Membership & Failure Detection

1 / 6

Core challenge

Liveness is inferred, not observed. Network noise makes failure detection probabilistic.

Goal: converge fast without excessive false positives.

2 / 6

Gossip digest sync

Digest entry: (member, incarnation, heartbeat)

For each member:

  • local newer -> send
  • remote newer -> request
  • equal -> no action

Incarnation must dominate heartbeat.

3 / 6

Phi accrual

phi = (elapsed / mean_interval) / ln(10)

  • phi increases as heartbeat silence grows
  • suspect when phi >= threshold
  • sample quality and threshold selection drive behavior
4 / 6

SWIM lifecycle

alive -> suspect -> dead

Need:

  • unique reporter counting
  • threshold/quorum
  • alive reset with newer incarnation
  • deterministic ordered transitions
5 / 6

Exercise progression

  1. Versioned digest sync planner
  2. Phi accrual evaluator
  3. SWIM transition capstone

Focus: robust state transitions under delay, duplication, and restarts.

6 / 6
Use arrow keys or click edges to navigate. Press H to toggle help, F for fullscreen.