Along the northern border wall stood seven beacon towers, each barely ten li from the next, linked only by drumbeats and firelight. Every night’s warning-fire signal had to follow one unified rhythm, set by a single chief signalman — if all seven towers kept their own beat, the signals would turn to noise, and friendly troops couldn’t tell a real warning from a false one. So at any given moment, the whole stretch of wall could recognize exactly one chief signalman.

While in office, the chief signalman struck a drum once every watch to tell the towers “I’m still here, follow my lead” — and every tower kept an ear out for that beat. If a tower went two watches straight without hearing it, its keeper would start to wonder whether the chief had met with trouble. But he wouldn’t jump in right away — he’d first wait out a stretch of silence of random length. If, during that wait, some other tower struck its own “I intend to lead” drum first, he’d stand down and let it be. Only if the random silence ran out with the drums still dead would he strike that drum himself, bump the term number he kept in his head up by one, and call out to the other six towers: I mean to be chief for term X — do you recognize me or not?

The rule was unbending: to actually take the seat, a tower had to gather seals of approval from at least four of the seven — a strict majority. Falling short meant the whole attempt counted for nothing. And from that point on, whoever held the higher term number won automatically — every other tower simply overwrote its old memory and fell in line, no second-guessing.

One year, a brutal winter snowstorm sealed the mountain pass and split the seven towers into two isolated clusters: three in the west, four in the east, cut off from each other for half a month. Keepers on both sides independently concluded the chief had gone silent, and both sides tried to elect a new one. The western three could never gather enough votes — even if all three voted for the same candidate, that only made three, one short of the four needed — so no matter how many rounds they tried, no one was ever confirmed, and the west simply sat idle, signaling suspended. The eastern four, on the other hand, managed to gather all four of their own seals cleanly, elected a new chief without a hitch, and kept signaling normally the whole time. When spring thaw restored contact and the drumbeat finally reached the west, the western keepers heard a term number higher than the one they remembered and yielded immediately, no argument — the whole wall was back down to a single chief signalman, with no chaos of two sides shouting conflicting signals into the night.

——By now you’ve probably recognized it: this rule of drumroll handovers, majority-only confirmation, and ever-increasing term numbers is describing the leader election mechanism in the Raft consensus algorithm, and how it relies on a quorum (majority) to guarantee that at most one legitimate leader ever exists at any given moment — which is exactly how etcd avoids split-brain during a network partition.

Concept explanation Raft is one of the most widely used consensus algorithms in distributed systems (it underpins etcd and Consul), and its core job is getting a cluster of nodes to agree on “who is the leader” at all times. Every node maintains a monotonically increasing term. A leader proves it’s alive through periodic heartbeats (AppendEntries with no log entries); once a follower goes past its election timeout without hearing one, it becomes a candidate, increments its term, and sends out RequestVote calls to the rest of the cluster — but it only becomes leader if it wins votes from a strict majority quorum of nodes. That “must be a majority” rule is what mathematically guarantees no two disjoint majorities can ever exist at the same time within one cluster, since any two majority sets of a group must overlap. This is exactly why, during a network partition, the minority side can never elect a leader while the majority side keeps functioning normally — and why etcd clusters are conventionally deployed with an odd number of nodes (3, 5, 7), with “majority of nodes alive” as the hard boundary of availability. The randomized election timeout exists specifically to prevent multiple candidates from starting elections simultaneously and splitting the vote so badly that no one ever wins.

Metaphor mapping

• Chief signalman → leader
• Term number → term
• A drumbeat each watch confirming “I’m still here” → heartbeat (AppendEntries)
• Two watches with no drumbeat → election timeout
• The random stretch of silence before acting → randomized election timeout (prevents split votes)
• Striking the “I intend to lead” drum, calling the other six → a candidate issuing RequestVote
• At least four of seven seals of approval → majority quorum
• The winter storm splitting seven towers into groups of three and four → network partition
• The western three can never scrape together four votes → the minority side can’t reach quorum, no leader elected
• The eastern four cleanly gather their own four votes → the majority side elects a new leader normally
• Higher term wins, old memory simply overwritten → a higher term overrides a stale one, guaranteeing convergence and no split-brain