Jupyter Notebook Python, Scala, R, Spark, Mesos Stack

[Git Hub] https://github.com/jupyter/docker-stacks/tree/master/all-spark-notebook

What it Gives You

• Jupyter Notebook 4.2.x
• Conda Python 3.x and Python 2.7.x environments
• Conda R 3.2.x environment
• Scala 2.10.x
• pyspark, pandas, matplotlib, scipy, seaborn, scikit-learn pre-installed for Python
• ggplot2, rcurl preinstalled for R
• Spark 1.6.0 for use in local mode or to connect to a cluster of Spark workers
• Mesos client 0.22 binary that can communicate with a Mesos master
• Unprivileged user jovyan (uid=1000, configurable, see options) in group users (gid=100) with ownership over /home/jovyan and /opt/conda
• tini as the container entrypoint and start-notebook.sh as the default command
• A start-singleuser.sh script for use as an alternate command that runs a single-user instance of the Notebook server, as required by JupyterHub
• Options for HTTPS, password auth, and passwordless sudo

Basic Use

---

The following command starts a container with the Notebook server listening for HTTP connections on port 8888 without authentication configured.

docker run -d -p 8888:8888 jupyter/all-spark-notebook

Using Spark Local Mode

---

This configuration is nice for using Spark on small, local data.

In a Python Notebook

1. Run the container as shown above.
2. Open a Python 2 or 3 notebook.
3. Create a SparkContext configured for local mode.

For example, the first few cells in a notebook might read:

import pyspark
sc = pyspark.SparkContext('local[*]')

# do something to prove it works
rdd = sc.parallelize(range(1000))
rdd.takeSample(False, 5)

In a R Notebook

1. Run the container as shown above.
2. Open a R notebook.
3. Initialize sparkR for local mode.
4. Initialize sparkRSQL.

For example, the first few cells in a R notebook might read:

library(SparkR)

sc <- sparkR.init("local[*]")
sqlContext <- sparkRSQL.init(sc)

# do something to prove it works
data(iris)
df <- createDataFrame(sqlContext, iris)
head(filter(df, df$Petal_Width > 0.2))

In an Apache Toree (Scala) Notebook

1. Run the container as shown above.
2. Open an Apache Toree (Scala) notebook.
3. Use the pre-configured SparkContext in variable sc.

For example:

val rdd = sc.parallelize(0 to 999)
rdd.takeSample(false, 5)

Connecting to a Spark Cluster on Mesos

---

This configuration allows your compute cluster to scale with your data.

1. Deploy Spark on Mesos.
2. Configure each slave with the --no-switch_user flag or create the jovyan user on every slave node.
3. 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.
4. Follow the language specific instructions below.

 In a Python Notebook

1. Open a Python 2 or 3 notebook.
2. Create a SparkConf instance in a new notebook pointing to your Mesos master node (or Zookeeper instance) and Spark binary package location.
3. Create a SparkContext using this configuration.

For example, the first few cells in a Python 3 notebook might read:

import os
# make sure pyspark tells workers to use python3 not 2 if both are installed
os.environ['PYSPARK_PYTHON'] = '/usr/bin/python3'

import pyspark
conf = pyspark.SparkConf()

# point to mesos master or zookeeper entry (e.g., zk://10.10.10.10:2181/mesos)
conf.setMaster("mesos://10.10.10.10:5050")
# point to spark binary package in HDFS or on local filesystem on all slave
# nodes (e.g., file:///opt/spark/spark-1.6.0-bin-hadoop2.6.tgz)
conf.set("spark.executor.uri", "hdfs://10.10.10.10/spark/spark-1.6.0-bin-hadoop2.6.tgz")
# set other options as desired
conf.set("spark.executor.memory", "8g")
conf.set("spark.core.connection.ack.wait.timeout", "1200")

# create the context
sc = pyspark.SparkContext(conf=conf)

# do something to prove it works
rdd = sc.parallelize(range(100000000))
rdd.sumApprox(3)

To use Python 2 in the notebook and on the workers, change the PYSPARK_PYTHON environment variable to point to the location of the Python 2.x interpreter binary. If you leave this environment variable unset, it defaults to python.

Of course, all of this can be hidden in an IPython kernel startup script, but "explicit is better than implicit." :)

In a R Notebook

1. Run the container as shown above.
2. Open a R notebook.
3. Initialize sparkR Mesos master node (or Zookeeper instance) and Spark binary package location.
4. Initialize sparkRSQL.

