Major Description
The Popen call to launch the py4j gateway specifies no stdout and stderr options, meaning logging from the JVM always goes to the parent process terminal.
https://github.com/apache/spark/blob/v2.1.1/python/pyspark/java_gateway.py#L77
It would be super handy if the launch_gateway function took an additional dict parameter called popen_kwargs and passed it along to the Popen calls. This API enhancement, for example, would allow Python applications to capture all stdout and stderr coming from Spark and process it programmatically, without resorting to reading from log files or other hijinks.
Example use:
import pyspark import subprocess from pyspark.java_gateway import launch_gateway # Make the py4j JVM stdout and stderr available without buffering popen_kwargs = { 'stdout': subprocess.PIPE, 'stderr': subprocess.PIPE, 'bufsiz': 0 } # Launch the gateway with our custom settings gateway = launch_gateway(popen_kwargs=popen_kwargs) # Use the gateway we launched sc = pyspark.SparkContext(gateway=gateway) # This could be done in a thread or event loop or ... # Written briefly / poorly here only as a demo while True: buf = gateway.proc.stdout.read() print(buf.decode('utf-8'))
To get access to the stdout and stderr pipes, the "proc" instance created in launch_gateway also needs to be exposed to the application. I'm thinking that stashing it on the JavaGateway instance that the function already returns is the cleanest from the client perspective, but means hanging an extra attribute off the py4j.JavaGateway object.
I can submit a PR with this addition for further discussion.
