Quiz

C — Nothing certain; retries must be idempotent.

A lost request, a crashed server, a lost reply and a very slow server all look the same to the caller: silence. The timeout only says "I stopped waiting". If the reservation was actually applied and the reply was lost, a naive retry reserves the unit twice. Making the operation idempotent (for example, with a unique request id that the server deduplicates) makes the retry safe.

D — 8.

On receipt, the rule is max(local, received) + 1 = max(3, 7) + 1 = 8. This guarantees that the receive event gets a larger timestamp than the send event, so "happened-before" is always reflected in the timestamps. Note the converse does not hold: a smaller Lamport timestamp does not prove causality — vector clocks are needed for that.

C — Partition Tolerance.

Network partitions — where messages between nodes are lost or delayed — happen in any real distributed system due to hardware failures, congestion, or physical distance. Because partitions cannot be prevented, the practical design choice is between C (refuse requests during partition) and A (continue serving, possibly stale data).

B — CP.

By refusing reads during a partition, the system prioritises consistency (no stale data) over availability. Once the partition heals, service resumes. ZooKeeper and etcd are classic CP examples.

C — Apache Cassandra and DNS.

Both continue serving requests during partitions, accepting that responses may be temporarily stale (eventual consistency). ZooKeeper and etcd are CP; PostgreSQL with synchronous replication and MongoDB with its default majority writes behave as CP. A single-node Redis or SQLite is not replicated at all, so CAP does not really apply to it — it sits in the "CA" corner only because there is no second node to be partitioned from.

B — Even without a partition: Latency vs Consistency trade-off.

CAP only describes behaviour during a partition. PACELC adds that even in normal operation, systems must trade off between lower latency (serve quickly, risk slightly stale reads) and stronger consistency (wait for replication agreement before responding). Cassandra is PA/EL; MongoDB with majority writes is PC/EC.

C — CP store.

For financial transactions, correctness is non-negotiable. A CP store refuses to process the payment during the partition (the customer gets an error, not a double charge). An AP store might process the payment on two nodes simultaneously before the partition heals, creating a duplicate transaction that cannot be safely undone. Last-write-wins (D) would silently discard one of the conflicting writes, and depends on clocks that drift.

C — Single-machine or tightly controlled LAN.

CA only works when partitions genuinely cannot occur — which is only realistic for a single server or a local area network under strict control. In any internet-connected or geographically distributed deployment, P is effectively mandatory, leaving the real choice between C and A.

D — PA/EL.

PA = during a Partition, favours Availability. EL = Else (normal operation), favours Latency. Apache Cassandra is the canonical PA/EL system — it always responds quickly and accepts eventual consistency in both scenarios.

B — No guaranteed termination in an async system with ≥1 faulty node.

FLP (Fischer, Lynch, Paterson, 1985) proves this fundamental limit: some unlucky ordering of delays can keep any deterministic algorithm undecided forever. Real algorithms respond by assuming partial synchrony (delays are eventually bounded) or by using randomisation. Paxos and Raft never violate safety, and use timeouts as a proxy for failure detection to make progress.

C — Follower election timeout expires.

Each follower maintains a randomised election timeout. If no heartbeat arrives before it expires, the follower assumes the leader has crashed and starts a new election by incrementing the term and requesting votes. Randomised timeouts prevent multiple followers from starting elections simultaneously.

A — 1.

A write needs a majority: ⌊4/2⌋ + 1 = 3 nodes. With 2 nodes down only 2 remain, which is not a majority. So 4 nodes tolerate ⌊(4-1)/2⌋ = 1 failure — the same as 3 nodes, with one more machine that can fail. That is why consensus clusters use odd sizes (3 or 5). An even cluster can still elect a leader (D is false); it just gains nothing.

D — W + R > N.

If write quorum W plus read quorum R exceeds total replicas N, at least one node that acknowledged the write must also participate in any read quorum. This overlap ensures the latest write is always visible in a quorum read, forming the basis of systems like Dynamo and Cassandra's tunable consistency (for example, QUORUM reads and writes with N = 3: 2 + 2 > 3).

B — 7.

The G-Counter merge takes the element-wise maximum of the slots: [max(4,2), max(1,3), max(0,0)] = [4, 3, 0], whose sum is 7. Because each node only increments its own slot, the maximum of a slot is its latest value; summing the two totals (C) would double-count, and the merge is commutative and idempotent, so the order of merges does not matter (D is false).