The result
The FLP impossibility, proved by Fischer Lynch and Paterson in 1985, states that in a fully asynchronous network no deterministic consensus protocol can guarantee termination if even a single node may crash.
Why it holds
In an asynchronous model there is no bound on message delay, so a node cannot distinguish a crashed peer from one whose messages are merely slow. The proof shows there is always a sequence of delays that keeps the protocol in an undecided bivalent state forever.
- No shared clock means no reliable timeout in theory
- One possible crash is enough to force the indecision
- This blocks termination, not agreement or validity
What it does not say
FLP does not say consensus is useless. It says you cannot have all three of safety liveness and fault tolerance with certainty under pure asynchrony.
How real systems escape it
- Partial synchrony assume delays are usually bounded
- Timeouts and failure detectors that may sometimes be wrong
- Randomization to break symmetry probabilistically
Paxos and Raft keep safety always and achieve liveness only when the network behaves.
Key idea
FLP proves no deterministic protocol can guarantee consensus termination under pure asynchrony with one crash, so real systems add timeouts or randomness to make progress in practice.