CRDT

The property is mathematical rather than procedural: the merge operation is commutative, associative and idempotent, so replicas that have seen the same set of updates in any order converge to identical state. No central authority decides, no locks are taken, and no user is asked to resolve anything.

That is what makes real-time collaborative editing work offline, and why the technique underpins most modern collaborative tooling.

The limitation to be honest about is that convergence is not the same as correctness. A CRDT guarantees everyone ends up with the same answer; it does not guarantee the answer is what a human intended, and it cannot enforce a global invariant. A counter that must never go below zero cannot be a plain CRDT counter, because two concurrent decrements each valid alone can sum to an invalid state — which is the same reason CRDTs cannot implement seat inventory.

The practical costs: metadata overhead can exceed the data itself, particularly for text types that must retain tombstones for deleted characters, and garbage collecting that metadata safely across replicas that may be offline for months is genuinely difficult.

The alternative worth comparing against is operational transformation — more compact, requires a central server to order operations, which is a different architecture rather than a worse one.