Crash-Stop vs Byzantine Failures
What you'll learn: You'll understand the two main types of node failures in distributed systems and why distinguishing them matters for consensus design.
When designing consensus algorithms, we must consider how nodes can fail. Not all failures are created equal, and the type of failure dramatically affects the complexity and cost of achieving consensus.
Crash-Stop Failures
Crash-stop (or fail-stop) failures are the "well-behaved" failures. A node simply stops working—it crashes and stays silent. It doesn't respond to requests, doesn't send messages, and doesn't participate further. Think of it like a power outage: the computer just goes dark.
The key property: a crashed node doesn't lie, corrupt data, or send misleading information. It just stops. This makes crash-stop failures relatively easier to handle—you detect silence (timeouts) and exclude the dead node from decisions.
Byzantine Failures
Byzantine failures are the nightmare scenario. A node behaves arbitrarily—it might send different values to different nodes, corrupt data, lie about its state, or even collude with other faulty nodes. The node could be compromised by malware, have hardware corruption, or be actively malicious.
The name comes from the "Byzantine Generals Problem," where generals must coordinate but some might be traitors sending conflicting orders.
Why it matters: Crash-stop failures require simpler algorithms (like Raft or Paxos). Byzantine failures demand much more complex protocols (like PBFT or blockchain consensus) that can tolerate active deception—at significant performance and communication cost.
Most internal distributed systems assume crash-stop failures. Byzantine fault tolerance is reserved for adversarial environments: blockchains, multi-organization systems, or military/aerospace applications where malicious behavior is a real threat.
Key Takeaway: Crash-stop failures mean nodes simply die; Byzantine failures mean nodes can lie, cheat, or behave maliciously—requiring fundamentally different (and costlier) consensus approaches.