Using Spot Instances
To help manage cloud resource costs, Cloudera supports Spot instances. Spot instances are Amazon EC2 instances that you can bid on. Unlike On-Demand Amazon EC2 instances, Spot instances only run as long as the price you bid exceeds the current Spot price. This allows you to add capacity to your workload at a low price.
Spot instances run just like On-Demand instances, except that they are not provisioned until the instance price falls below your bid. They also terminate automatically when the instance price exceeds or equals your bid price.
Planning for Spot Instances
It is normal for Spot instances on a cluster to disappear over time. However, Cloudera Manager does not see that these instances are terminated. If you restart a cluster that contains a Spot instance group, and the Spot instances have terminated, the restart fails. If you are modifying any group in the cluster that has lost Spot instances, do not select the Restart checkbox.
- Delete the entire group.
- Add more instances to the group.
- Delete unprovisioned instances from the group (only as part of adding more instances to the group).
- Retry (repair) existing instances.
- Change the bid price
- Delete all instances without adding more
The bid price for Spot instances is set in an instance template. This template is associated with a group. Although you can modify the group, you cannot change the bid price. Therefore, if you set the bid price too low for successful provisioning, you must delete the group where that price is set and create a new group with the higher bid price. You must also delete the current group and create a new one if you want to drop the bid price.
Specifying Spot Instances
To specify Spot instances, create a new instance template and use this template for your group. For more information, see the steps for adding a cluster in the Deploying Cloudera Manager and CDH on AWS topic.
Best Practices for Using Spot Instances
- Use a Spot instance worker group in conjunction with an On-Demand worker group. This ensures that the cluster can redo computational tasks run on Spot instances that could be terminated before the tasks are finished.
- Use Spot instances only in contexts where the loss of the instance can be tolerated, as in a worker group. Do not use Spot instances for master nodes or for data storage.
- Use a minimum count of 0 for Spot instance groups. If you use a number above 0, the cluster will likely enter a failed state. If the cluster fails, contact Cloudera support for help.