Skip to content

Consensus & replication · 2 of 6

Step 1 · The pattern

One primary, one replica

Writes go to the primary. It ships each change to a replica, which serves the reads. Splitting the work this way lets the database handle far more traffic.

  • request
  • change being replicated
  • done

In short

Replication lag: how it works and how it breaks

To handle more reads, a database is often run as one primary that accepts writes and several replicas that copy its changes and serve queries. The copying is usually asynchronous: the primary confirms a write first and ships it to the replicas afterwards. The delay, replication lag, is normally milliseconds, but it stretches to seconds under heavy writes, a long transaction or a slow network. During that delay a replica answers with slightly old data. The classic symptom is a user who saves something, is redirected to a page that reads from a replica, and does not see what they just saved. They save again, and now there are two. The usual fix is read-your-writes: for a short time after a user writes, send that user's reads to the primary, or to a replica known to have caught up to the write's position. Everyone else can keep reading from replicas.

How this comes up in interviews

"A user saves a comment, the page reloads, and the comment is missing. A moment later it appears. Your reads go to replicas. What happened and how do you fix it?"

Then they ask

  1. Why not send every read to the primary? What do you give up?
  2. You pin a user to the primary for a few seconds after a write. When is that timer wrong, and what is more precise?
  3. With several replicas, a user sees a comment and then loses it on refresh. What guarantee is missing, and how do you provide it?

A strong answer covers

  • Replication is usually asynchronous: the primary confirms a write first and ships it afterwards, so a replica serves slightly old data during the delay.
  • Lag is normally milliseconds but stretches to seconds under heavy writes, a long transaction or a slow network.
  • The classic symptom is a user who cannot see their own write and saves it again, which creates a duplicate.
  • Read-your-writes: for a short time after a user writes, send that user's reads to the primary or to a replica known to have caught up.
  • Everyone else keeps reading from replicas, so the primary does not carry all the read load.

Points to cover, not a script. Say them in your own words and tie them to a system you have worked on.

On one page: the track cheat sheet.

Last updated: Official documentation

Diagrams are our own simulations: the timings, rates and counts in them are illustrative, not measurements. Captions are original wording.