[Loading Your Data](Tutorial%3A-Loading-Your-Data-Part-2.html) and [All About Queries](Tutorial%3A-All-About-Queries.html) contain recipes to boot a small druid cluster on localhost. However, when it's time to run a more realistic setup—for production or just for testing production—you'll want to find a way to start the cluster on multiple hosts. This document describes two different ways to do this: manually, or as a cloud service via Apache Whirr.
You can provision individual servers, loading Druid onto each machine (or building it) and setting the required configuration for each type of node. You'll also have to set up required external dependencies. Then you'll have to start each node. This process is outlined in [Tutorial: The Druid Cluster](Tutorial%3A-The-Druid-Cluster.html).
[Apache Whirr](http://whirr.apache.org/) is a set of libraries for launching cloud services. For Druid, Whirr serves as an easy way to launch a cluster in Amazon AWS by using simple commands and configuration files (called *recipes*).
You'll need an AWS account, S3 Bucket and an EC2 key pair from that account so that Whirr can connect to the cloud via the EC2 API. If you haven't generated a key pair, see the [AWS documentation](http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-key-pairs.html) or see this [Whirr FAQ](http://whirr.apache.org/faq.html#how-do-i-find-my-cloud-credentials).
In order to run the test below, you'll also need two files that available only from a [standard install of Druid](http://druid.io/downloads.html) or the [Druid repo](https://github.com/druid-io/druid/tree/master/examples/bin/examples):
The Whirr recipe for Druid is the configuration file `$WHIRR_HOME/recipies/druid.properties`. You can edit this file to suit your needs; it is annotated and self-explanatory. Here are some hints about that file:
* You can choose the hardware used with `whirr.hardware-id` to a specific instance type (e.g., m1.large). By default druid.properties, m3.2xlarge (broker, historical, middle manager), m1.xlarge (coordinator, overlord), and m1.small (zookeeper, mysql) are used.
* If you don't choose an image via `whirr.image-id` (image must be compatible with hardware), you'll get plain vanilla Linux. Default druid.properties uses ami-018c9568 (Ubuntu 12.04).
* SSH keys (not password protected) must exist for the local user. If they are in the default locations, `${sys:user.home}/.ssh/id_rsa` and `${sys:user.home}/.ssh/id_rsa.pub`, Whirr will find them. Otherwise, you'll have to specify them with `whirr.private-key-file` and `whirr.public-key-file`.
* Two Druid cluster templates (see `whirr.instance-templates`) are provided: a small cluster running on a single EC2 instance, and a larger cluster running on multiple instances.
* You must specify the path to an S3 bucket. Otherwise the cluster won't be able to process tasks.
* To successfully submit the test task below, you'll need to specify the location of the `wikipedia_realtime.spec` in the property `whirr.druid.realtime.spec.path`.
* Specify Druid version only if [Druid extenions](Modules.html) are being used.
In order to configure each node, you can edit `services/druid/src/main/resources/functions/start_druid.sh` for JVM configuration and `services/druid/src/main/resources/functions/configure_[NODE_NAME].sh` for specific node configuration. For more information on configuration, see the [Druid configuration documentation](Configuration.html).
If both the instances and the Druid cluster launch successfully, a few minutes later other messages to STDOUT should follow with information returned from EC2, including the instance ID:
Note that Whirr will return an exception if any of the nodes fail to launch, and the cluster will be destroyed. To destroy the cluster manually, run the following command:
Now you can run an indexing task and a simple query to see if all the nodes have launched correctly. We are going to use a Wikipedia example again. For a realtime indexing task, run the following command:
where OVERLORD_PUBLIC_IP_ADDR should be available from the EC2 information logged to STDOUT, the Overlord port is 8080 by default, and `wikipedia_realtime_task.json` is discussed above.
To check the state of the overlord, open up your browser and go to `#{OVERLORD_PUBLIC_IP_ADDR}:#{PORT}/console.html`.
Next, go to `#{COORDINATOR_PUBLIC_IP_ADDR}:#{PORT}`. Click "View Information about the Cluster"->"Full Cluster View." You should now see the information about servers and segments. If the cluster runs correctly, Segment dimensions and Segment binaryVersion fields should be filled up. Allow few minutes for the segments to be processed.
