Primary-backup replication

Hook

One node in your cluster owns every write. It crashes at 3am. Which of your other nodes gets to take over — and what did you just lose?

Intuition

Click a backup to inspect it

The primary (top) has written 7 entries. Each backup (bottom) has replicated some prefix of that log — click one to see what promoting it right now would cost you. A backup that's behind doesn't have the missing writes anywhere else to recover them from.

Formalize

In primary-backup replication, one designated node — the primary — accepts every write, appends it to its own log, and ships that log to a set of backups. The backups apply entries in the same order, so each one's state is just an earlier prefix of the primary's.

The protocol rule for handling a primary failure:

Find the most caught-up backup
Compare every backup's replicatedUpTo — how many of the primary's log entries it has durably received.
Promote it
The backup with the largest replicatedUpTo becomes the new primary.
Count the damage
Any entry in the old primary's log beyond that backup's replicatedUpTo is permanently lost — there's no third copy to recover it from.

Primary log — the full, authoritative sequence of writes, in order.
replicatedUpTo — how far a given backup has replicated into that log.
Writes lost — primary log length − replicatedUpTo of whichever backup gets promoted.

Play

Best to promote: B2 — 0 write(s) would be lost. Click a backup to replicate one more write into it.

Click a lagging backup a few times to replicate it forward. Watch the "best to promote" readout update — and notice the writes-lost count shrink toward zero as a backup catches up. Replication lag isn't a bug; it's the gap that determines how much a crash costs you.

Worked example

The primary has written 7 entries: w1 … w7. At the moment it crashes, the three backups are at:

Read off the replication state
- B1 has replicated 5 entries
- B2 has replicated all 7
- B3 has replicated only 3
Pick the most up-to-date backup
B2, at replicatedUpTo = 7, is furthest along — it becomes the new primary.
Count writes lost for the winner
$7 - 7 = 0$ . Promoting B2 loses nothing.
Compare to the alternatives
- Promoting B1 instead would have lost $7 - 5 = 2$ writes (w6, w7)
- Promoting B3 would have lost $7 - 3 = 4$ writes (w4–w7)

Checkpoint

The primary just crashed. Click the backup that should be promoted — the one that loses the fewest writes.

Primary log has 7 writes — click a backup

Click a backup to promote it

Summary

Primary-backup replication is simple: one node takes writes, ships a log, backups replay it. But the primary is a single point of failure, and a crash always costs you whatever the best-replicated backup hadn't yet received. The next chapter's quorum systems are one way to bound that cost precisely instead of leaving it to chance.