Most clustered databases assume replicas crash or pause, they do not equivocate. Castro and Liskov's Practical Byzantine Fault Tolerance (OSDI 1999) targets the other threat: a replica may send different votes to different peers, replay old messages, or collude — up to f such faults — while the rest follow the protocol. The headline math is 3f+1 replicas to tolerate f Byzantine faults, not the 2f+1 of crash-fault Paxos.

That extra third is not a tax for fun. In a crash model, a majority overlap guarantees that two quorums share an honest, persistent acceptor. In a Byzantine model, a majority of 2f+1 could overlap only on a liar. You need quorums large enough that two quorums share an honest replica even after subtracting f traitors.

Three phases around a primary

PBFT uses a primary to assign sequence numbers. Clients send requests to the primary; the primary assigns a seq and broadcasts a pre-prepare. Backups broadcast prepare. When a replica has 2f matching prepares plus its own (a prepared certificate), it broadcasts commit. When it has 2f+1 commits, it executes the request and replies to the client. The client waits for f+1 identical replies so that at least one honest replica executed the same result.

text
pre-prepare  →  prepare (agree on order)  →  commit (agree to execute)
view-change  →  new primary if the old one is silent or diverging

View changes are the operational nightmare. If the primary is slow or malicious, backups collect evidence and move to a new view, carrying prepared requests so they are not lost. A malicious primary that equivocates but stays just fast enough is harder than a crash. Timeouts again sneak in; PBFT is practical because it assumes partial synchrony for liveness, like everyone else.

Where this shows up outside papers

Permissioned blockchains, some certificate transparency designs, and research filesystems (Castro and Liskov's NFS prototype in the paper) used PBFT or cousins. Public proof-of-work and proof-of-stake systems solve a different problem (open membership, sybil resistance) and should not be described as "just PBFT." HotStuff and later BFT protocols reduce phases or pipelining; they still pay 3f+1 in the classic authenticated-channel model.

Do not run PBFT in front of a Postgres that already has a crash-fault primary. Byzantine protocols are for when operators of replicas do not trust each other, or when a compromised host should not silently fork state. Inside one company's VPC with a shared IAM boundary, crash-fault Raft plus good identity is usually the rational threat model. Mixing them — "we're BFT because we have three AZs" — confuses independence of failures with independence of administrators.

Costs you will feel

MACs or signatures on each phase, extra replicas, and view-change complexity dominate. Speculative execution variants improve latency when the primary is honest. Checkpointing and garbage collection of logs are mandatory or memory dies.

If a vendor says "Byzantine fault tolerant" , ask: f of what, 3f+1 of what, and who holds the signing keys. Castro and Liskov gave a protocol you can implement. They did not give you a reason to ignore crash-fault tools when the adversary cannot send two different prepares.