distributed-systems

CAP theorem states you can only guarantee two of consistency, availability, and partition tolerance. Since network partitions are unavoidable, choose CP (like HBase or MongoDB with majority writes) for strong consistency, or AP (like Cassandra or DynamoDB) for availability with eventual consistency.

Raft is more practical and easier to implement, using leader election and log replication with clear state transitions between follower, candidate, and leader. Paxos is more general but notoriously difficult to implement correctly, so Raft is preferred for most systems like etcd.

Consistent hashing maps nodes and keys onto a hash ring, assigning each key to the closest node clockwise. Adding a node only redistributes keys in its immediate range rather than reshuffling everything, moving roughly 1/N of the data.

Use quorum-based decisions requiring a majority of nodes, such as 3 out of 5, to elect a leader or commit writes. Only the partition side holding a majority can make progress, preventing two leaders from operating simultaneously.

Use sagas when services must remain available during failures, since two-phase commit blocks if the coordinator crashes. Sagas execute local transactions with compensating actions, such as refunding a payment when inventory reservation fails.

Leaderless replication requires quorum tuning (W + R > N) to guarantee read-your-writes consistency, and concurrent writes need conflict resolution via vector clocks, last-write-wins, or CRDTs. It trades simplicity of failover for complexity in conflict handling.