[HDFS-9290] DFSClient#callAppend() is not backward compatible for slightly older NameNodes

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 2.7.1
• Fix Version/s: 2.8.0, 2.7.2, 3.0.0-alpha1
• Target Version/s: 2.7.2
• Hadoop Flags: Reviewed
• Assignee: Tony Wu
• Reporter: Tony Wu
• Created: 22/Oct/15 19:57
• Updated: 05/Jan/17 22:50
• Resolved: 23/Oct/15 21:49

Description

HDFS-7210 combined 2 RPC calls used at file append into a single one. Specifically getFileInfo() is combined with append(). While backward compatibility for older client is handled by the new NameNode (protobuf). Newer client's append() call does not work with older NameNodes. One will run into an exception like the following:

java.lang.NullPointerException
        at org.apache.hadoop.hdfs.DFSOutputStream.isLazyPersist(DFSOutputStream.java:1741)
        at org.apache.hadoop.hdfs.DFSOutputStream.getChecksum4Compute(DFSOutputStream.java:1550)
        at org.apache.hadoop.hdfs.DFSOutputStream.<init>(DFSOutputStream.java:1560)
        at org.apache.hadoop.hdfs.DFSOutputStream.<init>(DFSOutputStream.java:1670)
        at org.apache.hadoop.hdfs.DFSOutputStream.newStreamForAppend(DFSOutputStream.java:1717)
        at org.apache.hadoop.hdfs.DFSClient.callAppend(DFSClient.java:1861)
        at org.apache.hadoop.hdfs.DFSClient.append(DFSClient.java:1922)
        at org.apache.hadoop.hdfs.DFSClient.append(DFSClient.java:1892)
        at org.apache.hadoop.hdfs.DistributedFileSystem$4.doCall(DistributedFileSystem.java:340)
        at org.apache.hadoop.hdfs.DistributedFileSystem$4.doCall(DistributedFileSystem.java:336)
        at org.apache.hadoop.fs.FileSystemLinkResolver.resolve(FileSystemLinkResolver.java:81)
        at org.apache.hadoop.hdfs.DistributedFileSystem.append(DistributedFileSystem.java:336)
        at org.apache.hadoop.hdfs.DistributedFileSystem.append(DistributedFileSystem.java:318)
        at org.apache.hadoop.fs.FileSystem.append(FileSystem.java:1164)

The cause is that the new client code is expecting both the last block and file info in the same RPC but the old NameNode only replied with the first. The exception itself does not reflect this and one will have to look at the HDFS source code to really understand what happened.

We can have the client detect it's talking to a old NameNode and send an extra getFileInfo() RPC. Or we should improve the exception being thrown to accurately reflect the cause of failure.

Issue Links

• is broken by: HDFS-7210 Avoid two separate RPC's namenode.append() and namenode.getFileInfo() for an append call from DFSClient