Interesting. This assumes there is a way to do consistent remote reads, that is some way to implement B.read(value=y, tx=T). Additionally, this works without us

Interesting. This assumes there is a way to do consistent remote reads, that is some way to implement B.read(value=y, tx=T). Additionally, this works without using a global sync point, as per your other message. Naive implementation: * B just blocks the read call until T is ready for execution, then returns the value of y. * The coordinator C pushes the transactions to both A and B until they independently ACK, so B never blocks forever. * Assuming serializable A and B. Would be interesting to see how this performs compared with vanilla 2PC. The one downside I see is that if B vanishes for a prolonged amount of time, A is stuck waiting on its read call without a mechanism to drop the distributed transaction T and go on with its other duties.