Tutorial: Load batch data using Apache Hadoop

This tutorial shows you how to load data files into Apache Druid using a remote Hadoop cluster.

For this tutorial, we’ll assume that you’ve already completed the previous batch ingestion tutorial using Druid’s native batch ingestion system and are using the micro-quickstart single-machine configuration as described in the quickstart.

Install Docker

This tutorial requires Docker to be installed on the tutorial machine.

Once the Docker install is complete, please proceed to the next steps in the tutorial.

Build the Hadoop docker image

For this tutorial, we’ve provided a Dockerfile for a Hadoop 2.8.5 cluster, which we’ll use to run the batch indexing task.

This Dockerfile and related files are located at quickstart/tutorial/hadoop/docker.

From the apache-druid-0.20.0 package root, run the following commands to build a Docker image named “druid-hadoop-demo” with version tag “2.8.5”:

  1. cd quickstart/tutorial/hadoop/docker
  2. docker build -t druid-hadoop-demo:2.8.5 .

This will start building the Hadoop image. Once the image build is done, you should see the message Successfully tagged druid-hadoop-demo:2.8.5 printed to the console.

Setup the Hadoop docker cluster

Create temporary shared directory

We’ll need a shared folder between the host and the Hadoop container for transferring some files.

Let’s create some folders under /tmp, we will use these later when starting the Hadoop container:

  1. mkdir -p /tmp/shared
  2. mkdir -p /tmp/shared/hadoop_xml

Configure /etc/hosts

On the host machine, add the following entry to /etc/hosts:

  1. 127.0.0.1 druid-hadoop-demo

Start the Hadoop container

Once the /tmp/shared folder has been created and the etc/hosts entry has been added, run the following command to start the Hadoop container.

  1. docker run -it -h druid-hadoop-demo --name druid-hadoop-demo -p 2049:2049 -p 2122:2122 -p 8020:8020 -p 8021:8021 -p 8030:8030 -p 8031:8031 -p 8032:8032 -p 8033:8033 -p 8040:8040 -p 8042:8042 -p 8088:8088 -p 8443:8443 -p 9000:9000 -p 10020:10020 -p 19888:19888 -p 34455:34455 -p 49707:49707 -p 50010:50010 -p 50020:50020 -p 50030:50030 -p 50060:50060 -p 50070:50070 -p 50075:50075 -p 50090:50090 -p 51111:51111 -v /tmp/shared:/shared druid-hadoop-demo:2.8.5 /etc/bootstrap.sh -bash

Once the container is started, your terminal will attach to a bash shell running inside the container:

  1. Starting sshd: [ OK ]
  2. 18/07/26 17:27:15 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
  3. Starting namenodes on [druid-hadoop-demo]
  4. druid-hadoop-demo: starting namenode, logging to /usr/local/hadoop/logs/hadoop-root-namenode-druid-hadoop-demo.out
  5. localhost: starting datanode, logging to /usr/local/hadoop/logs/hadoop-root-datanode-druid-hadoop-demo.out
  6. Starting secondary namenodes [0.0.0.0]
  7. 0.0.0.0: starting secondarynamenode, logging to /usr/local/hadoop/logs/hadoop-root-secondarynamenode-druid-hadoop-demo.out
  8. 18/07/26 17:27:31 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
  9. starting yarn daemons
  10. starting resourcemanager, logging to /usr/local/hadoop/logs/yarn--resourcemanager-druid-hadoop-demo.out
  11. localhost: starting nodemanager, logging to /usr/local/hadoop/logs/yarn-root-nodemanager-druid-hadoop-demo.out
  12. starting historyserver, logging to /usr/local/hadoop/logs/mapred--historyserver-druid-hadoop-demo.out
  13. bash-4.1#

The Unable to load native-hadoop library for your platform... using builtin-java classes where applicable warning messages can be safely ignored.

Accessing the Hadoop container shell

To open another shell to the Hadoop container, run the following command:

  1. docker exec -it druid-hadoop-demo bash

Copy input data to the Hadoop container

From the apache-druid-0.20.0 package root on the host, copy the quickstart/tutorial/wikiticker-2015-09-12-sampled.json.gz sample data to the shared folder:

  1. cp quickstart/tutorial/wikiticker-2015-09-12-sampled.json.gz /tmp/shared/wikiticker-2015-09-12-sampled.json.gz

Setup HDFS directories

In the Hadoop container’s shell, run the following commands to setup the HDFS directories needed by this tutorial and copy the input data to HDFS.

