Pick 2 of 3: Consistency, Availability, Partition Tolerance
Since P is inevitable → Choose CP or AP
CP: ZooKeeper, HBase, MongoDB (strong consistency, may reject requests)
AP: Cassandra, DynamoDB, CouchDB (always available, eventual consistency)
Model Guarantee Example
Linearizable Real-time ordering ZooKeeper Sequential Total order, no real-time Raft Causal Cause-effect preserved MongoDB Read-your-writes See own writes User profiles Eventual Converges eventually Cassandra
Algorithm Fault Tolerance Complexity Use
Paxos f < n/2 Hard to understand Theory Raft f < n/2 Understandable etcd, CockroachDB ZAB f < n/2 Similar to Raft ZooKeeper PBFT f < n/3 (Byzantine) O(n²) Blockchain
Follower → (timeout) → Candidate → (majority votes) → Leader
Leader → (term higher) → Follower
Candidate → (higher term seen) → Follower
Log Replication: Leader receives → appends → replicates → commits (majority ack)
Strategy Latency Consistency Data Loss Risk
Synchronous High Strong None Asynchronous Low Eventual Possible Semi-sync Medium Hybrid Minimal Chain Variable Strong Depends on position Quorum (NRW) Configurable Tunable N-W+R > N ensures consistency
N = total replicas
W = write quorum
R = read quorum
Strong consistency if: W + R > N
Example: N=3, W=2, R=2 → guaranteed consistency
Strategy Pros Cons
Hash Even distribution Range queries hard Range Range queries efficient Hotspots possible Consistent Hashing Minimal redistribution Complexity
Ring of 0 to 2^32 - 1
Server → hash → position on ring
Key → hash → walk clockwise → nearest server
Virtual nodes: multiple positions per server for balance
Heartbeat: Periodic "I'm alive" messages
Timeout: No heartbeat → suspect failure
Gossip: Peer-to-peer state propagation
Phi Accrual: Adaptive failure detector (Cassandra)
Protocol Blocking Rounds Use
2PC Yes (coordinator failure) 2 Traditional DB 3PC No (theoretically) 3 Rarely used Saga No N (compensating) Microservices TCC No 3 Business transactions
Guarantee Meaning Implementation
At-most-once May lose messages Fire and forget At-least-once May duplicate Retry + ack Exactly-once No loss, no dup Idempotent + dedup
System Type Consensus CAP
Kafka Log/messaging ISR (Raft-like) AP→CP Cassandra Wide-column Gossip AP ZooKeeper Coordination ZAB CP etcd KV store Raft CP CockroachDB SQL Raft CP DynamoDB KV/document Vector clocks AP
Always discuss failure modes (network partition, node crash, split brain)
Explain consistency trade-offs clearly
Know when to use CP vs AP
Draw the architecture with data flow arrows
Mention monitoring, alerting, and recovery