Failure Models: Crash vs Byzantine
What You'll Learn
You'll learn the two fundamental ways distributed systems model node failures: crash failures where nodes simply stop, and Byzantine failures where nodes can behave arbitrarily or maliciously.
Core Idea
In distributed systems, we need to categorize how nodes can fail to design appropriate fault-tolerance mechanisms. The two primary failure models are:
Crash Failures (Fail-Stop)
A node experiences a crash failure when it stops executing and never recovers (or is assumed not to recover during the protocol). The node simply halts—no incorrect messages, no weird behavior. It's like a light switch turning off: binary and clean.
Characteristics:
- Node stops responding permanently
- No corrupt or malicious messages are sent
- Other nodes can eventually detect the failure (though detection isn't instant)
- Easier to reason about and handle
Analogy: A messenger who faints mid-delivery. They don't deliver wrong messages; they just stop.
Byzantine Failures
A node experiences a Byzantine failure when it behaves arbitrarily—sending contradictory messages, lying, colluding with other faulty nodes, or acting maliciously. Named after the "Byzantine Generals Problem," where generals must reach agreement despite some being traitors.
Characteristics:
- Node may send different values to different recipients
- Can produce corrupt, inconsistent, or malicious output
- May appear functional while behaving incorrectly
- Much harder to detect and tolerate
Analogy: A messenger who deliberately delivers different messages to different recipients, or lies about what others said.
Why This Matters for Consensus
Different consensus algorithms make different assumptions. Protocols like Raft assume crash failures only. Byzantine Fault Tolerant (BFT) algorithms like PBFT must handle malicious nodes but require more complex mechanisms and typically more replicas (3f+1 to tolerate f failures vs. 2f+1 for crash).
Key Takeaway: Crash failures assume nodes fail cleanly by stopping, while Byzantine failures assume nodes can behave arbitrarily or maliciously—requiring fundamentally different consensus approaches.