Three ways to
agree on state.

A DDS here follows Fluid Framework’s collaborative-object model. Every client applies its own ops optimistically, sends them to a central sequencer, and receives one authoritative order back. Because every replica replays the same ordered stream, they all land in the same state.

The conflict policy applies on top of that order: last-write-wins for a map, first-writer-wins for a claim, FIFO for a queue, quorum for a pact. The order is the source of truth; the policy just decides what the order means.

Simple and cheap, but it needs a sequencer, and offline correctness is only as good as the policy you layer on.

A CRDT is designed so its merge is commutative, associative, and idempotent. Feed two replicas the same set of updates in any order, any number of times, and they converge. No central authority is needed to referee.

watershed still delivers CRDT ops over the same sequenced stream, but their correctness does not depend on it. That is why the demo can feed a duplicate delta straight to the kernel and watch the merge absorb it with no double-count. Commutativity, associativity, and idempotence provide that safety; an order-based DDS relies on the runtime deduping by sequence number.

Its merge can tolerate duplicate and reordered state or deltas, but offline editing still needs durable local storage and a reconnect protocol. Metadata costs depend on the type: some keep per-replica counters, while removable collections retain tags or tombstones.

OT takes the opposite tack from CRDTs. Instead of designing merges that commute, it transforms each op against the concurrent ops it did not see, rewriting indices and positions until any apply order lands in the same place.

watershed ships two OT kernels on the same single-op-in-flight client protocol. json_ot is a faithful port of the ottypes json0 algebra for structured JSON documents. SharedRichText runs that protocol over quill-delta's rich-text algebra (retain/insert/delete spans, attribute patches, embeds) for collaborative Quill editors. In both, concurrent ops have their positions transformed so every replica converges identically.

Minimal per-op metadata and a natural fit for ordered sequences and rich text (SharedSequence now covers similar ground by merge). The cost: transform functions must be correct for every pair of op types.