Dropbox - the case study

The whole system on one board, the sync problem along the top and the disks that finally hold your bytes along the bottom. Every box on it gets a scene below.

Before you start. You need one fact: Dropbox keeps a copy of your files on your own machine and another copy on its machines, and those two copies drift apart whenever you are offline. That is the whole of the background; every other word in the story is explained the first time it is used.

The idea in 30 seconds

Sync is not copying. It is reconciling two histories that were both right, because a laptop in a bag cannot be locked, polled or out-voted.
The root bug was a missing noun: files had no identity apart from their path, so a move was a delete plus an add, and a hiccup between the two lost the file.
A missing noun is the one thing refactoring cannot fix, and it is the only argument that justifies a rewrite. Slow, ugly and hated are all fixable in place.
Two snapshots cannot tell “they added it” from “I deleted it”. The Synced Tree is the merge base that breaks that tie.
The Planner rebuilds the whole plan from the three trees every time, so the plan is never wrong, because it is never stored. Nobody can imagine the edge cases, so a seeded random tester imagines them.
Underneath, a file is a list of fingerprints and the chunks never change, which is what lets the storage layer stay simple enough to own outright.

Read the story · Skip to the numbers