Can I run a pyspark jupyter notebook in cluster deploy mode?

Question

Context: The cluster configuration is as follows:

• Everything is running withing docker files.
• node1: spark master
• node2: jupyter hub (where I run my notebooks as well)
• node3-7: spark worker nodes
• I can telnet and ping from my worker nodes to node2 and vice versa using the default ports for spark

Problem: I'm trying to create a spark session in pyspark jupyter notebook which runs in cluster deploy mode. I'm trying to get the driver to run on a node that isn't the node where the jupyter notebook is run from. Right now I can run jobs on the cluster but only with the driver running on node2.

After a lot of digging I found this stackoverflow post which claims that if you run an interactive shell with spark you can only do so in local deploy mode (where the driver is located on the machine you are working on). That post goes on to say that something like jupyter hub as result also won't work in cluster deploy mode but I am unable to find any documentation that can confirm this. Can someone confirm if jupyter hub can run in cluster mode at all?

My attempt at running the spark session in cluster deployed mode:

from pyspark.sql import SparkSession

spark = SparkSession.builder\
.enableHiveSupport()\
.config("spark.local.ip",<node 3 ip>)\
.config("spark.driver.host",<node 3 ip>)\
.config('spark.submit.deployMode','cluster')\
.getOrCreate()

Error:

/usr/spark/python/pyspark/sql/session.py in getOrCreate(self)
    167                     for key, value in self._options.items():
    168                         sparkConf.set(key, value)
--> 169                     sc = SparkContext.getOrCreate(sparkConf)
    170                     # This SparkContext may be an existing one.
    171                     for key, value in self._options.items():

/usr/spark/python/pyspark/context.py in getOrCreate(cls, conf)
    308         with SparkContext._lock:
    309             if SparkContext._active_spark_context is None:
--> 310                 SparkContext(conf=conf or SparkConf())
    311             return SparkContext._active_spark_context
    312 

/usr/spark/python/pyspark/context.py in __init__(self, master, appName, sparkHome, pyFiles, environment, batchSize, serializer, conf, gateway, jsc, profiler_cls)
    113         """
    114         self._callsite = first_spark_call() or CallSite(None, None, None)
--> 115         SparkContext._ensure_initialized(self, gateway=gateway, conf=conf)
    116         try:
    117             self._do_init(master, appName, sparkHome, pyFiles, environment, batchSize, serializer,

/usr/spark/python/pyspark/context.py in _ensure_initialized(cls, instance, gateway, conf)
    257         with SparkContext._lock:
    258             if not SparkContext._gateway:
--> 259                 SparkContext._gateway = gateway or launch_gateway(conf)
    260                 SparkContext._jvm = SparkContext._gateway.jvm
    261 

/usr/spark/python/pyspark/java_gateway.py in launch_gateway(conf)
     93                 callback_socket.close()
     94         if gateway_port is None:
---> 95             raise Exception("Java gateway process exited before     sending the driver its port number")
     96 
     97         # In Windows, ensure the Java child processes do not linger after Python has exited.

Exception: Java gateway process exited before sending the driver its port number

Answer 1

1. One "supported" way to indirectly use yarn-cluster mode in Jupyter is through Apache Livy

Basically, Livy is a REST API service for Spark cluster.

Jupyter has a extension "spark-magic" that allows to integrate Livy with Jupyter

An example of Jupyter with Spark-magic bound (driver runs in the yarn cluster and not locally in this case, as mentioned above):

2. Another way to use YARN-cluster mode in Jupyter is to use Jupyter Enterprise Gateway https://jupyter-enterprise-gateway.readthedocs.io/en/latest/kernel-yarn-cluster-mode.html#configuring-kernels-for-yarn-cluster-mode

3. There are also commercial options that normally use one of the methods I listed above. Some of our users for example are on IBM DSX (aka IBM Watson Studio Local) use Apache Livy - option number one above.

Answer 2

You cannot use cluster mode with PySpark at all:

Currently, standalone mode does not support cluster mode for Python applications.

and even if you could cluster mode is not applicable in interactive environment:

case (_, CLUSTER) if isShell(args.primaryResource) =>
  error("Cluster deploy mode is not applicable to Spark shells.")
case (_, CLUSTER) if isSqlShell(args.mainClass) =>
  error("Cluster deploy mode is not applicable to Spark SQL shell.")

Answer 3

I'm not an expert in PySpark, but did you try to change the kernel.json file of pyspark jupyter kernel ?

Maybe you can add the option deploy mode cluster in it

"env": {
  "SPARK_HOME": "/your_dir/spark",
  "PYTHONPATH": "/your_dir/spark/python:/your_dir/spark/python/lib/py4j-0.9-src.zip",
  "PYTHONSTARTUP": "/your_dir/spark/python/pyspark/shell.py",
  "PYSPARK_SUBMIT_ARGS": "--master local[*] pyspark-shell"
 }

And you change this line :

"PYSPARK_SUBMIT_ARGS": "--master local[*] pyspark-shell"

with your cluster master ip and --deploy-mode cluster

Not sure that will change anything, but maybe that works, and I would love to know too !

Good luck

EDIT : I found this maybe that can help yuo even though it's from 2015

link jupyter pyspark cluster