Design Twitter

Question 1

Who calls this system?

Accepted Answer

A quarter of a billion people, opening the timeline fifteen times a day and posting twice. Start from the caller and that asymmetry is immediate.

Question 2

What do the apps connect to?

Accepted Answer

A gateway holds authentication and rate limiting, and is where the read path and the write path stop being the same traffic.

Question 3

Posting and reading are 7,000× apart in volume. Are they the same service?

Accepted Answer

They should not be. Reading is constant, latency-critical and cacheable; posting is rare, bursty and tolerant of a second of delay. Splitting them lets a fan-out backlog slow posts down without touching anybody’s timeline.

Question 4

An account with fifty million followers posts. What happens inside that request?

Accepted Answer

Nothing that touches fifty million timelines. The write records the post and puts the fan-out on a queue; workers grind through it afterwards. Doing it inline turns one popular post into a site-wide outage, and doing it asynchronously turns it into a backlog.

Question 5

Where does an assembled timeline live between opens?

Accepted Answer

In memory, per user, pre-merged. Rebuilding a timeline from the posts table on every open means a query per followed account, forty-three thousand times a second. A materialised timeline turns the read into one lookup.

Question 6

Where do posts and the follow graph actually live?

Accepted Answer

Timelines are derived and can be rebuilt. The posts and the edges of the graph cannot be — they are the truth every cache is a view of.

Question 7

Your client reads the posts store directly. Who applies blocks and private accounts?

Accepted Answer

A timeline is a permissioned, personalised view. Every rule about who may see what lives in a service, and a direct connection skips all of them.

Question 8

How many timeline reads a second are you designing for?

Accepted Answer

250M users × 15 opens is 3.75B a day, and 3.75e9 / 86,400 is about 43,000 a second on average — several times that at peak.

Question 9

And how many posts a second?

Accepted Answer

500,000,000 / 86,400 is about 5,800 a second. Tiny next to the reads — until you multiply each one by its author’s follower count, which is where the real write amplification hides.

Question 10

Push posts into follower timelines, or pull them at read time?

Accepted Answer

Push makes reads trivial and makes a 100M-follower post catastrophic. Pull makes writes trivial and makes every read a fan-in query. Real systems do both — push for the long tail, pull for the hot accounts, merged at read time — and being able to say why is the whole question.

Blueprint — system design