Multi-node transactions with the weakest consistency model known to humanity. Can replication really be this easy?
In-memory map stays from v1, still no persistence
Process writes locally first, broadcast async to other nodes
No need for consensus, just fire-and-forget writes
Reads still return nil for missing keys (feature, not bug!)
Only blocks G0 (dirty writes) via local locking
Async Replication: Fire off writes, don't wait for acknowledgments
Total Availability: Network partitions? No problem, keep operating
Global Dirty Write Prevention: Local locks prevent G0 within nodes
Topology Tracking: Learn about all nodes for broadcasting
No Cost Reads: Read whatever's locally available, it's valid
Pros:
Still pretty simple implementation
No coordination overhead for writes
No need to wait for replicas
Totally available even during partitions
Will pass the tests!
Cons:
Weak consistency guarantees - basically anything goes
Only protects against G0, allowing all kinds of fun anomalies
No guarantees about writes making it to other nodes
Nodes can diverge during partitions
Real read uncommitted would need write locking between nodes (overkill for spec)
Read uncommitted is amusing - it's the "consistency for people who hate consistency" model. Looking forward to the read committed challenge where we'll have to prevent dirty reads too.
ref:
G0 (dirty write): a cycle of transactions linked by write-write dependencies. For instance, transaction T1 appends 1 to key x, transaction T2 appends 2 to x, and T1 appends 3 to x again, producing the value [1, 2, 3].
