[HDFS-9361] Default block placement policy causes TestReplaceDataNodeOnFailure to fail intermittently

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Component/s: namenode
• Assignee: Unassigned
• Reporter: Wei-Chiu Chuang
• Created: 02/Nov/15 23:04
• Updated: 24/Nov/15 23:15
• Resolved: 23/Nov/15 18:34

Description

TestReplaceDatanodeOnFailure sometimes fail (See HDFS-6101).
(For background information, the test case set up a cluster with three data nodes, add two more data nodes, remove one data nodes, and verify that clients can correctly recover from the failure and set up three replicas)

I traced down and found that some times a client only set up a pipeline with only two data nodes, which is one less than configured in the test case, even though the test case configures to always replace failed nodes.

Digging into the log, I saw:

2015-11-02 12:07:38,634 [IPC Server handler 8 on 50673] WARN  blockmanagement.BlockPlacementPolicy (BlockPlacementPolicyDefault.java:chooseTarget(355)) - Failed to place enough replicas, still in nee
d of 1 to reach 3 (unavailableStorages=[], storagePolicy=BlockStoragePolicy{HOT:7, storageTypes=[DISK], creationFallbacks=[], replicationFallbacks=[ARCHIVE]}, newBlock=true)
org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicy$NotEnoughReplicasException: [
Node /rack0/127.0.0.1:32931 [
  Datanode 127.0.0.1:32931 is not chosen since the rack has too many chosen nodes .
]
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseRandom(BlockPlacementPolicyDefault.java:723)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseRemoteRack(BlockPlacementPolicyDefault.java:624)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseTargetInOrder(BlockPlacementPolicyDefault.java:429)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseTarget(BlockPlacementPolicyDefault.java:342)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseTarget(BlockPlacementPolicyDefault.java:220)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseTarget(BlockPlacementPolicyDefault.java:105)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockPlacementPolicyDefault.chooseTarget(BlockPlacementPolicyDefault.java:120)
        at org.apache.hadoop.hdfs.server.blockmanagement.BlockManager.chooseTarget4NewBlock(BlockManager.java:1727)
        at org.apache.hadoop.hdfs.server.namenode.FSDirWriteFileOp.chooseTargetForNewBlock(FSDirWriteFileOp.java:299)
        at org.apache.hadoop.hdfs.server.namenode.FSNamesystem.getAdditionalBlock(FSNamesystem.java:2457)
        at org.apache.hadoop.hdfs.server.namenode.NameNodeRpcServer.addBlock(NameNodeRpcServer.java:796)
        at org.apache.hadoop.hdfs.protocolPB.ClientNamenodeProtocolServerSideTranslatorPB.addBlock(ClientNamenodeProtocolServerSideTranslatorPB.java:500)
        at org.apache.hadoop.hdfs.protocol.proto.ClientNamenodeProtocolProtos$ClientNamenodeProtocol$2.callBlockingMethod(ClientNamenodeProtocolProtos.java)
        at org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:637)
        at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:976)
        at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:2305)
        at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:2301)
        at java.security.AccessController.doPrivileged(Native Method)
        at javax.security.auth.Subject.doAs(Subject.java:415)
        at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1669)
        at org.apache.hadoop.ipc.Server$Handler.run(Server.java:2299)

So from the log, it seems the policy causes the pipeline selection to give up on the data node.
I wonder whether this is appropriate or not. If the load factor exceeds certain threshold, but the file is insufficient of replicas, should it accept it as is, or should it attempt to acquire more replicas?

I am filing this JIRA for discussion. I am very unfamiliar with block placement, so I may be wrong about my hypothesis.

(Edit: I turned on DEBUG option for Log4j, and changed the logging message a bit to make it show the stack trace)