🏴‍☠️ Failure Detection

Possible through 2 subprotocols

• Failure detection → “I think server A failed”
• Dissemination → “I need to tell everyone else that server A failed”

Desired properties

• Completeness → every failure is detected → aim for guarantee
• Accuracy → no mistaken detections → partial or probabilistic
• Speed
• Scale

Impossible to have both completeness and accuracy in any lossy network, never know if it was a server crash vs. slow network vs. etc… [Chandra and Toueg]

MTTF (mean time to failure) is proportional to scaling datacenter

Frequency increases linearly to scaling datacenter

🫀 Heartbeating

Heartbeating is a mechanism for failure detection where nodes increment their own heartbeat and send them to neighbors → ping means “I am alive, here is my heartbeat”