ELI5 · Distributed systems

Quorum.

Nothing counts until most copies say yes — and any two majorities always share a witness.

Five flatmates, one rule: a restaurant booking stands once three of you have thumbed it up in the group chat. Can the flat end up double-booked for Friday? It cannot — any two groups of three drawn from five must share a person, and that person already said yes to the first booking. The overlap does the policing, and nobody has to.

Distributed systems run on the same arithmetic. Data lives on several replicas, and waiting for all of them is fragile, because one dead machine jams everything, while trusting just one is reckless. So operations wait for a majority. Any majority overlaps any other, which means the system cannot be talked into two contradictory states.

  1. MOTION
    1

    Each member keeps a copy of the record; the motion on the table is a write asking to change it.

  2. to pass: 3 of 5 more than half, or nothing
    2

    Three of five is the smallest count that no rival group can also reach.

  3. Carried. Tell the absentees.
    3 yes — it carries
    3

    Two members never voted, and the motion stands anyway — one slow machine cannot jam the works.

  4. shared two majorities always overlap
    4

    Pick any two majorities from the same five: somebody sits in both, and remembers what they signed.

  5. 3 — acts 2 — waits
    5

    A partition cannot mint extra members — at most one side can scrape together three.

  6. R + W > N overlap = fresh reads
    6

    Reads play the same overlap game: make readers and writers intersect, and at least one copy consulted is current.

A committee that needs a majority to act — so two halves of a split can never both decide.

The overlap does the remembering

Majorities work without anyone seeing the whole cluster. Each node counts only the acknowledgements it can reach, and arithmetic supplies the guarantee: if a write reached three of five, any future majority read includes at least one node that took that write. There is no central referee — the referee is the pigeonhole principle. It also explains odd-numbered clusters: five machines tolerate two failures, six still tolerate only two, so the sixth buys no extra safety at all.

R and W are a slider, not a law

Quorum stores let you pick how many replicas a write must reach (W) and how many a read must consult (R). Keep R + W above N and every read overlaps the latest write — strong consistency. Drop R to 1 and reads turn fast and nearly always available, at the price of occasionally being a little stale. Cassandra and its relatives let you choose per query, so the shopping cart and the page-view counter need not pay the same price.

Pairs well with Consensus (Raft)

The real version Read & write quorum simulator →
Found this useful?