CRDT (Conflict-free Replicated Data Type)

A data structure designed so that concurrent updates from multiple nodes always converge to the same result regardless of order, without coordination. For applicable data types, this makes the cap-theorem|CAP tradeoff disappear: the math handles consistency, so the network only needs to deliver updates eventually.

The two guarantees

For a data structure to be a CRDT:

Commutative — order of operations doesn’t matter (a ∘ b == b ∘ a).
Idempotent — duplicate messages don’t corrupt state (a ∘ a == a).

Guarantee both and concurrent writes on different nodes always converge once they exchange state. No locking, no consensus protocol, no central authority.

Common types

Type	Use case
G-Counter (grow-only)	Page views, likes — increments only
PN-Counter (positive-negative)	Inventory, balances — increments and decrements
OR-Set (observed-remove)	Tags, members — add/remove with last-add-wins on conflict
LWW-Register (last-writer-wins)	Single-value fields with a tiebreaker (timestamp + node id)

(Sequence CRDTs for collaborative text — RGA, Yjs’s YATA, egwalker — are a richer family layered on top.)

Worked example: YouTube view counter

Multiple servers in different regions incrementing the same counter simultaneously, no locks:

Each node tracks its own increments locally.
Merge by summing per-node contributions.
Addition is commutative → no race condition possible.
It mathematically cannot lose a count.

What CRDTs do not solve

Uniqueness constraints — “this email must be globally unique” still needs coordination at write time. CRDTs don’t replace consensus for invariants that span the whole system.
Most business operations — cross-service workflows (charge card → ship order → notify) are not CRDTs. Use the saga-pattern there.
Arbitrary application state — every type that needs conflict-free merge has to be designed for it; you can’t sprinkle CRDT-ness on existing objects.

Real-world adoption (from the source video)

System	CRDT status
Redis Enterprise (geo-replication)	True CRDTs, no central coordinator
GoodNotes	Ships automerge to millions of monthly users
Figma (canvas)	Influenced, not true CRDTs — central server serializes, last-writer-wins per property
Figma (text editing, 2025)	Uses egwalker — constructs a CRDT structure transiently to resolve text conflicts, then discards it
Notion	CRDT-influenced
Apple iCloud	CRDT-influenced

Libraries

yjs — most-deployed JS/TS CRDT framework, dominant in collaborative-editor space
automerge — JSON-like CRDT, document-oriented, GoodNotes-grade production use
egwalker — text-only sequence CRDT algorithm (Josephg, 2024)