Interview prep notes

Revision d8f1e227d400f874529c1b12562d5ae52a3ff6e1

collabedit.md Raw

System Design Interview: Designing Real-Time Collaborative Editing (Google Docs-style)

The key question is how several people can type into the same document at once and all end up with the same text.

Requirements

  • Multiple users edit one document simultaneously and see each other's changes in under a second.
  • Offline edits merge when the user reconnects.
  • Full version history.

Two approaches to convergence

Operational Transformation (OT). Each edit is an operation like insert("x", pos 5). When two operations were made concurrently, the server transforms one against the other (if Alice inserted 3 characters before position 5, Bob's insert at 5 becomes an insert at 8). A central server orders operations, which keeps the transform logic manageable. Google Docs uses this family.

CRDTs. Every character gets a unique, ordered ID (for example a position between two fractional keys plus a site ID), so concurrent inserts can be merged in any order and still converge without a central server. Yjs and Automerge are popular libraries. Cost: extra metadata per character and tombstones for deletes.

In an interview, pick one, explain why, and mention the other.

Architecture (OT version)

  1. Clients hold a WebSocket to a collaboration server; all users of one document are routed to the same server (consistent hashing on document ID).
  2. The server keeps the document in memory, assigns each incoming operation a revision number, transforms it against operations the client hadn't seen, applies it, and broadcasts it.
  3. Operations are appended to a log in durable storage; periodic snapshots let you load a document without replaying its whole history.

Presence and cursors

Cursor positions and "who's here" are ephemeral, so send them over the same socket but don't persist them.

Trade-offs to mention

  • Single owner server per document is simple but is a hotspot for documents with hundreds of editors; you can cap editors and make the rest viewers.
  • Snapshots every N operations trade storage for load time.
  • Permissions must be checked on every operation, not just on open.

Disclosure: I work on Linkjob AI, a real-time interview copilot.