News for hackers. Stuff that matters.
Re:It’s Time to Move on from Two Phase Commit by pacala on Friday January 25, 2019 @02:20PM (#19001434)
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.

Re:It’s Time to Move on from Two Phase Commit by abadid on Friday January 25, 2019 @02:50PM (#19001791)
I understood and agree with all of your stated assumptions. However, I didn't follow your naive implementation. I would definitely caution against using a coordinator.
Re:It’s Time to Move on from Two Phase Commit by pacala on Friday January 25, 2019 @02:59PM (#19001898)
The transaction must enter the system somewhere. The 'coordinator' is that entry point, with a durable log so we don't inadvertently lose the transaction. It's only job is to push the transaction to all the workers involved.