Details
Major Description
"Confined recovery" mentioned in Pregel paper can be used to improve the cost and latency of recovery.
In addition to the existing HDFS checkpoints,1) the tasks log outgoing messages to local filesystem for each superstep (See disk queue). When a task fails, 2) it reverts to the last checkpoint. 3) Other tasks re-send messages sent to failed task at each superstep occurring after the last checkpoint.
Today we write the checkpointed messages to HDFS. We want these files to be written on local filesystem. There should be a way these files could be moved across to optimize the fault recovery process.