Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Information Provided
-
1.2.1, 1.4.1, 1.6.0, 1.6.1
-
None
-
None
Description
[Edit]
Editing the description because I can't comment on JIRA since yesterday, because of the anti-spam measure. Well, we got down to the problem, and it is not a Spark issue. tldr; the type of problem you would want to use containers to avoid, and we are going to be using Mesos moving forward. Explaining the details is going to be tough because there were too many moving parts. There is only one learning I can share about Spark: when you build spark with -Phive it makes a copy of the Hadoop configurations into the Spark configurations folder, however those configurations are overridden by the configurations inherited through HADOOP_CONF_DIR. I might have gotten this the other way around, but make sure you use one or the other and do it consistently across your cluster.
I am going to close the issue as Information Given, because Sean has been very generous with his time and his comments helped us pin point the root cause of the problem. Thank you very much Sean.
[End Edit]
We have started to frequently see Spark apps failing because of a NoClassDefFoundError thrown despite that the dependency had been added to the ClassLoader just before it was thrown. The Executor.run method adds the JAR containing the class but then a NoClassDefFoundError is thrown subsequently. We see log messages from updateDependencies indicating that the JAR is fetched and added to the class loader. Upon inspection of the worker dir, the JAR is there, it is not corrupted, and it contains the class that could not be found in the class loader.
We first saw this when we started writing streaming apps, and we thought it is something specific to streaming apps. However, this was wrong as the same problem happened with several batch apps.
We first saw this on a Standalone cluster, and we though that it might be a problem caused by the lack of resource manager. Now we installed Mesos and the problem still happens.
I tried to create a POC Spark App that demonstrates the problem, but I couldn't reliably reproduce it. The problem would still happen in other apps, but it didn't happen in the POC app even though I made it structurally the same as any other app we run. The problem seems to be environmental, specially because we found a work around for it.
The work around we found is setting SPARK_CLASSPATH on the executor nodes to a local copy of the dependency. The problem still happens if we set the 'spark.executor.extraClassPath' or 'spark.driver.extraClassPath' or set SPARK_CLASSPATH on the driver node. However, if the SPARK_CLASSPATH is set on the executor node, then the problem doesn't happen because the JAR doesn't need to be added to the class loader by Executor#updateDependencies.
Other symptoms of the problem are the following:
1) Even though there is a 'log4j.properties' in the 'spark.executor.extraClassPath', the first line of the stderr of the worker says "Using Spark's default log4j profile: org/apache/spark/log4j-defaults.properties" The log4j.properties file that is shipped with the job is totally neglected.
2) Any configuration files on 'spark.executor.extraClassPath' are neglected. I am mentioning this because log4j.properties is loaded very early on and in a static call, which might sway the troubleshooting into wrong directions.
Here is the specific example in our case:
> grep NoClassDef workers/app-20160414111328-0043/0/stderr
Caused by: java.lang.NoClassDefFoundError: Could not initialize class org.apache.hadoop.hbase.protobuf.ProtobufUtil
Caused by: java.lang.NoClassDefFoundError: Could not initialize class org.apache.hadoop.hbase.protobuf.ProtobufUtil
.. SEVERAL ATTEMPTS ...
Caused by: java.lang.NoClassDefFoundError: Could not initialize class org.apache.hadoop.hbase.protobuf.ProtobufUtil
Even though, in the same application worker dir:
> for j in workers/app-20160414111328-0043/0/*.jar ; do jar tf $j | grep ProtobufUtil ; done;
org/apache/hadoop/hbase/protobuf/ProtobufUtil$1.class
org/apache/hadoop/hbase/protobuf/ProtobufUtil.class
There are other examples, specially for configurations not being found. I think the SPARK-12279 can also be caused by the same root cause.
We have been seeing this in several of our clusters and several engineers had spent days looking into why their applications suffer from this. We rebuilt our infrastructure (always on AWS EC2 nodes) and tested many hypotheses, including things that are non-sensical, and we still can't find anything that reliably reproduces the problem. The only reliable piece of information is that setting SPARK_CLASSPATH on the executor nodes prevents the problem from happening, because then the dependencies are included in the -cp parameter of the java command running the CoarseGrainedExecutorBackend .
We would appreciate if someone more knowledgeable in Spark internals take a look, and we can help by providing as much details as possible.
Attachments
Issue Links
- incorporates
-
-
- Resolved
-