[ https://issues.apache.org/jira/browse/HDFS-13145?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16375172#comment-16375172 ]
Erik Krogen edited comment on HDFS-13145 at 2/24/18 12:55 AM: -------------------------------------------------------------- v0 LGTM overall. Simple fix. Verified that the test fails without your change. Considering that HDFS-10519 is in 3.0, I'm thinking we should target this for branch-3.0 and up? Minor nit: Can we extend the comment above the if-statement explaining that if the log isn't in progress, all edits are guaranteed to be committed? And also maybe a comment on the test case explaining the real world situation it tests, i.e. replace: {code} // Do recovery from a separate QJM. {code} with something like {code} // Do recovery from a separate QJM, like during failover {code} It wasn't immediately obvious to me what was happening. I don't have strong feelings about this one, though. Maybe it's already clear and I just missed it :) Last nit: Should we add a {{checkRecovery()}} to the test? was (Author: xkrogen): v0 LGTM. Simple fix. Verified that the test fails without your change. Considering that HDFS-10519 is in 3.0, I'm thinking we should target this for branch-3.0 and up? > SBN crash when transition to ANN with in-progress edit tailing enabled > ---------------------------------------------------------------------- > > Key: HDFS-13145 > URL: https://issues.apache.org/jira/browse/HDFS-13145 > Project: Hadoop HDFS > Issue Type: Bug > Components: ha, namenode > Reporter: Chao Sun > Assignee: Chao Sun > Priority: Major > Attachments: HDFS-13145.000.patch > > > With edit log in-progress edit log tailing enabled, {{QuorumOutputStream}} > will send two batches to JNs, one normal edit batch followed by a dummy batch > to update the commit ID on JNs. > {code} > QuorumCall<AsyncLogger, Void> qcall = loggers.sendEdits( > segmentTxId, firstTxToFlush, > numReadyTxns, data); > loggers.waitForWriteQuorum(qcall, writeTimeoutMs, "sendEdits"); > > // Since we successfully wrote this batch, let the loggers know. Any > future > // RPCs will thus let the loggers know of the most recent transaction, > even > // if a logger has fallen behind. > loggers.setCommittedTxId(firstTxToFlush + numReadyTxns - 1); > // If we don't have this dummy send, committed TxId might be one-batch > // stale on the Journal Nodes > if (updateCommittedTxId) { > QuorumCall<AsyncLogger, Void> fakeCall = loggers.sendEdits( > segmentTxId, firstTxToFlush, > 0, new byte[0]); > loggers.waitForWriteQuorum(fakeCall, writeTimeoutMs, "sendEdits"); > } > {code} > Between each batch, it will wait for the JNs to reach a quorum. However, if > the ANN crashes in between, then SBN will crash while transiting to ANN: > {code} > java.lang.IllegalStateException: Cannot start writing at txid 24312595802 > when there is a stream available for read: ...... > at > org.apache.hadoop.hdfs.server.namenode.FSEditLog.openForWrite(FSEditLog.java:329) > at > org.apache.hadoop.hdfs.server.namenode.FSNamesystem.startActiveServices(FSNamesystem.java:1196) > at > org.apache.hadoop.hdfs.server.namenode.NameNode$NameNodeHAContext.startActiveServices(NameNode.java:1839) > at > org.apache.hadoop.hdfs.server.namenode.ha.ActiveState.enterState(ActiveState.java:61) > at > org.apache.hadoop.hdfs.server.namenode.ha.HAState.setStateInternal(HAState.java:64) > at > org.apache.hadoop.hdfs.server.namenode.ha.StandbyState.setState(StandbyState.java:49) > at > org.apache.hadoop.hdfs.server.namenode.NameNode.transitionToActive(NameNode.java:1707) > at > org.apache.hadoop.hdfs.server.namenode.NameNodeRpcServer.transitionToActive(NameNodeRpcServer.java:1622) > at > org.apache.hadoop.ha.protocolPB.HAServiceProtocolServerSideTranslatorPB.transitionToActive(HAServiceProtocolServerSideTranslatorPB.java:107) > at > org.apache.hadoop.ha.proto.HAServiceProtocolProtos$HAServiceProtocolService$2.callBlockingMethod(HAServiceProtocolProtos.java:4460) > at > org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:447) > at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:989) > at org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:851) > at org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:794) > at java.security.AccessController.doPrivileged(Native Method) > at javax.security.auth.Subject.doAs(Subject.java:422) > at > org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1836) > at org.apache.hadoop.ipc.Server$Handler.run(Server.java:2490) > 2018-02-13 00:43:20,728 INFO org.apache.hadoop.util.ExitUtil: Exiting with > status 1 > {code} > This is because without the dummy batch, the {{commitTxnId}} will lag behind > the {{endTxId}}, which caused the check in {{openForWrite}} to fail: > {code} > List<EditLogInputStream> streams = new ArrayList<EditLogInputStream>(); > journalSet.selectInputStreams(streams, segmentTxId, true, false); > if (!streams.isEmpty()) { > String error = String.format("Cannot start writing at txid %s " + > "when there is a stream available for read: %s", > segmentTxId, streams.get(0)); > IOUtils.cleanupWithLogger(LOG, > streams.toArray(new EditLogInputStream[0])); > throw new IllegalStateException(error); > } > {code} > In our environment, this can be reproduced pretty consistently, which will > leave the cluster with no running namenodes. Even though we are using a 2.8.2 > backport, I believe the same issue also exist in 3.0.x. -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org