docker-stacks/docs/using/specifics.md

Image Specifics

This page provides details about features specific to one or more images.

Apache Spark™

Specific Docker Image Options

• -p 4040:4040 - The jupyter/pyspark-notebook and jupyter/all-spark-notebook images open SparkUI (Spark Monitoring and Instrumentation UI) at default port 4040, this option map 4040 port inside docker container to 4040 port on host machine . Note every new spark context that is created is put onto an incrementing port (ie. 4040, 4041, 4042, etc.), and it might be necessary to open multiple ports. For example: docker run -d -p 8888:8888 -p 4040:4040 -p 4041:4041 jupyter/pyspark-notebook.

Build an Image with a Different Version of Spark

You can build a pyspark-notebook image (and also the downstream all-spark-notebook image) with a different version of Spark by overriding the default value of the following arguments at build time.

• Spark distribution is defined by the combination of the Spark and the Hadoop version and verified by the package checksum, see Download Apache Spark for more information.
  • spark_version: The Spark version to install (3.0.0).
  • hadoop_version: The Hadoop version (3.2).
  • spark_checksum: The package checksum (BFE4540...).
• Spark is shipped with a version of Py4J that has to be referenced in the PYTHONPATH.
  • py4j_version: The Py4J version (0.10.9), see the tip below.
• Spark can run with different OpenJDK versions.
  • openjdk_version: The version of (JRE headless) the OpenJDK distribution (11), see Ubuntu packages.

For example here is how to build a pyspark-notebook image with Spark 2.4.6, Hadoop 2.7 and OpenJDK 8.

# From the root of the project
# Build the image with different arguments
docker build --rm --force-rm \
    -t jupyter/pyspark-notebook:spark-2.4.6 ./pyspark-notebook \
    --build-arg spark_version=2.4.6 \
    --build-arg hadoop_version=2.7 \
    --build-arg spark_checksum=3A9F401EDA9B5749CDAFD246B1D14219229C26387017791C345A23A65782FB8B25A302BF4AC1ED7C16A1FE83108E94E55DAD9639A51C751D81C8C0534A4A9641 \
    --build-arg openjdk_version=8 \
    --build-arg py4j_version=0.10.7

# Check the newly built image
docker images jupyter/pyspark-notebook:spark-2.4.6

# REPOSITORY                 TAG                 IMAGE ID            CREATED             SIZE
# jupyter/pyspark-notebook   spark-2.4.6         7ad7b5a9dbcd        4 minutes ago       3.44GB

# Check the Spark version
docker run -it --rm jupyter/pyspark-notebook:spark-2.4.6 pyspark --version

# Welcome to
#       ____              __
#      / __/__  ___ _____/ /__
#     _\ \/ _ \/ _ `/ __/  '_/
#    /___/ .__/\_,_/_/ /_/\_\   version 2.4.6
#       /_/
#                         
# Using Scala version 2.11.12, OpenJDK 64-Bit Server VM, 1.8.0_265

Tip: to get the version of Py4J shipped with Spark:

• Build a first image without changing py4j_version (it will not prevent the image to build it will just prevent Python to find the pyspark module),
• get the version (ls /usr/local/spark/python/lib/),
• set the version --build-arg py4j_version=0.10.7.

Note: At the time of writing there is an issue preventing to use Spark 2.4.6 with Python 3.8, see this answer on SO for more information.

docker run -it --rm jupyter/pyspark-notebook:spark-2.4.6 ls /usr/local/spark/python/lib/ 
# py4j-0.10.7-src.zip  PY4J_LICENSE.txt  pyspark.zip
# You can now set the build-arg
# --build-arg py4j_version=

Usage Examples

The jupyter/pyspark-notebook and jupyter/all-spark-notebook images support the use of Apache Spark in Python, R, and Scala notebooks. The following sections provide some examples of how to get started using them.

Using Spark Local Mode

Spark local mode is useful for experimentation on small data when you do not have a Spark cluster available.

