A recovery mechanism of MapReduce model based on transferring dependency
Zhuyu Li, Fucong Li, Xiangxin Li · Advances in computer science research · 2015
MapReduce is a popular big data processing tool, but there exists a main node failure problem.In order to make the restarted JobTracker recover jobs site quickly, there are mainly two recovery mechanisms: synchronization mechanism and backup mechanism.In synchronization mechanism, TaskTracker sends synchronous message to JobTracker within required time in order to recover jobs without considering dependency between Map and Reduce tasks, which leads to congestion.Using backup mechanism, JobTracker recovers jobs based on backup records ignoring assigned tasks between backup intervals (which is called zombie tasks).That will result in repeated task assignment.We come up with a new backup mechanism based on transferring dependency, with the addition of dependency list mechanism and zombie task recovery mechanism.The new mechanism effectively solves problems of congestion and duplicated distribution of zombie tasks, making the jobs continue running from the breakpoint.