Consistency & Reconciliation

1 / 8

Core challenge

Replication improves latency/availability, but replicas diverge under delay + failure.

Need both:

  • correctness model
  • repair strategy
2 / 8

Model hierarchy

Strongest to weakest:

  1. Linearizable
  2. Causal
  3. Eventual

A read can be causal but not linearizable.

3 / 8

Bounded staleness

  • SLO-style guarantee: lag <= window
  • choose lowest latency among eligible replicas
  • if none eligible: escalate/fail by policy
4 / 8

Reconciliation mechanics

For each key:

  • VC relation: A/B/equal/conflict
  • deterministic resolution policy (often LWW fallback)

Conflict detection and conflict resolution are separate concerns.

5 / 8

Multi-version read repair

  • keep non-dominated versions
  • repair dominated replicas
  • collapse exact duplicates deterministically

Do not treat concurrent versions as stale.

6 / 8

Session guarantees

Per-client guarantees:

  • RYW
  • MR
  • MW
  • WFR

Client-visible correctness can be strong even when global model is weaker.

7 / 8

Exercise progression

  1. Model classifier
  2. Bounded staleness planner
  3. Anti-entropy reconciler
  4. Multi-version repair planner
  5. Session guarantees analyzer capstone
8 / 8
Use arrow keys or click edges to navigate. Press H to toggle help, F for fullscreen.