Drain ECS instances before scaling down in EC2
We have 2 independent auto scaling systems: EC2 auto scaling groups which scales instances number & ECS auto scaling which scales tasks number. This setup may work fine when scaling up. Still, if new tasks need more instances to start up, it must wait for them, but it’s ok to wait a little.
But things soon become disaster when EC2 instances terminates before ECS tasks draining out.
Actually, ECS instances will not auto drain to be ready for terminating instances in EC2, so don’t expect anything. We have to set that up manually.
An AWS official blog shows us how to do this, and it has a nice demo for newly created EC2 & ECS setup. But what if we just want to set up for our running ECS instances? Here it is.
Firstly, make sure you know what lifecycle hooks are. Then follow these steps:
2, Set role for our lambda function to have these permissions:
- autoscaling:CompleteLifecycleAction - logs:CreateLogGroup - logs:CreateLogStream - logs:PutLogEvents - ec2:DescribeInstances - ec2:DescribeInstanceAttribute - ec2:DescribeInstanceStatus - ec2:DescribeHosts - ecs:ListContainerInstances - ecs:SubmitContainerStateChange - ecs:SubmitTaskStateChange - ecs:DescribeContainerInstances - ecs:UpdateContainerInstancesState - ecs:ListTasks - ecs:DescribeTasks - sns:Publish
3, Setup an SNS trigger for the function, choose a name for it.
4, Setup a service role so Auto Scaling Groups can push to SNS, using this guide. (This is quite interesting because you have to set up a role so that a service will have some kind of permission).
5, Create a lifecycle hook by CLI, since the GUI is not fully supported yet:
aws autoscaling put-lifecycle-hook --lifecycle-hook-name EcsWebScaleDown --auto-scaling-group-name ecs-web --lifecycle-transition autoscaling:EC2_INSTANCE_TERMINATING --heartbeat-timeout 900 --notification-target-arn arn:aws:sns:ap-southeast-1:XXXXXXXXXXXX:EcsInstanceDrain --role-arn arn:aws:iam::XXXXXXXXXXXX:role/AutoScalingNotificationRole
It should be good now, as advertised :)