Apache Cassandra is the archetype of high-throughput, horizontally scalable, masterless distributed databases. Inspired by Amazon’s Dynamo paper and Google’s Bigtable storage model, Cassandra dispenses with leader-follower architectures in favor of a completely decentralized peer-to-peer token ring. Every node is architecturally identical; any node can coordinate any read or write, guaranteeing zero single points of failure.
This architectural study deconstructs Cassandra’s internal systems:
- Consistent Hashing Token Ring: Murmur3 64-bit token distribution, virtual nodes (vnodes), and token-range replica placement across datacenter racks.
- Gossip & The Accrual Failure Detector: Asynchronous peer state convergence, generation heartbeats, and probabilistic network health modeling.
- Storage Engine Internals: Append-only CommitLogs for durable writes, concurrent in-memory MemTables, immutable on-disk SSTables, row-level Bloom filters, partition indexes, and Tombstone garbage collection.
- Tunable Consistency & Repair Mechanics: Quorum intersection proofs (), Hinted Handoff buffers, asynchronous Read Repair, and anti-entropy Merkle tree background repairs.
This module provides the blueprint for building and operating petabyte-scale masterless distributed storage.