纱线群集模式下Spark作业的ClassNotFoundException

Question

So I am trying to run a Spark job on Yarn-cluster mode kicked off via Oozie workflow, but have been encountering the following error (relevant stacktrace below)

java.sql.SQLException: ERROR 103 (08004): Unable to establish connection.
    at org.apache.phoenix.exception.SQLExceptionCode$Factory$1.newException(SQLExceptionCode.java:388)
    at org.apache.phoenix.exception.SQLExceptionInfo.buildException(SQLExceptionInfo.java:145)
    at org.apache.phoenix.query.ConnectionQueryServicesImpl.openConnection(ConnectionQueryServicesImpl.java:296)
    at org.apache.phoenix.query.ConnectionQueryServicesImpl.access$300(ConnectionQueryServicesImpl.java:179)
    at org.apache.phoenix.query.ConnectionQueryServicesImpl$12.call(ConnectionQueryServicesImpl.java:1917)
    at org.apache.phoenix.query.ConnectionQueryServicesImpl$12.call(ConnectionQueryServicesImpl.java:1896)
    at org.apache.phoenix.util.PhoenixContextExecutor.call(PhoenixContextExecutor.java:77)
    at org.apache.phoenix.query.ConnectionQueryServicesImpl.init(ConnectionQueryServicesImpl.java:1896)
    at org.apache.phoenix.jdbc.PhoenixDriver.getConnectionQueryServices(PhoenixDriver.java:180)
    at org.apache.phoenix.jdbc.PhoenixEmbeddedDriver.connect(PhoenixEmbeddedDriver.java:132)
    at org.apache.phoenix.jdbc.PhoenixDriver.connect(PhoenixDriver.java:151)
    at java.sql.DriverManager.getConnection(DriverManager.java:664)
    at java.sql.DriverManager.getConnection(DriverManager.java:208)
    ...
Caused by: java.io.IOException: java.lang.reflect.InvocationTargetException
    at org.apache.hadoop.hbase.client.ConnectionFactory.createConnection(ConnectionFactory.java:240)
    at org.apache.hadoop.hbase.client.ConnectionManager.createConnection(ConnectionManager.java:414)
    at org.apache.hadoop.hbase.client.ConnectionManager.createConnectionInternal(ConnectionManager.java:323)
    at org.apache.hadoop.hbase.client.HConnectionManager.createConnection(HConnectionManager.java:144)
    at org.apache.phoenix.query.HConnectionFactory$HConnectionFactoryImpl.createConnection(HConnectionFactory.java:47)
    at org.apache.phoenix.query.ConnectionQueryServicesImpl.openConnection(ConnectionQueryServicesImpl.java:294)
    ... 28 more
Caused by: java.lang.reflect.InvocationTargetException
    at sun.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method)
    at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:62)
    at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45)
    at java.lang.reflect.Constructor.newInstance(Constructor.java:422)
    at org.apache.hadoop.hbase.client.ConnectionFactory.createConnection(ConnectionFactory.java:238)
    ... 33 more
Caused by: java.lang.UnsupportedOperationException: Unable to find org.apache.hadoop.hbase.ipc.controller.ClientRpcControllerFactory
    at org.apache.hadoop.hbase.util.ReflectionUtils.instantiateWithCustomCtor(ReflectionUtils.java:36)
    at org.apache.hadoop.hbase.ipc.RpcControllerFactory.instantiate(RpcControllerFactory.java:58)
    at org.apache.hadoop.hbase.client.ConnectionManager$HConnectionImplementation.createAsyncProcess(ConnectionManager.java:2317)
    at org.apache.hadoop.hbase.client.ConnectionManager$HConnectionImplementation.<init>(ConnectionManager.java:688)
    at org.apache.hadoop.hbase.client.ConnectionManager$HConnectionImplementation.<init>(ConnectionManager.java:630)
    ... 38 more
Caused by: java.lang.ClassNotFoundException: org.apache.hadoop.hbase.ipc.controller.ClientRpcControllerFactory
    at java.net.URLClassLoader.findClass(URLClassLoader.java:381)
    at java.lang.ClassLoader.loadClass(ClassLoader.java:424)
    at sun.misc.Launcher$AppClassLoader.loadClass(Launcher.java:331)
    at java.lang.ClassLoader.loadClass(ClassLoader.java:357)
    at java.lang.Class.forName0(Native Method)
    at java.lang.Class.forName(Class.java:264)
    at org.apache.hadoop.hbase.util.ReflectionUtils.instantiateWithCustomCtor(ReflectionUtils.java:32)
    ... 42 more

Some background information:

• The job runs on spark 1.4.1 (specified correct spark.yarn.jar field in the spark.conf file).
• oozie.libpath is set to the hdfs directory in which the jar of my program resides.
• org.apache.hadoop.hbase.ipc.controller.ClientRpcControllerFactory, the class not found, exists in phoenix-4.5.1-HBase-1.0-client.jar. I've specified this jar in spark.driver.extraClassPath and spark.executor.extraClassPath in my spark.conf file. I've also added the phoenix-core dependency in my pom file, so that the class exists in my shaded project jar as well.

