for a brief introduction to streams
spark can be run in many ways. I did on AWS EMR, Dataproc on GCP, Computing Instance on GCP and Serverless Databricks.
- standalone, on GCP
here, you should do some steps on GCP (step-1,2,3,4,5) and on your local machine (step-6)
also, open ports: 7070, 4040 on GCP instance firewall - standalone, on your local machine
- standalone, on your local machine w/ anaconda and pyspark package installation (step-6)
- master-slave, on GCP
here, you should do some steps on GCP (step-1,2,3,4,5,8) and on your local machine (step-6)
also, open ports: 7070, 8080, 8081 on GCP instance firewall - master-slave, on AWS EMR
look for details under /aws-emr-jupiter-notebooks/README.md file. You have to pay, either use or not use, but 2 optimization is available.
* EMR instances are %50 discounted compared to EC2 equivalant instances
* Leverage spot instances for more cost decrease.
* Use transient instances for job kind things - master-slave, on GCP Dataproc
very similar to AWS, except more robust and more faster and much better experience in GCP.
* Do not forget to add Jupiter component installation and open 8123 port on firewall !
* Leverage preemptible instances for more cost decrease.
* Use transient instances for job kind things - master-slave, on your local machine
- Serverless Databricks on Azure
look for details under /databricks-jupiter-notebooks/README.md file. You only pay,when you process. - Serverless Databricks on AWS
Basicly,
- copy your dataset
- get your spark cluster
- do your things in jupiter
- submit your job to cluster
0- create a GCP (ubuntu 18.04) instance on GCP console, then connect with that server via appropriate SSH ways.
gcloud compute --project .... ssh --zone .... ....
1- download tansudasli/spark-sandbox files to GCP instance via
git clone https://github.com/tansudasli/spark-sandbox.git
, then
cd spark-sandbox
2- then give run permisson to install_cloudera_stack.sh file
chmod +x install_apache_spark.sh
3- and run below script to install python3x, java8 and spark 2.4
./install_apache_spark.sh
if you face w/ connection or downloading issues, run it again after delete unnecessary folders.
4- test pyspark in standalone, on GCP
pyspark
for python or spark-shell
for scala
and then, in the shell, type sc.version
* at this stage you may access your spark over `IP:4040` to see jobs and storages etc.
5- download movielens sample data set.
sudo apt install unzip
wget http://files.grouplens.org/datasets/movielens/ml-100k.zip
unzip ml-100k.zip
for latest & largest data-set, you may use http://files.grouplens.org/datasets/movielens/ml-latest.zip url.
6- you may want to write pyspark, and other staffs on your local machine without install spark. There are many ways for that, but i prefer anaconda!
- install everything seperately on local (vscode or another IDE, python3, pip3) and jupiyer-notebook and pyspark
- don't install pyspark, you will see some imports errors in your IDE and also you won't test code interactively, but that's ok. Run your code on GCP instance where you installed spark.
- install anaconda on local (w/ conda package manager), then leverage jupiter-notebooks and install pyspark
for anaconda
brew cask install anaconda
install anaconda w/ brew on Mac
echo 'export PATH="/usr/local/anaconda3/bin/:$PATH"' >> .zshrc
change .profile if not using zsh-terminal!
cd ~/anaconda3/bin
conda update -n base -c defaults conda
conda create --name apache-spark python=3
conda activate apache-spark
conda install -c conda-forge pyspark
* on VSCode, do not forget to switch python interpreter to anaconda python version!
8- you may want to test master-slave, on GCP, then run below commands.
./sbin/start-master.sh
./sbin/start-slave.sh spark://IP:7077
pyspark --master spark://IP:7077
9- to run jupiternotebook on GCP instance,
sudo pip3 install runipy
runipy spark-sandbox/ratings-histogram.ipynb