The Store

Model Answer

Use a Raft-replicated state machine: all writes go through the leader, which replicates to a majority before acknowledging. Reads from the leader provide linearizability (or use ReadIndex protocol to avoid log round-trips). Partition tolerance: the majority partition continues serving both reads and writes; the minority partition becomes unavailable for writes (choosing CP in CAP terms). This is the etcd / CockroachDB model. For higher availability at the cost of consistency, use quorum reads/writes (DynamoDB style) where W + R > N.

Model Answer

With N replicas, requiring W write acknowledgments and R read responses such that W + R > N ensures every read quorum intersects every write quorum by at least one node. Example: N=3, W=2, R=2 (2+2>3). A write must reach 2 of 3 nodes. A read queries 2 of 3 nodes. At least 1 node is in both sets — that node has the latest write, and a read that returns the highest version across R responses will return it. This does not guarantee linearizability (concurrent writes can still cause conflicts), but prevents reading a value older than the last quorum write.

Model Answer

DynamoDB uses conditional writes and optimistic locking via version numbers. Without a condition, the last writer wins (based on wall clock, which is susceptible to clock skew). With conditional expressions (e.g., version = expected_version), write is rejected if the version does not match, forcing the client to re-read and retry. DynamoDB's original design used vector clocks (described in the 2007 Dynamo paper), but the production service moved to simpler conflict resolution because clients found vector clock reconciliation complex to implement correctly.

Model Answer

A key-value store persists data durably and provides strong consistency guarantees (e.g., etcd, DynamoDB). A distributed cache stores data in memory for fast access, typically with eviction and optional durability. Redis can function as both: with AOF (append-only file) and/or RDB snapshots enabled, Redis provides durability, but its consistency model (single-threaded, eventually consistent replication) differs from systems like etcd. For critical data, use a durable store with replication; use Redis as a cache in front of it.

Model Answer

Read repair happens when a quorum read returns different values from different replicas. The coordinator detects the inconsistency and sends a write to the stale replicas to bring them up to date. This is background repair triggered by read traffic. It solves replica divergence caused by temporary node failures: if a node was down during a write, it missed the update; the next quorum read that includes this node detects the divergence and repairs it. Cassandra also runs anti-entropy repair (Merkle tree-based) as a scheduled background process to catch divergence that is not detected by reads.