Lecture spine

This is a topic roadmap for the term. The assignment schedule and milestones are stable; the exact ordering of topics can shift (e.g., for a guest lecture) without changing what is assessed.

Readings: see Readings for the week-by-week mapping.

Modeling: see TLA+ / PlusCal track (lightweight) for how formal specs fit into the course.

Three arcs

1. Universal agreement: transactions and atomic commit (2PC/3PC) — what breaks under failure, and what assumptions are needed to recover.
2. Partial agreement: quorums (set quorum first), view changes, and crash-fault consensus (Viewstamped Replication / Paxos).
3. Adversarial agreement: Byzantine consensus (pBFT) and incentive-driven ledger systems.

Two recurring lenses

• Failure-first: start from the failure, then derive the protocol mechanism that prevents or recovers from it.
• Optimization & Goodhart: metrics are projections; optimizing one property can erase other properties. We use this lens both for distributed systems (e.g., incentives, availability vs correctness) and for GenAI-assisted work (polish vs evidence).

Week-by-week sketch (W01–W12)

Reading week sits between W06 and W07. The final week of class meetings (W13) is reserved for low-risk activities (reviews/demos).

• W01: Course framing (failure-first, evidence chain) + GenAI/Goodhart + DS Labs scaffold
• W02: Communication semantics (timeouts/retries, at-most-once) and why “exactly-once” is hard
• W03: Transactions under failure: 2PC (blocking) + invariants; introduce TLA+/PlusCal as a tool
• W03 (slot): Guest lecture on modeling and formal verification (TLA+/PlusCal)
• W04: 3PC and synchrony assumptions; failure detectors; DP1 design due by start of week
• W05: Replication: primary/backup + view change; set quorum intuition; DP1 peer review
• W06: Leader election & “ballots before Paxos”: Lamport’s bakery algorithm as a warm-up; then partial consensus (viewstamped → Paxos safety); DP1 implementation + reflection setup
• W07: Paxos mechanics and Multi-Paxos; DP1 reflection; DP2 design (leveraging DP1 lessons)
• W08: Paxos in systems: reconfiguration, performance, observability; optional follow-up on TLA+ modeling
• W09: Mid-course Goodhart revisit: measurement, incentives, fault injection, and “evidence vs polish”
• W10: Byzantine failures: pBFT and quorum intersection under adversarial behavior
• W11: Ledgers: Nakamoto-style consensus and incentive design; optimization tradeoffs (sourdough analogy)
• W12: Avoiding consensus: CRDTs/eventual consistency; capstone evaluation clinic (failure plans)