Hi All, We had one of the requirement of supporting zeppelin deployment over Kubernetes cluster and running spark jobs over zeppelin through spark interpreter.
* As a first step, we have cloned the zeppelin source code having last commit id: 685eb9249d1c1d821ce57f1ed0559f1539dfbe69 * While building source code with the above commit id we faced following issues: 1. after cloning the source to particular directory, there are some modifications needed. Otherwise user will get the errors related to 'npm' is not installed properly and the directory where actually the zeppelin code is cloned the folder should be three level inner folder of root folder of the machine where actually it has been cloned. Otherwise internally code will be not able to create plugin folder after three level up from the cloned folder. <Refer the pom.xml zeppelin-distribution of zeppelin source code> 2. In our environment proxy urls should be added to .bowerrc file of the zeppelin-web project of the zeppelin source code. This can be updated in the build user guide. 3. All profiles are not correctly supported by the zeppelin code. Some specific profiles needs to be added during the build time. < -Pyarn -Ppyspark >So the following maven goal needs to be executed from zeppelin parent directory to perform zeppelin code build. 4. maven command to execute from the zeppelin parent directory. mvn clean install -DskipTests -Drat.skip=true -Pspark-2.2 -Phadoop2 -Pscala-2.11 -s settings.xml 5. after executing the above mvn command build was successful on all 49 sub module packages of zeppelin. 6. Then cd into zeppelin-distribution and execute following mvn goal. mvn org.apache.maven.plugins:maven-assembly-plugin:3.0.0:single -P apache-release. * After building the source code we took https://github.com/apache/zeppelin/pull/2637 this PR ,added to the code and built the source code. * And when the zeppelin is deployed to kubernates cluster, and spark job is being submitted, drivers and executors came up but were referring to the spark interpreter class SparkInterpreterLauncher.java and not SparkK8SInterpreterLauncher.java of above mentioned PR * After reviewing the PR, we found that there is one condition in 'interpretersetting.java' that checks if deploy mode is cluster and then only launches SparkK8SInterpreterLauncher.java otherwise launches SparkInterpreterLauncher.java code. * Setting the value from the Zeppelin GUI did not work. * So we modified interpreter-setting.json file of zeppelin-0.9.0-SNAPSHOT/interpreter/spark to include spark.submit.deployMode as cluster * After these changes, when the driver is spawned it gives the following exception: ClassNotFoundException:org.apache.zeppelin.interpreter.remote.RemoteInterpreterServer as it is using spark-interpreter-0.9.0-SNAPSHOT.jar during spark submit job, * To resolve the above issue, interpreter.sh was modified to execute the spark submit command using zeppelin-interpreter-0.9.0-SNAPSHOT.jar * when above mentioned jar is used to spawn spark submit command one more exception was encountered * ClassNotFoundException cerner.ether class not found * to resolve ClassNotFoundException cerner.ether class error spark-interpreter-0.9.0-SNAPSHOT.jar needs to be mentioned in classpath using --jars option during spark submit. * After integrating all above changes, zeppelin submit of spark jobs started using SparkK8SInterpreterLauncher.java. Now our proposal is after doing all above modifications should it go for upstream for merging with the zeppelin source code or not. Please let us know if the above procedure we followed is proper way or not. One more question when will the above mentioned PR is planned for merging with the source code. Some changes may be needed for the K8_URL handling. We can help with community developers so that these changes can be integrated to the zeppelin source code. Regards Naveen