In Python

In a Python notebook.

from pyspark.sql import SparkSession

# Spark session & context
spark = SparkSession.builder.master('local').getOrCreate()
sc = spark.sparkContext

# Sum of the first 100 whole numbers
rdd = sc.parallelize(range(100 + 1))
rdd.sum()
# 5050

In R

In a R notebook with SparkR.

library(SparkR)

# Spark session & context
sc <- sparkR.session("local")

# Sum of the first 100 whole numbers
sdf <- createDataFrame(list(1:100))
dapplyCollect(sdf,
              function(x) 
              { x <- sum(x)}
             )
# 5050

In a R notebook with sparklyr.

library(sparklyr)

# Spark configuration
conf <- spark_config()
# Set the catalog implementation in-memory
conf$spark.sql.catalogImplementation <- "in-memory"

# Spark session & context
sc <- spark_connect(master = "local", config = conf)

# Sum of the first 100 whole numbers
sdf_len(sc, 100, repartition = 1) %>% 
    spark_apply(function(e) sum(e))
# 5050

In Scala

Spylon kernel instantiates a SparkContext for you in variable sc after you configure Spark options in a %%init_spark magic cell.

%%init_spark
# Configure Spark to use a local master
launcher.master = "local"

// Sum of the first 100 whole numbers
val rdd = sc.parallelize(0 to 100)
rdd.sum()
// 5050

Connecting to a Spark Cluster in Standalone Mode

Connection to Spark Cluster on Standalone Mode requires the following set of steps:

0. Verify that the docker image (check the Dockerfile) and the Spark Cluster which is being deployed, run the same version of Spark.
1. Deploy Spark in Standalone Mode.
2. Run the Docker container with --net=host in a location that is network addressable by all of your Spark workers. (This is a Spark networking requirement.)
   • NOTE: When using --net=host, you must also use the flags --pid=host -e TINI_SUBREAPER=true. See https://github.com/jupyter/docker-stacks/issues/64 for details.

Note: In the following examples we are using the Spark master URL spark://master:7077 that shall be replaced by the URL of the Spark master.

In Python

The same Python version need to be used on the notebook (where the driver is located) and on the Spark workers. The python version used at driver and worker side can be adjusted by setting the environment variables PYSPARK_PYTHON and / or PYSPARK_DRIVER_PYTHON, see Spark Configuration for more information.

from pyspark.sql import SparkSession

# Spark session & context
spark = SparkSession.builder.master('spark://master:7077').getOrCreate()
sc = spark.sparkContext

# Sum of the first 100 whole numbers
rdd = sc.parallelize(range(100 + 1))
rdd.sum()
# 5050

In R

In a R notebook with SparkR.

library(SparkR)

# Spark session & context
sc <- sparkR.session("spark://master:7077")

# Sum of the first 100 whole numbers
sdf <- createDataFrame(list(1:100))
dapplyCollect(sdf,
              function(x) 
              { x <- sum(x)}
             )
# 5050

In a R notebook with sparklyr.

library(sparklyr)

# Spark session & context
# Spark configuration
conf <- spark_config()
# Set the catalog implementation in-memory
conf$spark.sql.catalogImplementation <- "in-memory"
sc <- spark_connect(master = "spark://master:7077", config = conf)

# Sum of the first 100 whole numbers
sdf_len(sc, 100, repartition = 1) %>% 
    spark_apply(function(e) sum(e))
# 5050

In Scala

Spylon kernel instantiates a SparkContext for you in variable sc after you configure Spark options in a %%init_spark magic cell.

%%init_spark
# Configure Spark to use a local master
launcher.master = "spark://master:7077"

// Sum of the first 100 whole numbers
val rdd = sc.parallelize(0 to 100)
rdd.sum()
// 5050

Tensorflow

The jupyter/tensorflow-notebook image supports the use of Tensorflow in single machine or distributed mode.

Single Machine Mode

import tensorflow as tf

hello = tf.Variable('Hello World!')

sess = tf.Session()
init = tf.global_variables_initializer()

sess.run(init)
sess.run(hello)

Distributed Mode

import tensorflow as tf

hello = tf.Variable('Hello Distributed World!')

server = tf.train.Server.create_local_server()
sess = tf.Session(server.target)
init = tf.global_variables_initializer()

sess.run(init)
sess.run(hello)