[HDDS-2347] XCeiverClientGrpc's parallel use leads to NPE

• Type: Improvement
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 0.5.0
• Component/s: Ozone Client
• Labels: pull-request-available
• Assignee: István Fajth
• Reporter: István Fajth
• Created: 22/Oct/19 11:25
• Updated: 24/Feb/20 20:55
• Resolved: 30/Oct/19 07:02

Description

This issue came up when testing Hive with ORC tables on Ozone storage backend, I so far I could not reproduce it locally within a JUnit test but the issue.

I am attaching a diff file that shows what logging I have added in XCevierClientGrpc and in KeyInputStream to get the results that made me arrive to the following understanding of the scenario:

• Hive starts a couple of threads to work on the table data during query execution
• There is one RPCClient that is being used by these threads
• The threads are opening different stream to read from the same key in ozone
• The InputStreams internally are using the same XCeiverClientGrpc
• XCeiverClientGrpc throws the following NPE intermittently:

Caused by: java.lang.NullPointerException
        at org.apache.hadoop.hdds.scm.XceiverClientGrpc.sendCommandAsync(XceiverClientGrpc.java:398)
        at org.apache.hadoop.hdds.scm.XceiverClientGrpc.sendCommandWithRetry(XceiverClientGrpc.java:295)
        at org.apache.hadoop.hdds.scm.XceiverClientGrpc.sendCommandWithTraceIDAndRetry(XceiverClientGrpc.java:259)
        at org.apache.hadoop.hdds.scm.XceiverClientGrpc.sendCommand(XceiverClientGrpc.java:242)
        at org.apache.hadoop.hdds.scm.storage.ContainerProtocolCalls.getBlock(ContainerProtocolCalls.java:118)
        at org.apache.hadoop.hdds.scm.storage.BlockInputStream.getChunkInfos(BlockInputStream.java:169)
        at org.apache.hadoop.hdds.scm.storage.BlockInputStream.initialize(BlockInputStream.java:118)
        at org.apache.hadoop.hdds.scm.storage.BlockInputStream.read(BlockInputStream.java:224)
        at org.apache.hadoop.ozone.client.io.KeyInputStream.read(KeyInputStream.java:173)
        at org.apache.hadoop.fs.ozone.OzoneFSInputStream.read(OzoneFSInputStream.java:52)
        at org.apache.hadoop.fs.FSInputStream.read(FSInputStream.java:75)
        at org.apache.hadoop.fs.FSInputStream.readFully(FSInputStream.java:121)
        at org.apache.hadoop.fs.FSDataInputStream.readFully(FSDataInputStream.java:112)
        at org.apache.orc.impl.ReaderImpl.extractFileTail(ReaderImpl.java:555)
        at org.apache.orc.impl.ReaderImpl.<init>(ReaderImpl.java:370)
        at org.apache.hadoop.hive.ql.io.orc.ReaderImpl.<init>(ReaderImpl.java:61)
        at org.apache.hadoop.hive.ql.io.orc.OrcFile.createReader(OrcFile.java:105)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator.populateAndCacheStripeDetails(OrcInputFormat.java:1708)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator.callInternal(OrcInputFormat.java:1596)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator.access$2900(OrcInputFormat.java:1383)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator$1.run(OrcInputFormat.java:1568)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator$1.run(OrcInputFormat.java:1565)
        at java.security.AccessController.doPrivileged(Native Method)
        at javax.security.auth.Subject.doAs(Subject.java:422)
        at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1876)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator.call(OrcInputFormat.java:1565)
        at org.apache.hadoop.hive.ql.io.orc.OrcInputFormat$SplitGenerator.call(OrcInputFormat.java:1383)
        at java.util.concurrent.FutureTask.run(FutureTask.java:266)

I have two proposals to fix this issue, one is the easy answer to put synchronization to the XCeiverClientGrpc code, the other one is a bit more complicated, let me explain below.

Naively I would assume that when I get a client SPI instance from XCeiverClientManager, that instance is ready to use. In fact it is not, and when the user of the SPI instance sends the first request that is the point when the client gets essentially ready. Now if we put synchronization to this code, that is the easy solution, but my pragmatic half screams for a better solution, that ensures that the Manager essentially manages the clients that is giving to it's users, and the clients themselves are not getting ready by accident.
I am working on a proposal that moves things around a bit, and looking for possible other solutions that does not feel hacky as I feel with the easy solution.

I am attaching the followings:

• a diff that shows the added extended logging in XCeiverClientGrpc and KeyInputStream.
• a job log snippet from a Hive query that shows the relevant output from the extensive logging added by the diff in a cluster.
• later a proposal for the fix I need to work on a bit more

Issue Links

• is related to: HDDS-2362 XCeiverClientManager issues
• links to: GitHub Pull Request #81