The last time Hackerfall tried to access this page, it returned a not found error. A cached version of the page is below, or click here to continue anyway


EPaxos eliminates leader and solves the associated problems. It seems that Paxos saga (Paxos, Disk-, Cheap-, Fast-, Generalized-, Vertical- Paxoses, Mencius and others) is over and we have found the holy grail, havent we?

Actually there a lot of undiscovered yet EPaxos-like systems with different operational characteristics. Let me demonstate it. EPaxoss authors wrote:

The main goals when designing EPaxos were: (1) optimal commit latency in the wide area, (2) optimal load balancing across all replicas, to achieve high throughput, and (3) graceful performance degradation when some replicas become slow or crash

They achieves the goals by dealing with a set of objects and doing updates concurrently if they are non-interfering. It sounds just like transaction. By this time it should be clear how build next 700 Paxos algorithms:

  1. Run a bunch of single degree Paxos state machines
  2. Peak a favorite transactions model to support multi-object updates e.g. Sagas, RAMP transactions, Yabandehs transactions [1] [2] [3] or 2PC
  3. Remove redundancy and merge transactions with Paxos into one algorithm
  4. Write a paper on new algorithm

Personally I'm fascinated by sagas since they allow to implement different distributed system like banking system and to support transaction that would require impressive one round trip in most cases and two in the worst case.

BTW, If youre developing a real-world system then you probably want to:

  1. Make system works in a dynamic environment
  2. Support sharding

Continue reading on