Leveling up our transaction system with an extra guarantee: no dirty reads. We've upgraded our consistency model while keeping total availability.
Each key stores a list of versioned values
Values track their committed state
Reads only see committed values
Writes live in transaction until commit time
Reads only see the latest committed value
Writes are staged in the transaction object
Atomic commit applies all transaction writes
Committed writes are replicated async to other nodes
Support for transaction abort (required by spec)
Versioned Values: Only committed values are visible to reads
Transaction Staging: Buffer all writes until commit
Atomic Commits: All writes become visible at once
Async Replication: Fire-and-forget keeps us totally available
No Wait-for-Ack: We keep going even if nodes are partitioned
Pros:
Still relatively simple design
Remains totally available during partitions
Prevents all anomalies (G0, G1a, G1b, G1c) as required
No distributed coordination needed
Final form! Complete consistency path from single node → read-uncommitted → read-committed
Cons:
Storage overhead from versioning
Still weak consistency (no repeatable reads)
Values eventually replicate, but no guarantees when
Partition recovery requires no special handling
Transaction ID generation is ultra-simple (but works!)
The subtle art of keeping enough consistency to pass the test while staying totally available. i think this was the last challenge, so ill see you when i see you!
ref: Read Committed prohibits:
G0 (dirty write): transaction write conflicts
G1a (aborted read): reading data from aborted transactions
G1b (intermediate read): reading partial/intermediate transaction results
G1c (circular information flow): transactions creating circular dependencies
