Distributed Systems: Consistency Models¶

The CAP Theorem¶

In distributed systems, you can only have 2 of 3: - Consistency: All nodes see the same data - Availability: Every request gets a response - Partition Tolerance: System works despite network partitions

Consistency Models (Strongest to Weakest)¶

1. Linearizability (Strongest)¶

Operations appear to execute atomically
Real-time ordering respected
Example: Single-server systems

2. Sequential Consistency¶

Operations appear in some total order
Each process's operations appear in program order
Example: Most traditional databases

3. Causal Consistency¶

Causally related operations ordered
Concurrent operations may be seen in different orders
Example: Version vectors, vector clocks

4. Eventual Consistency (Weakest)¶

All replicas eventually converge
No ordering guarantees
Example: DNS, CDNs

Practical Considerations¶

Model	Latency	Availability	Use Case
Linearizable	High	Low	Financial transactions
Sequential	Medium	Medium	User sessions
Causal	Low	High	Social feeds
Eventual	Lowest	Highest	Caching, CDN

Implementation Patterns¶

Quorum Systems¶

N = Total replicas
W = Write quorum
R = Read quorum

Strong consistency: W + R > N
Example: N=3, W=2, R=2

Vector Clocks¶

Track causality across nodes:

Node A: {A:1, B:0}
Node B: {A:0, B:1}
After sync: {A:1, B:1}

Real-World Examples¶

Spanner: Linearizable (TrueTime)
DynamoDB: Eventually consistent (default)
CockroachDB: Serializable
Cassandra: Tunable consistency

目录