Covers 1:03:50–1:09:50. The engineering challenges of removing centralized cloud dependency, illustrated by the decentralized access control problem.

The vision

  • Get away from dependency on centralized cloud services. Cloud services may still be involved (e.g. syncing between phone and laptop) but no single service should be a required trust anchor: run multiple cloud services on multiple providers side by side, sync with whichever responds first — if one disappears, the others remain. 1:04:02 1:04:27

Case study: access control without a central server

  • The problem sounds trivial — grant and revoke collaborator access to a document — and is trivial in a centralized model (roles checked by one server). 1:04:50
  • In a multi-provider or peer-to-peer setting: a user's edit permission is revoked concurrently with that user making an edit. Some devices see edit-then-revocation and accept the edit; others see revocation-then-edit and drop it → devices become permanently inconsistent. 1:05:18 1:05:43
sequenceDiagram
    participant U as Revoked user
    participant D1 as Device 1
    participant D2 as Device 2
    Note over D1: sees edit → revocation
    U->>D1: edit
    D1->>D1: accepts edit
    Note over D2: sees revocation → edit
    U->>D2: edit (arrives after revocation)
    D2->>D2: drops edit (unauthorized)
    Note over D1,D2: permanently inconsistent
  • A consensus protocol could decide the order, but consensus is messy: quorum votes, nodes must be online. The team is solving it without consensus, while preserving high availability, offline work, and serverless peer-to-peer sync — solvable, and close to solved in Automerge, the CRDT library Martin works on. 1:06:08 1:06:44
  • Clocks don't help: a revoked user who wants to vandalize a document can backdate their edits with forged timestamps, undermining clock-based ordering. Actions from end-user devices must be treated as potentially malicious. 1:07:17 1:07:39

Why research takes the hard road

  • Startups rightly take the easy pragmatic route (centralized server — it makes business sense). Research can afford the idealistic stance of solving the harder decentralized problem; solving it creates a building block and optionality for anyone who later wants decentralized tech without inventing it from scratch. 1:08:11 1:08:37 1:09:09