需求

主程序拆成多个子模块方便复用：util.py, module1.py, module2.py, main.py。

Solution

对于 main.py 依赖的 util.py, module1.py, module2.py，需要先压缩成一个 .zip 文件，再通过 spark-submit 的 --py--files 选项上传到 yarn，mail.py 才能 import 这些子模块。命令如下：

$ spark-submit --master=yarn --deploy-mode=cluster --jars elasticsearch-hadoop-5.3.1.jar --py-files deps.zip main.py

Oozie spark-action

需要在 <spark-opts> 后面加上相应的 ** --py-files ** 选项:

<spark-opts>
  ${OTHER_OPTS} --py-files hdfs://${HDFS_HOST}:${HDFS_PORT}/${DEP_PATH}/deps.zip
</spark-opts>

软件版本

Spark 1.6.2，Oozie 4.2 测试 OK.

参考链接

Third party packages: If you require a third party package to be installed on the executors, do so when setting up the node (for example with pip).
Scripts: If you require custom scripts, create a *.zip (or *.egg) file containing all of your dependencies and ship it to the executors using--py-filescommand line option. Two things to look out for:
Make sure that you also include ‘nested dependencies’
Make sure that your *.py files are at the top level of the *.zip file.