Dataproc Client

This repository provides a startup script and instructions for configuring a GCP VM Instance to submit jobs to a remote Dataproc cluster.

Supported clients:

• Zeppelin
• Hue
• Hive
• Beeline
• Hadoop / HDFS
• Spark Submit
• Spark Shell

Instructions

1. Change working directory to repository root

cd dataproc-client-vm-startup

3. Modify metadata.config

Replace the following placeholders:

• YOUR_TARGET_CLUSTER The server Dataproc Cluster name (DON'T include the '-m') where the client VM will point to to run Hive and Spark, for example 'hiver-server'
• YOUR_PROJECT The project name
• YOUR_CLIENT The Client Cluster name (DON"T include the '-m'), for example "test-client-dp-demo"
• YOUR_SERVICE_ACCOUNT The service account (in email format like '[email protected]`). It should have the required access to GCP APIs/services, for example, Dataproc, Compute Engine, GCS.
• YOUR_BUCKET The GCS bucket name that will be used to transfer the scripts to the client VM
• YOUR_STORAGE_CLASS The storage class for the GCS bucket, multi_regional recommended as there might be different regions where the client VM might be
• YOUR_REGION The region where you want to establish the client VM
• YOUR_ZONE The zone where you want to establish the client VM
• YOUR_CLIENT_TAG The network tag to be used for the client VM
• YOUR_NETWORK The network to be used for the client VM

4. Run setup script. This will create a Dataproc Client VM that is pointed to the first server.

create-dproclient-vm.sh

./create-dproclient-vm.sh

5. Create an image using the just created single-nodel Dataproc VM disk, with name for example "dproclient-image"

6. Change working directory to from-img sub-folder

cd dataproc-client-vm-startup/from-img

7. Modify metadata.config

Replace the following placeholders:

• YOUR_TARGET_CLUSTER The new server Dataproc Cluster name (DON'T include the '-m') where you want the new client VM to point to to run Hive and Spark, for example 'hiver-server-bigger'
• YOUR_IMAGE The Dataproc single-node master VM image created in step-5 that will be used to create the compute engine as client VM, for example "test-client-dp-demo-m-image"
• OLD_TARGET_CLUSTER The old server target cluster name that was used to create the image (DON'T include the '-m'), for example "hive-server"
• IMG_MACHINE_TYPE The machine type of the VM that was used to create the image. If use the current default settings in create-dproclient-vm.sh, then it should be "n1-standard-1".
• IMG_BOOT_DISK_SIZE The boot disk size of the VM that was used to create the image. If use the current default settings in create-dproclient-vm.sh, then it should be "200GB".
• IMG_BOOT_DISK_TYPE The boot disk type of the VM that was used to create the image. If use the current default settings in create-dproclient-vm.sh, then it should be "pd-standard".
• YOUR_PROJECT The project name
• YOUR_CLIENT The Client VM name
• YOUR_REGION The region where you want to establish the client VM
• YOUR_ZONE The zone where you want to establish the client VM
• YOUR_CLIENT_TAG The network tag to be used for the client VM
• YOUR_NETWORK The network to be used for the client VM
• YOUR_SERVICE_ACCOUNT The service account (in email format like '[email protected]`)

8. Run setup script. This will create another VM (Compute Egngine) that is pointed to the new server.

create-dproclient-from-img.sh

./create-dproclient-from-img.sh

9. Access Zeppelin web UI hosted on the just created Client VM

http://localhost:8080

Create a JDBC interpreter connecting Zeppelin with the remote Dataproc cluster:

• default.driver -> org.apache.hive.jdbc.HiveDriver
• default.url -> jdbc:hive2://<target_dataproc_cluster_name-m>:10000/
• artifacts -> org.apache.hive:hive-jdbc:0.14.0 & org.apache.hadoop:hadoop-common:2.6.0

10. Access Hue Web UI hosted on the just created Client VM

http://localhost:8888

Note that as the new client VM has startup script that needs to run during execution, even though processes may look complete in the terminal or the Cloud Console may appear to have the VM set up completely, the startup script may still be running (you can check serial log port output to see the status). So it might require more time for the configurations and UI to be fully done.

How this works

The create-dproclient-vm.sh contains the the following

• create a Cloud storage bucket used to store those configurations files passed to the client dataproc .
• gcloud command that starts a Dataproc cluster
• gcloud command that will SSH into the Dataproc client VM and run util/install-client.sh

The util/install-client.sh constains the following

• Installation of Hue and Zeppeline on it. The zeppelin.sh and hue.sh scripts are the initialization-actions bash files provided by Google, which will initilze Zeppelin and Hue

• replace the hostnames in the Hadoop, Yarn, Hive, Spark files with the hostname with the hostname of the remote target server dataproc cluster. Thus, when starting hive, spark-shell, , spark-submit, the session will be created on the remote cluster instead of locally.

