB of data, 500K reads/s and 100K writes/s, replication factor 3, p99 get under 10ms, node failures weekly. State a clear consistency contract and cover partitioning, replication, quorums, conflict resolution, hinted handoff and repair, the storage engine, and tombstones.

Requirements

Partition data with consistent hashing and virtual nodes, replicating across failure domains
Define the consistency contract: quorum settings, what a read can return, and what the store does on conflict
Explain write durability: commit log, replication and what an acknowledgement means
Keep replicas in sync with hinted handoff, read repair and anti-entropy, and handle deletes safely

Workspace