Chapter 29: Consistency and Consensus (Raft, Paxos)
In distributed systems, how do multiple nodes agree on the state of data? This is the consensus problem.
Consistency Models
| Model | Guarantee | Example |
|---|---|---|
| Strong (Linearizable) | All reads see the latest write | Single-node PostgreSQL |
| Sequential | All nodes see operations in same order | ZooKeeper |
| Causal | Causally related operations are ordered | Some distributed DBs |
| Eventual | All nodes converge eventually | DNS, Cassandra (default) |
Raft Consensus Algorithm
Used by etcd, CockroachDB, TiDB. Elects a leader; leader replicates log entries to followers.
Raft cluster (3 nodes):
┌────────┐ ┌──────────┐ ┌──────────┐
│ Leader │────►│ Follower │ │ Follower │
│ │────►│ │ │ │
└────────┘ └──────────┘ └──────────┘
1. Client sends write to Leader
2. Leader appends to its log
3. Leader replicates to followers
4. Once majority (2/3) acknowledge → committed
5. Leader responds to client
If leader dies → followers elect new leader (election timeout)
Two-Phase Commit (2PC)
For transactions spanning multiple nodes:
// Phase 1: PREPARE
Coordinator → all participants: "Can you commit?"
All participants: lock resources, write to WAL, respond YES/NO
// Phase 2: COMMIT (if all said YES)
Coordinator → all participants: "COMMIT"
All participants: make changes permanent, release locks
// Problem: if coordinator crashes between phases, participants are stuck💡 Practical Impact
You rarely implement consensus yourself. You use databases that implement it internally (CockroachDB, YugabyteDB, TiDB) or coordination services (etcd, ZooKeeper). Understanding the concepts helps you reason about consistency guarantees and failure modes.
Key Takeaways
- Consensus = getting distributed nodes to agree on state
- Raft: leader-based, majority quorum, used in modern distributed DBs
- Strong consistency is expensive (latency of slowest node in quorum)
- Eventual consistency is fast but requires application-level conflict handling
- 2PC enables distributed transactions but has blocking failure modes