  1. cd /usr/local/hadoop/bin
  2. ./hdfs dfs -mkdir /druid
  3. ./hdfs dfs -mkdir /druid/segments
  4. ./hdfs dfs -mkdir /quickstart
  5. ./hdfs dfs -chmod 777 /druid
  6. ./hdfs dfs -chmod 777 /druid/segments
  7. ./hdfs dfs -chmod 777 /quickstart
  8. ./hdfs dfs -chmod -R 777 /tmp
  9. ./hdfs dfs -chmod -R 777 /user
  10. ./hdfs dfs -put /shared/wikiticker-2015-09-12-sampled.json.gz /quickstart/wikiticker-2015-09-12-sampled.json.gz

If you encounter namenode errors when running this command, the Hadoop container may not be finished initializing. When this occurs, wait a couple of minutes and retry the commands.

Configure Druid to use Hadoop

Some additional steps are needed to configure the Druid cluster for Hadoop batch indexing.

Copy Hadoop configuration to Druid classpath

From the Hadoop container’s shell, run the following command to copy the Hadoop .xml configuration files to the shared folder:

  1. cp /usr/local/hadoop/etc/hadoop/*.xml /shared/hadoop_xml

From the host machine, run the following, where {PATH_TO_DRUID} is replaced by the path to the Druid package.

  1. mkdir -p {PATH_TO_DRUID}/conf/druid/single-server/micro-quickstart/_common/hadoop-xml
  2. cp /tmp/shared/hadoop_xml/*.xml {PATH_TO_DRUID}/conf/druid/single-server/micro-quickstart/_common/hadoop-xml/

Update Druid segment and log storage

In your favorite text editor, open conf/druid/single-server/micro-quickstart/_common/common.runtime.properties, and make the following edits:

Disable local deep storage and enable HDFS deep storage

  1. #
  2. # Deep storage
  3. #
  4. # For local disk (only viable in a cluster if this is a network mount):
  5. #druid.storage.type=local
  6. #druid.storage.storageDirectory=var/druid/segments
  7. # For HDFS:
  8. druid.storage.type=hdfs
  9. druid.storage.storageDirectory=/druid/segments

Disable local log storage and enable HDFS log storage

  1. #
  2. # Indexing service logs
  3. #
  4. # For local disk (only viable in a cluster if this is a network mount):
  5. #druid.indexer.logs.type=file
  6. #druid.indexer.logs.directory=var/druid/indexing-logs
  7. # For HDFS:
  8. druid.indexer.logs.type=hdfs
  9. druid.indexer.logs.directory=/druid/indexing-logs

Restart Druid cluster

Once the Hadoop .xml files have been copied to the Druid cluster and the segment/log storage configuration has been updated to use HDFS, the Druid cluster needs to be restarted for the new configurations to take effect.

If the cluster is still running, CTRL-C to terminate the bin/start-micro-quickstart script, and re-run it to bring the Druid services back up.

Load batch data

We’ve included a sample of Wikipedia edits from September 12, 2015 to get you started.

To load this data into Druid, you can submit an ingestion task pointing to the file. We’ve included a task that loads the wikiticker-2015-09-12-sampled.json.gz file included in the archive.

Let’s submit the wikipedia-index-hadoop.json task:

  1. bin/post-index-task --file quickstart/tutorial/wikipedia-index-hadoop.json --url http://localhost:8081

Querying your data

After the data load is complete, please follow the query tutorial to run some example queries on the newly loaded data.

Cleanup

This tutorial is only meant to be used together with the query tutorial.

If you wish to go through any of the other tutorials, you will need to:

  • Shut down the cluster and reset the cluster state by removing the contents of the var directory under the druid package.
  • Revert the deep storage and task storage config back to local types in conf/druid/single-server/micro-quickstart/_common/common.runtime.properties
  • Restart the cluster

This is necessary because the other ingestion tutorials will write to the same “wikipedia” datasource, and later tutorials expect the cluster to use local deep storage.

Example reverted config:

  1. #
  2. # Deep storage
  3. #
  4. # For local disk (only viable in a cluster if this is a network mount):
  5. druid.storage.type=local
  6. druid.storage.storageDirectory=var/druid/segments
  7. # For HDFS:
  8. #druid.storage.type=hdfs
  9. #druid.storage.storageDirectory=/druid/segments
  10. #
  11. # Indexing service logs
  12. #
  13. # For local disk (only viable in a cluster if this is a network mount):
  14. druid.indexer.logs.type=file
  15. druid.indexer.logs.directory=var/druid/indexing-logs
  16. # For HDFS:
  17. #druid.indexer.logs.type=hdfs
  18. #druid.indexer.logs.directory=/druid/indexing-logs

Further reading

For more information on loading batch data with Hadoop, please see the Hadoop batch ingestion documentation.