Design an ad click aggregator

Question 1

Who reports a click?

Accepted Answer

The browser or app the ad was shown in — which means the event can be duplicated, delayed, or forged, and every design decision downstream follows from that.

Question 2

What receives 116,000 events a second?

Accepted Answer

A collector that does almost nothing: validate, stamp, and hand off. Any work it does synchronously is work it does a hundred thousand times a second, and any dependency it has is a dependency that can lose you clicks.

Question 3

Where do clicks go between arriving and being counted?

Accepted Answer

A durable partitioned log. It decouples a spiky ingest from a steady aggregator, lets several consumers read the same events for different purposes, and — most importantly — lets you replay yesterday when you find the counting was wrong. Without it, a bug in aggregation is permanent data loss.

Question 4

What turns ten billion events into a number per campaign per minute?

Accepted Answer

A stream processor holding per-minute windows in memory and emitting them when the window closes. Querying raw events for a dashboard is the thing this design exists to avoid — you aggregate once on the way in, not per viewer.

Question 5

Where do two years of per-minute counts live?

Accepted Answer

10M campaigns × per-minute buckets is a lot of small rows written once and read by time range and campaign — a columnar or time-series store, not the transactional database serving ad delivery.

Question 6

You discover the aggregator has been double-counting for a week. Can you fix the numbers?

Accepted Answer

Only if the raw events were kept. Cheap immutable storage of the original stream is what makes a batch recompute possible, and it is the difference between a bug and an incident you cannot undo.

Question 7

An advertiser refreshes their dashboard every ten seconds. What do they hit?

Accepted Answer

The current minute’s counts, held in memory. The analytics store is built for range queries over history, not for thousands of dashboards asking for the same live number.

Question 8

How many clicks a second is 10B a day?

Accepted Answer

10,000,000,000 / 86,400 is about 116,000 a second. It is the number that says every click cannot be a row insert, and that the counting has to happen in a stream.

Question 9

A click arrives ten minutes late, and another arrives twice. What are the counts?

Accepted Answer

Advertisers are billed on these numbers, so a duplicate is money you did not earn and a dropped event is money you did not charge. The honest answer is at-least-once delivery with de-duplication on a click id, windows closed on a watermark with a grace period, and a batch recompute from the raw log for anything that arrives after that.

Blueprint — system design