Hi, I looked at HAMA-734 and HAMA-704.patch-v1 today, and it seems quite invalid.
HAMA-734's concept is creating a additional storage for persisting messages. The problem is when there are huge amount of messages to be persisted, like VertexIndexedMessageStorage in HAMA-704.patch-v1 case. And in the worst case, the size of storage can be continuously increased, until starting to read the messages from there. So, we have to consider the memory-efficient storage again, like DiskQueue or Spilling Queue. Moreover, every messages should be polled from incoming queue, and should be determined whether to be stored in queue or additional storage. Simply, I like to add two queue behaviours: Regular and Persistent. If the queue behaviour is regular, all of the unconsumed messages will be lost (reset in every superstep). In persistent case, unconsumed messages will be persisted in incoming queue. If there are no feedbacks, I'll close the HAMA-734 as a invalid and start to work on Queue implementations. -- Best Regards, Edward J. Yoon @eddieyoon