For example, the first few cells in a R notebook might read:

library(SparkR)

# point to mesos master or zookeeper entry (e.g., zk://10.10.10.10:2181/mesos)\
# as the first argument
# point to spark binary package in HDFS or on local filesystem on all slave
# nodes (e.g., file:///opt/spark/spark-1.6.0-bin-hadoop2.6.tgz) in sparkEnvir
# set other options in sparkEnvir
sc <- sparkR.init("mesos://10.10.10.10:5050", sparkEnvir=list(
    spark.executor.uri="hdfs://10.10.10.10/spark/spark-1.6.0-bin-hadoop2.6.tgz",
    spark.executor.memory="8g"
    )
)
sqlContext <- sparkRSQL.init(sc)

# do something to prove it works
data(iris)
df <- createDataFrame(sqlContext, iris)
head(filter(df, df$Petal_Width > 0.2))

In an Apache Toree (Scala) Notebook

1. Open a terminal via New -> Terminal in the notebook interface.
2. Add information about your cluster to the SPARK_OPTS environment variable when running the container.
3. Open an Apache Toree (Scala) notebook.
4. Use the pre-configured SparkContext in variable sc.

The Apache Toree kernel automatically creates a SparkContext when it starts based on configuration information from its command line arguments and environment variables. You can pass information about your Mesos cluster via the SPARK_OPTS environment variable when you spawn a container.

For instance, to pass information about a Mesos master, Spark binary location in HDFS, and an executor options, you could start the container like so:

docker run -d -p 8888:8888 -e SPARK_OPTS '--master=mesos://10.10.10.10:5050 \ --spark.executor.uri=hdfs://10.10.10.10/spark/spark-1.6.0-bin-hadoop2.6.tgz \ --spark.executor.memory=8g' jupyter/all-spark-notebook

Note that this is the same information expressed in a notebook in the Python case above. Once the kernel spec has your cluster information, you can test your cluster in an Apache Toree notebook like so:

// should print the value of --master in the kernel spec
println(sc.master)

// do something to prove it works
val rdd = sc.parallelize(0 to 99999999)
rdd.sum()

 Connecting to a Spark Cluster on Standalone Mode

---

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

1. Verify that the docker image (check the Dockerfile) and the Spark Cluster which is being deployed, run the same version of Spark.
2. Deploy Spark on Standalone Mode.
3. 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.)
4. • 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.
5. The language specific instructions are almost same as mentioned above for Mesos, only the master url would now be something like spark://10.10.10.10:7077

Notebook Options

You can pass Jupyter command line options through the start-notebook.sh command when launching the container. For example, to set the base URL of the notebook server you might do the following:

docker run -d -p 8888:8888 jupyter/all-spark-notebook start-notebook.sh --NotebookApp.base_url=/some/path

You can sidestep the start-notebook.sh script entirely by specifying a command other than start-notebook.sh. If you do, the NB_UID and GRANT_SUDO features documented below will not work. See the Docker Options section for details.

Docker Options

You may customize the execution of the Docker container and the Notebook server it contains with the following optional arguments.

• -e PASSWORD="YOURPASS" - Configures Jupyter Notebook to require the given password. Should be conbined with USE_HTTPS on untrusted networks.
• -e USE_HTTPS=yes - Configures Jupyter Notebook to accept encrypted HTTPS connections. If a pem file containing a SSL certificate and key is not provided (see below), the container will generate a self-signed certificate for you.
• -e NB_UID=1000 - Specify the uid of the jovyan user. Useful to mount host volumes with specific file ownership. For this option to take effect, you must run the container with --user root. (The start-notebook.sh script will su jovyan after adjusting the user id.)
• -e GRANT_SUDO=yes - Gives the jovyan user passwordless sudo capability. Useful for installing OS packages. For this option to take effect, you must run the container with --user root. (The start-notebook.sh script will su jovyan after adding jovyan to sudoers.) You should only enable sudo if you trust the user or if the container is running on an isolated host.
• -v /some/host/folder/for/work:/home/jovyan/work - Host mounts the default working directory on the host to preserve work even when the container is destroyed and recreated (e.g., during an upgrade).
• -v /some/host/folder/for/server.pem:/home/jovyan/.local/share/jupyter/notebook.pem - Mounts a SSL certificate plus key for USE_HTTPS. Useful if you have a real certificate for the domain under which you are running the Notebook server.
• -p 4040:4040 - Opens the port for the Spark Monitoring and Instrumentation UI. 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. docker run -d -p 8888:8888 -p 4040:4040 -p 4041:4041 jupyter/all-spark-notebook

SSL Certificates

The notebook server configuration in this Docker image expects the notebook.pem file mentioned above to contain a base64 encoded SSL key and at least one base64 encoded SSL certificate. The file may contain additional certificates (e.g., intermediate and root certificates).

If you have your key and certificate(s) as separate files, you must concatenate them together into the single expected PEM file. Alternatively, you can build your own configuration and Docker image in which you pass the key and certificate separately.

For additional information about using SSL, see the following:

• The docker-stacks/examples for information about how to use Let's Encrypt certificates when you run these stacks on a publicly visible domain.
• The jupyter_notebook_config.py file for how this Docker image generates a self-signed certificate.

The Jupyter Notebook documentation for best practices about running a public notebook server in general, most of which are encoded in this image.

