[calmcoconut]

Summary

This PR addresses a quality audit of all 18 multiple-choice questions in Chapter 10 (Consistency & Consensus). The audit found that 14 of 18 questions had at least one distractor that was trivially eliminable — either nonsensical, factually absurd, or directly contradicting common knowledge — allowing students to guess the correct answer through elimination rather than genuine understanding.

Problem

Weak distractors undermine the quiz's pedagogical purpose. When a student can eliminate 2 of 4 options immediately (e.g., "ML algorithm that predicts latency" or "once per calendar day"), the question becomes a 50/50 guess rather than a test of knowledge.

Changes

All replaced distractors use real-but-wrong alternatives drawn from concepts in the chapter or adjacent systems:

Question	Issue	Fix
Q1 — Linearizability definition	D was vague nonsense ("transaction leases")	Replaced with a monotonic-reads description that mimics the correct answer
Q2 — Stale replica cause	B blamed client device clock skew	Replaced with server-side NTP skew, a real plausible cause
Q5 — Linearizability + Serializability	A was "weak eventual consistency" (opposite); D was unrelated 2PC	Replaced with causal consistency and snapshot isolation (real models)
Q8 — NOT linearizable method	D was "single-node database" (trivially correct)	Replaced with leaderless quorum w+r>n — forces precise understanding of why multi-leader fails categorically
Q10 — ID ordering without coordinator	Correct answer was "None of the above" (anti-pattern)	Reframed question; Snowflake-style IDs are now the correct answer with nuanced explanation
Q11 — What logical clocks count	B (CPU cycles) and D (active nodes) were absurd	Replaced with "messages sent/received" and "version of most recent write"
Q13 — Lamport timestamps & uniqueness	A falsely claimed Lamport timestamps have no node IDs	Replaced with plausible network reordering argument
Q14 — HLC advantage	A was UUID (completely unrelated)	Replaced with GPS/atomic clock option that conflates HLC with Spanner TrueTime
Q16 — Timestamp oracle definition	A was "ML algorithm for latency"	Replaced with stateless UUID generator service
Q17 — Spanner's approach	C was "single master in NA"; D was "abandons linearizability"	Replaced with Paxos-based timestamp service and MVCC-only options
Q19 — FLP impossibility	C was "eventual consistency is impossible" (obviously false)	Replaced with a near-miss restatement about arbitrary delays
Q21 — State machine replication	C (heterogeneous OS kernels) and D (random routing) were absurd	Replaced with snapshotting and unordered fan-out — real replication patterns
Q23 — Consensus split-brain prevention	C (shared disk) and D (manual operators) were implausible	Replaced with lease-based and pre-vote protocol options
Q24 — Unclean leader election	D described serializability/disk space (unrelated)	Replaced with description of clean leader election — the exact opposite, a perfect trap
Q26 — ZooKeeper/etcd use case	A (petabytes of clickstream) and D (image storage) were absurd	Replaced with distributed cache and cross-shard transaction coordination
Q28 — etcd linearizable reads	D was "throttled to once per calendar day" (absurd)	Replaced with TiKV-style read-index protocol — a real alternative implementation

Not Changed

• Q7 (CAP Consistency definition) — Already excellent. All four options (ACID consistency, eventual consistency, linearizability, monotonic reads) are real, credible models a student could plausibly confuse.
• Q4 (Linearizability vs. Serializability) — Option A is the mirror-flip of the correct answer, already a strong trap.
• All write (open-ended) questions, flashcards, and misconception explanations are unchanged.

Testing

No logic changes — all correct index values are preserved where the correct answer position was unchanged, and updated where the correct answer moved (Q10, Q28). Please verify answer indices match the new option ordering.