Observations so far:

• adding an extra field in my spark.conf file spark.driver.userClassPathFirst and setting it to true gets rid of the classnotfound exception. However, it also prevents me from initializing a spark context (null pointer exception). From googling around it seems that including this field messes up classpaths, so may not be the way to go about it since I cannot even initialize a spark context this way.
• I noticed that in the oozie stdout log, I do not see the classpath of the phoenix jar. So maybe for some reason spark.driver.extraClassPath and spark.executor.extraClassPath aren't actually picking up the jar as an extraClassPath? I do know that I'm specifying the correct jar file path, as other jobs have spark.conf files with the same parameters.
• I found a hacky way to make the phoenix jar show up in the classpath (in the oozie stdout log) by copying the jar to the same directory as where my program jar resides. This works whether or not spark.executor.extraClassPath is changed to point to the new jar location. However, the classnotfound exception persists, even though I clearly see the ClientRpcControllerFactory jar when I unzip the jar)

Other things I've tried:

• I tried using the sparkConf.setJars() and sparkContext.addJar() methods, but still encountered the same error
• added the jar in the spark.driver.extraClassPath field in my job properties file, but it hasn't seemed to help (Spark docs indicated that this field is necessary when running in client mode, so may not be relevant for my case)

Any help/ideas/suggestions would be greatly appreciated.

Answer 1

I use CDH 5.5.1 + Phoenix 4.5.2 (both installed with parcels) and faced the same problem. I think the problem disappeared after I switched to client mode. I can't verify this because I am getting other error with cluster mode now.

I tried to trace Phoenix source code and found some interesting things. Hope Java / Scala expert identify the root cause.

1. The PhoenixDriver class was loaded. This showed the jar was found initially. After layers of Class Loader / context switch (?), the jar lost from the classpath.

2. If I Class.forName() a non-existing class in my program, there is no need to call sun.misc.Launcher$AppClassLoader.loadClass(Launcher.java:331) . The stack is like:

    java.lang.ClassNotFoundException: NONEXISTINGCLASS at java.net.URLClassLoader.findClass(URLClassLoader.java:381) at java.lang.ClassLoader.loadClass(ClassLoader.java:424) at java.lang.ClassLoader.loadClass(ClassLoader.java:357) at java.lang.Class.forName0(Native Method) at java.lang.Class.forName(Class.java:264) 

3. I copied Phoenix code into my program for testing. I still get the ClassNotFoundExcpetion if I call ConnectionQueryServicesImpl.init (ConnectionQueryServicesImpl.java:1896) . However, a call to ConnectionQueryServicesImpl.openConnection (ConnectionQueryServicesImpl.java:296) returned usable HBase connection. So it seems PhoenixContextExecutor was causing the loss of the jar, but I don't know how.

Source code of Cloudera Phoenix 4.5.2 : https://github.com/cloudera-labs/phoenix/blob/phoenix1-4.5.2_1.2.0/phoenix-core/src/main/java/org/apache/

(Not sure whether I should post a comment... but I have no reputation anyway)

Answer 2

So I managed to fix my issue and get my job to run. My solution is very hacky, but will post it here in case it may help others in the future.

Basically, the problem as I understand it was that the org.apache.hadoop.hbase.util.ReflectionUtils class, which is responsible for finding the ClientRpcControllerFactory class, was being loaded from some cloudera directory in the cluster instead of from my own jar. When I set spark.driver.userClassPathFirst to true, it prioritized loading the ReflectionUtils class from my jar, and so was able to location the ClientRpcControllerFactory class. But that messed up some other classpaths and kept giving me a NullPointerException when I tried to initialize a SparkContext, so I looked for another solution.

I tried to figure out if it was possible to exclude all default cdh jars from being included in my classpath, but found that the value in spark.yarn.jar was pulling in all these cdh jars, and I definitely needed to specify that jar.

So the solution was to include all classes under org.apache.hadoop.hbase from the Phoenix jar into spark-assembly jar (the jar that spark.yarn.jar pointed to), which got rid of the original exception and did not give me a NPE when trying to initialize a SparkContext. I found that now the ReflectionUtils class was being loaded from the spark-assembly jar, and since the ClientRpcControllerFactory was also included in that jar, it was able to find it. After this, I encountered a few more classNotFoundExceptions for Phoenix classes, so I put those classes into the spark-assembly jar as well.

Finally, I had a java.lang.RuntimeException: hbase-default.xml File Seems to be for and old Version of HBase problem. I found that my application jar contained such a file, but changing hbase.defaults.for.version.skip to true didn't do anything. So I included another hbase-default.xml file in the spark-assembly jar with the skip flag to true, and it finally worked.

Some observations:

• I noticed that my spark-assembly jar was completely missing an org.apache.hadoop.hbase directory. A coworker told me that usually I should expect to find an hbase directory in my spark-assembly jar, so maybe I was working with a bad spark-assembly jar. Edit: I checked a spark-assembly jar that I newly downloaded (v1.5.2) and it doesn't have it, so maybe the apache.hadoop.hbase package is not included in it.
• ClassNotFoundExceptions and classloader problems are hard to debug.