Notesreplication🌱 seedling
Last Write Wins
Also: LWW · Last Writer Wins
Resolve conflicting versions by keeping the one with the highest timestamp. It always converges, and it silently discards every other concurrent write.
Why it exists
It's the simplest rule that makes replicas agree: no application code, no siblings, no metadata beyond a timestamp.
How it works
- Every write carries a timestamp (client or coordinator clock, or a logical clock).
- On conflict, each replica keeps the higher timestamp; ties break deterministically.
- "Last" is defined by clocks that disagree (Clock Skew): a write that really happened later but was stamped by a slow clock loses.
Trade-offs
| Gains | Costs |
|---|---|
| trivial convergence | acknowledged writes vanish without error |
| no merge logic | correctness depends on clocks |
| small metadata | concurrent updates are lost, not merged |
In practice
- Cassandra applies LWW per cell, using write timestamps.
- DynamoDB global tables resolve concurrent multi-Region writes per item by "last writer wins".
- Jepsen, 2013: Riak with LWW lost 1,434 of 2,000 acknowledged writes (71.7%) under concurrent writes with no partition; a CRDT set lost none.
Pitfalls
- Safe only when keys are written once (immutable events with unique IDs) or losing a concurrent write is acceptable.
- Using it for counters or collections: you get one replica's total, not the sum.
Connections
- Part of
- Write Conflict
- Contrasts with
- CRDTVersion Vector
Recall
These cards are also in your review deck, scheduled with spaced repetition.