The Node That Keeps Answering, and Lies
Subtitle: Silence is easy to spot. A convincing lie is not.
- Panel 1 - Name the fault model:
- It defines how a node may fail
- Every consensus design assumes one
- Fail-stop means crash, then silence
- A missing heartbeat exposes it
- Survivors agree the node is gone
- Panel 2 - The live liar:
- Byzantine means up, but untrustworthy
- Replica A hears X; replica B hears Y
- Health checks can remain green
- One observer sees liar or slow link
- Absence detects nothing
- Panel 3 - Count, with the condition:
- Crash faults f need 2f+1 replicas
- Oral-message BFT needs 3f+1
- For f=1: three nodes versus four
- Signatures can change the bound
- Never quote 3f+1 unconditionally
- Panel 4 - Falsify your assumption:
- Can one live node send conflicting data?
- Can bad bytes still look plausible?
- If yes, fail-stop consensus is not enough
- Raft-family clusters assume fail-stop
- Pay for BFT when peers cannot trust