Byzantine consensus
Overview
Crash-fault consensus assumes a failed node stops and says nothing false. Byzantine consensus makes no such assumption: a faulty node may send different messages to different peers, lie about what it received, and actively work against the outcome. The consequence is arithmetic — tolerating f Byzantine faults requires 3f+1 replicas rather than 2f+1, and every guarantee must hold against an adversary choosing the worst messages at the worst moments.
This skill covers that model and the machinery built on it: why 3f+1 is the minimum, how the three-phase prepare-commit structure lets replicas converge on one order despite equivocation, how view change replaces a primary that may itself be the fault, and how a client decides a reply is trustworthy without trusting any single node. It ends at the boundary of the crash-fault family, which belongs to another skill.
Primary sources: https://pmg.csail.mit.edu/papers/osdi99.pdf and https://en.wikipedia.org/wiki/Byzantine_fault.
When to use
- Reading or evaluating a system that claims Byzantine fault tolerance.
- Reasoning about why a deployment needs 3f+1 replicas and what f it buys.
- Following a view change when the primary is suspected faulty.
- Deciding between signed messages and MACs for message authentication.
When NOT to use
- For crash-fault protocols of the Raft or Paxos family (use ac-consensus-protocols).
- For quorum intersection in majority systems (use ac-consensus-protocols).
- For hardening the endpoints themselves (use ac-security-engineer).
Subskill Index
| Leaf | Domain segment | Scope (one line) |
|---|---|---|
subskills/byzantine-versus-crash.md |
software-engineering/byzantine-versus-crash |
The failure model: what a Byzantine node may do that a crashed one cannot. |
subskills/replica-arithmetic.md |
software-engineering/replica-arithmetic |
Why 3f+1 is the minimum, and the quorum sizes it produces. |
subskills/three-phase-commit.md |
software-engineering/three-phase-commit |
Pre-prepare, prepare, and commit, and what each round eliminates. |
subskills/view-changes.md |
software-engineering/view-changes |
Replacing a suspected primary without losing committed work. |
subskills/client-quorums.md |
software-engineering/client-quorums |
How a client trusts a reply without trusting any single node. |
subskills/signatures-versus-macs.md |
software-engineering/signatures-versus-macs |
Authentication cost, message counts, and the recovery post-surgery. |
subskills/byzantine-record.md |
software-engineering/byzantine-record |
The written record of f, quorums, authentication, and evidence. |
Status and honest limits
Byzantine tolerance is expensive: more replicas, more message rounds, more cryptographic work, and a threat model that must be stated to mean anything. A deployment that never names its f has not chosen a fault tolerance, and a system that mixes the crash-fault vocabulary with Byzantine claims is usually one or the other. Nothing here recommends a protocol, and no Byzantine system tolerates faults in the clients or in the code every replica shares — the model covers the replicas, not the software's own correctness.