Segment: delivering billions of messages exactly once

What you'll be able to do

1. 1Predict where duplicate deliveries come from and how many reach downstream — a lost confirm, a worker crash after the work, a hiding timeout shorter than the work, a producer resending after an unclear answer.
2. 2Choose a guarantee for a job from what losing it or doing it twice would cost, and read a broker's "exactly-once" feature for what it covers — the queue, not the consumer's side effects.
3. 3Explain how a stable idempotency key and a record of handled keys, written together with the work in one transaction, turn at-least-once delivery into effectively-once processing — and why checking first and recording afterwards leaves a gap.
4. 4Predict which repeats a deduplication window misses — a late redelivery, a dead-letter replay, a producer resend after the keys expired — and size the window from how long the broker can redeliver.

Inside the lesson

steps

5

questions

6

simulations

2

You guess first, then run the system and see if you were right.

Step 1 · The setup