I want to have a scale-in policy where nodes are scaled-in when the CPU is under 30%.
How can I control that the individual instance selected for scaling-in is the one with lowest CPU or at least a CPU under 30%?
Add a comment |
1 Answer
By default, Auto Scaling first identifies which Availability Zone has the most instances (or, if they are equal, it picks a random AZ). Then, it uses the Termination Policy to determine which instance to control which instance Auto Scaling terminates during scale-in.
However, Auto Scaling does not select an instance based upon how 'busy' that particular is -- after all, that instance might be more busy, or less busy, a few seconds later.
Fortunately, Auto Scaling uses connection draining to allow in-flight requests to complete before the instance is terminated. Therefore, in theory, it doesn't matter that the instance is temporarily busy.
If you have long-running tasks on an instance that you don't want interrupted, you can configure Auto Scaling Lifecycle Hooks to move instances into a Terminating:Wait state. The instances will not receive any new traffic. Your application can then signal when the long-running task has completed (eg copying log files to S3, or finishing video rendering) and Auto Scaling will then terminate the instance.
Finally, if you want more fine-grained control over which instance will be terminated, you (or your application) can specifically Detach EC2 Instances From Your Auto Scaling Group via the management console or a describe-auto-scaling-instances API call.