Reads that outgrew one database — Architecture trainer

We’re very sorry, but system design is better done on a computer.

Send yourself the link and pick it up at your desk.

Scenario

A news site. Article pages are read about 6,000 times a second at peak, and writes run at about 200 a second — editors publishing and correcting, readers commenting, and a view counter on every article. Articles, comments and counters all live in one SQL database, which is now the ceiling: it sits at 90% CPU all evening, and a page read that used to take 20 ms takes 300 ms.

Requirements

peak is 6,000 article reads and 200 writes a second
the primary database serves at most 600 reads a second
an editor sees their own change the moment they save it
a reader may see an article up to 5 seconds out of date
no acknowledged write is lost if the primary is lost
writes resume within 2 minutes of losing the primary

Constraints

one region, three availability zones
the SQL database stays the source of truth; no move to another store
article pages are personalised — the reader's name, and which comments they have voted on — so one cached copy cannot be served to everybody
two engineers, no dedicated ops

Hints

0/6 requirements ticked