Design WhatsApp

Question 1

Who is sending messages?

Accepted Answer

The phone is half of this design — it holds the connection and it dedupes on redelivery.

Question 2

A message arrives for a phone. How does it get there without the phone asking?

Accepted Answer

Polling cannot deliver in a second at this scale. Something has to hold a long-lived connection per device and know which server holds which one — that layer is its own service, and it is the piece most designs miss.

Question 3

The recipient is offline. Where does the message wait?

Accepted Answer

A message is accepted before it is delivered. The queue is what makes the send fast and the delivery retryable — without one, sending is coupled to whether the other phone happens to be awake.

Question 4

Where do undelivered messages survive a restart?

Accepted Answer

A queue in memory loses a day of messages when the process dies. Anything not yet delivered has to be durable.

Question 5

Which of your servers is holding this user’s connection right now?

Accepted Answer

Routing a message to the right connection means looking up device-to-server mapping on every single message. That lookup is tiny, constant, and changes whenever a phone reconnects — an in-memory store, not a database.

Question 6

Your client reads the message store directly. What stops it reading someone else’s conversation?

Accepted Answer

Every read has to be checked against who is asking. A direct line to the store skips that check entirely.

Question 7

How many messages a second is 40B a day?

Accepted Answer

40,000,000,000 / 86,400 is about 460,000 messages a second on average, and peaks are several times that. It is the number that rules out anything synchronous on the send path.

Question 8

What delivery and ordering guarantees are you offering, and what do they cost?

Accepted Answer

Exactly-once delivery over a flaky mobile network is not achievable — the honest answer is at-least-once plus de-duplication on the client. Ordering is per conversation rather than global, because a global order means a single writer.

Blueprint — system design