Legend:

🔵 Quite sure

🟠 With doubts

⚫ TBD

Disputes & Open Questions

Concurrency remains an open problem. Multiple peers may modify the same data concurrently, leading to conflicts that must be resolved deterministically.
Event linearization vs. decentralization. Today, servers may linearize events to simplify consistency, but this introduces centralization. Can we preserve correctness without relying on a single authority?
Optimistic concurrency control. Assume operations succeed locally and resolve conflicts later. This requires robust conflict-resolution semantics.
Time conflicts. Wall-clock time is unreliable across devices. Logical clocks (e.g. vector clocks or Lamport clocks) may be necessary.

Blob vs. operations. Should we store raw state snapshots, operation logs, or a hybrid?
CRDTs as a foundation. CRDTs can help resolve conflicts without coordination.
- Operation-based CRDTs vs. state-based CRDTs remains an open choice.
- State-based CRDTs require enough metadata/history to converge correctly.
Deletion semantics. Deletion implies tombstones.
- How long do we keep them?
- Can we garbage-collect safely without breaking convergence?

Per-device vector clocks. Tracking causality per device could enable better merges and conflict resolution.
Signed operations. Operations can be cryptographically signed and distributed peer-to-peer, removing the need for a central server to establish authorship or validity.

Subscription model. Should subscriptions be merged or deduplicated across devices and peers?
Vault-based notifications. One option is storing notification state (or a notification server reference) inside the user’s vault.
Centralization tradeoff. Email and push notifications likely require some form of central service—this may be unavoidable.

Invocation-based authentication. Should all actions be authorized via explicit invocations?
Session keys. Alternatively, short-lived session keys could reduce overhead for frequent operations.
Hybrid approach? Invocations for delegation, session keys for performance.

Private documents between peers. How far can we go without servers?
Zero-knowledge techniques. ZK proofs or encrypted capabilities may allow sharing metadata or permissions without revealing content.
Selective disclosure. Can peers prove authorization without exposing document contents or full histories?