• swap the /etc/hue/conf/hue.ini hive_server_host setting to the target dataproc cluster instead of the local and also uncommented the hive_server_port configuration that are needed to have hue running query successfully.

The from-img/create-startup.sh will create the from-img/startup-script.sh that include contents of the following

• replace the old target (server) cluster names in the Hadoop, Yarn, Hive, Spark files with the new cluster name of the remote target server dataproc cluster. Thus, when starting hive, spark-shell, , spark-submit, the session will be created on the remote cluster instead of locally.

• swap the /etc/hue/conf/hue.ini hive_server_host setting to the new target dataproc cluster instead of the old target.

• Note: The Zeppelin configurations needed to allow Zeppelin to connect with the target dataproc hive server are set in step #9 above through web server UI.

The from-img/create-dproclient-from-img.sh will do the following

• read the from-img/metadata.config to get the metadata and run the from-img/create-startup.sh to create the from-img/startup-script.sh

• gcloud command that starts a new client VM (compute engine) using image

Note: The following error can be ignored:

• When first time openning Hue web UI, the below error will pop up but can be ignored. There should be no problem running Hive queres that run on the server machine. Solr server could not be contacted properly: HTTPConnectionPool(host='......', port=8983): Max retries exceeded with url: /solr/admin/info/system?user.name=hue&doAs=admin&wt=json (Caused by NewConnectionError(': Failed to establish a new connection: [Errno 111] Connection refused',))

Troubleshooting

1. Make sure you have the server with reasonable computation power. If the server is too small, the spark shell and spart submit may take too long or have YARN warnings/errors. Below is an example of creation of a Dataproc Server with the reasonable size and related settings like tags, etc.

gcloud dataproc clusters create hive-server-bigger-2 --region us-east1 --subnet default --zone "" --service-account=hive-demo@<your_project>.iam.gserviceaccount.com --master-machine-type n1-standard-1 --master-boot-disk-size 500  --num-workers 2   --worker-machine-type n1-standard-1 --worker-boot-disk-size 500 --image-version 1.3-deb9 --scopes 'https://www.googleapis.com/auth/cloud-platform' --tags hive-server --project <your_project>

2. Unmatched python versions on server and client sides

ERROR org.apache.spark.deploy.yarn.ApplicationMaster: User class threw exception: java.io.IOException: Cannot run program "/opt/conda/default/bin/python": error=2, No such file or directory

spark-submit may fail even spark-shell or pyspark run correctly due to unmatched versions on server and client sides.

Run which python and python --version to check if the python versions match.

One example of mis-match may be as follow:

• Server side: Python 2.7.13 | /usr/bin/python
• Client side: Python 3.6.5 :: Anaconda, Inc.| /opt/conda/default/bin/python .

It may be a simple fix to match the python env on both machines after the client was created but it's suggested to select image with the same python package and same settings with the server during creation of the client. So the gcloud image-version flag in create-dproclient-vm.sh should be "preview" if the server is using python3.5 with conda, or 1.3-deb9 (default as of March 2019) or older images that are using Python2.7.13 with python path /usr/bin/python.

Testing

1. Find the Dataproc Server cluster name and the image version to be referenced in the automation script

2. Decide the network tag to be used for the Dataproc client VM and use that to set up the firewall rules for the following ports to be opened.

• Client to master connection: for example, Hive: 10000 (hive server),9083 (hive metastore), spark: 7077 (spark master port), 7078 (spark worker port)
• Hue web server: 8888 - user terminal to client
• Zeppelin web server: 8080 - user terminal to client
• Spark web server: 18080 (spark master web ui port), - user temrinal to master

3. Create Client Dataproc VM with the above instructions

4. Validate the success criterion have been met

• Test Hive CLI: User should be able to query tables stored on master Dataproc hive server
• Test Spark Shell ( spark-shell for scala and pyspark for python): Note that it might take a while to initiate the shell. The job submitted on client VM should be shown on master Dataproc spark history-server UI (http://<spark_host_ip>:18080) - but it may only be shown after exit() the shell.
• Test spark-submit: The job submitted on client VM should be shown on master Dataproc spark history-server UI (http://<spark_host_ip>:18080)
• Test running Hive Query from Zeppelin: User should be able to query tables stored on master Dataproc hive server
• Test running Hive Query from Hue: User should be able to query tables stored on master Dataproc hive server

To-do

• Automate the image creation and simple the steps for usinig the package.

• Image Version as input metadata that matches the server to be used to create the image

• Currently opend all ports from client to server

Disclaimer

This is not an official Google project.