Design Google Docs

Question 1

What is the browser doing here?

Accepted Answer

More than displaying. It applies its own edit immediately so typing feels instant, then reconciles when the server answers — which is exactly why two clients can diverge and need merging.

Question 2

How does one person’s keystroke reach the other four editors within a second?

Accepted Answer

Not by polling. Every editor holds an open connection, and something has to own those connections and know which ones are on which document. That layer is its own service because it scales with connections, not with documents.

Question 3

Two edits arrive for the same character position. Who resolves that?

Accepted Answer

One authority per document, applying operations in a defined order. Without a single serialisation point per document, two servers accept conflicting edits and no amount of clever merging afterwards recovers the intent.

Question 4

The server holding the document dies mid-session. What survives?

Accepted Answer

Whatever was written down. The document is an append-only log of operations plus periodic snapshots, so recovery is replaying a few seconds of edits — not asking five people what they had typed.

Question 5

Where does "who is in this document right now, and where is their cursor" live?

Accepted Answer

Presence changes many times a second, is worthless a moment later, and is read on every broadcast. That is an in-memory store — putting cursor positions in a database is writing 2.5 million rows a second that nobody will ever read again.

Question 6

Who else needs to know about an edit besides the other editors?

Accepted Answer

Search indexing, version history, offline export and notifications all care about edits, and none of them should be able to add latency to somebody’s keystroke. Publish the stream and let them consume it.

Question 7

Your browser writes to the operation log directly. What stops it writing to a document it cannot open?

Accepted Answer

Sharing rules, view-only access and the merge itself are all decisions. A client that appends straight to the log bypasses every one of them and can corrupt the document for everyone.

Question 8

How many edit operations a second are you designing for at peak?

Accepted Answer

500,000 concurrent editors at 5 edits a second is 2.5 million operations a second. That is the number that says an edit cannot be a database transaction, and that most of this system has to live in memory.

Question 9

How do two concurrent edits to the same position end up agreeing?

Accepted Answer

Last-write-wins loses somebody’s sentence. The two real answers are operational transformation — the server rewrites each incoming operation against those already applied — and CRDTs, whose operations are designed to commute so any order converges. Knowing that both exist, and what each costs, is what this question is asking.

Blueprint — system design