----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/160/ -----------------------------------------------------------
Review request for hadoop-mapreduce, Dhruba Borthakur and Ramkumar Vadali. Summary ------- Raided file introduce extra dependencies on the blocks on the same stripe. Therefore we need a new way to place the blocks. It is desirable that raided file satisfies the following two conditions: a. Replicas on the same stripe should be on different machines (or racks) b. Replicas of the same block should be on different racks MAPREDUCE-1831 will try to delete the replicas on the same stripe and the same machine (a). But in the mean time, it will try to maintain the number of distinct racks of one block (b). We cannot satisfy (a) and (b) at the same time with the current logic in BlockPlacementPolicyDefault.chooseTarget(). One choice we have is to change BlockPlacementPolicyDefault.chooseTarget(). However, this placement is in general good for all files including the unraided ones. It is not clear to us that we can make this good for both raided and unraided files. So we propose this idea that when raiding the file. We create one more off-rack replica (so the replication=4 now). Than we delete two blocks using the policy in MAPREDUCE-1831 after that (replication=2 now). This way we can rearrange the replicas to satisfy (a) and (b) at the same time. Diffs ----- trunk/src/contrib/raid/src/java/org/apache/hadoop/raid/DistRaid.java 1040840 trunk/src/contrib/raid/src/java/org/apache/hadoop/raid/RaidNode.java 1040840 trunk/src/contrib/raid/src/test/org/apache/hadoop/raid/TestRaidNode.java 1040840 Diff: https://reviews.apache.org/r/160/diff Testing ------- Thanks, Scott