 

Conda Environments

The default Python 3.x Conda environment resides in /opt/conda. A second Python 2.x Conda environment exists in /opt/conda/envs/python2. You can switch to the python2 environment in a shell by entering the following:

source activate python2

You can return to the default environment with this command:

source deactivate

The commands jupyter, ipython, python, pip, easy_install, and conda (among others) are available in both environments. For convenience, you can install packages into either environment regardless of what environment is currently active using commands like the following:

# install a package into the python2 environment
pip2 install some-package

conda install -n python2 some-package

# install a package into the default (python 3.x) environment
pip3 install some-packageconda 

install -n python3 some-package

JupyterHub

---

JupyterHub requires a single-user instance of the Jupyter Notebook server per user. To use this stack with JupyterHub and DockerSpawner, you must specify the container image name and override the default container run command in your jupyterhub_config.py:

# Spawn user containers from this image
c.DockerSpawner.container_image = 'jupyter/all-spark-notebook'

# Have the Spawner override the Docker run command
c.DockerSpawner.extra_create_kwargs.update({
    'command': '/usr/local/bin/start-singleuser.sh'
})

Docker Hub 이미지를 이용한 Spark 설치하기

Apacher Spark 이미지

Docker Hub 이미지 중 가장많은 별포인트를 받은 아래의 Spark 이미지를 설치한다.

[Docker Hub] https://hub.docker.com/r/jupyter/all-spark-notebook/

[Git Hub] https://github.com/jupyter/docker-stacks

아래 명령어로 docker hub에서 이미지를 가져온다.

위의 명령어를 실행하면 아래와 같이 이미지를 다운받고 extract 하는 과정을 볼 수가 있다.

Image 다운이 다 되었으면, 아래의 명령어를 통해 다운 받은 Docker Image를 확인할 수가 있다.

제대로 설치가 되었다면 아래와 유사한 목록 리스트가 나타나야 한다.

컨테이너를 만들고 실행시키기 위해서 아래의 명령어를 이용한다.

docker ps 명령을 통해 해당 컨테이너가 잘 실행되었는지를 확인하자.

방금 만든 컨테이너가 보인다면, 해당 서비스가 잘 구동되어 있음을 나타내는 것이다.

이제 웹상에서 해당 machine의 ip와 port(ex) 8888)를 친후 ipython-notebook을 실행해보자.

[Docke Hub] https://hub.docker.com/r/sequenceiq/spark/

Apache Spark on Docker

This repository contains a Docker file to build a Docker image with Apache Spark. This Docker image depends on our previous Hadoop Docker

image, available at the SequenceIQ GitHub page.
The base Hadoop Docker image is also available as an official Docker image.

##Pull the image from Docker Repository

docker pull sequenceiq/spark:1.6.0

Building the image

docker build --rm -t sequenceiq/spark:1.6.0 .

Running the image

• if using boot2docker make sure your VM has more than 2GB memory
• in your /etc/hosts file add $(boot2docker ip) as host 'sandbox' to make it easier to access your sandbox UI
• open yarn UI ports when running container

  docker run -it -p 8088:8088 -p 8042:8042 -h sandbox sequenceiq/spark:1.6.0 bash
  

  or

  docker run -d -h sandbox sequenceiq/spark:1.6.0 -d

Versions

Hadoop 2.6.0 and Apache Spark v1.6.0 on Centos

Testing

There are two deploy modes that can be used to launch Spark applications on YARN.

YARN-client mode

In yarn-client mode, the driver runs in the client process, and the application master is only used for requesting resources from YARN.

# run the spark shell
spark-shell \
--master yarn-client \
--driver-memory 1g \
--executor-memory 1g \
--executor-cores 1

# execute the the following command which should return 1000
scala> sc.parallelize(1 to 1000).count()

YARN-cluster mode

In yarn-cluster mode, the Spark driver runs inside an application master process which is managed by YARN on the cluster, and the client can go away after initiating the application.

Estimating Pi (yarn-cluster mode):

# execute the the following command which should write the "Pi is roughly 3.1418" into the logs
# note you must specify --files argument in cluster mode to enable metrics
spark-submit \
--class org.apache.spark.examples.SparkPi \
--files $SPARK_HOME/conf/metrics.properties \
--master yarn-cluster \
--driver-memory 1g \
--executor-memory 1g \
--executor-cores 1 \
$SPARK_HOME/lib/spark-examples-1.6.0-hadoop2.6.0.jar

Estimating Pi (yarn-client mode):

# execute the the following command which should print the "Pi is roughly 3.1418" to the screen
spark-submit \
--class org.apache.spark.examples.SparkPi \
--master yarn-client \
--driver-memory 1g \
--executor-memory 1g \
--executor-cores 1 \
$SPARK_HOME/lib/spark-examples-1.6.0-hadoop2.6.0.jar