Replication

The three ways to keep the same data in step on more than one machine. The copy clerks (database replication) tells the same three-way split as a story first; this lesson is the full version, in order.

Before you start. The previous lesson, Transactions, was one machine’s promise: a group of writes lands whole or not at all, even with two people touching the same row at once. This lesson sends changes to the same piece of data to several machines instead of one, and a new question comes up: exactly what gets sent, and what a reader sees on a copy that has not caught up yet.

The idea in 30 seconds

One machine, the leader, is usually the only one allowed to write; every other copy, a follower, just replays its log, in order.
Followers spread out reads, not writes - every write still has to reach the leader, then every follower.
A follower is a moment behind by design, not by fault. That moment is replication lag, and a busy leader or a slow follower stretches it.
Your own write vanishing the instant after you make it is read-your-writes: written to the leader, read from a follower that had not caught up.
More than one leader, or no leader at all, trades that one clean answer for availability: multi-leader keeps writes local and risks conflicts; leaderless never refuses a write and gets its safety from a vote instead of a machine.
A quorum is that vote: enough copies written and enough copies read that the two sets are guaranteed to share one.

Read the story · Skip to the numbers