ASG Rolling Updates on AWS with Terraform

The cloud is great. Isn’t the cloud great? …and because you’re reading this you know about running compute instances in the cloud. If you’re unlucky enough to still be dealing with code deployed to machines instead of containers, you might be in the painful world of shipping code by automated or manual move and re-start procedures. Perhaps you are lucky enough to be in a containerized environment, but not on Kubernetes. Maybe you’re using Nomad to manage your containers and thus you’re still producing custom machine images etc… If so, you’re aware of the concept of machines, machine groups, and load balancers. If not, a quick primer:

Load Balancers direct requests from a “single point” to an internal pool of resources, Auto Scaling Groups (ASG) (or Managed Instance Groups in GCP). Load Balancers are typically exposed to the outside world and often used for SSL termination, bridging traffic from the outside world to an internal private network. In addition to the Load Balancers there are the network definitions / rules, ASG definitions, machine images (on AWS known as AMIs), and Launch Configurations which bridge machine images and the type of machine to the ASG.

One great feature of these group is the ability to roll image updates across the “fleet”. Unlike Googles Managed Instance Groups, AWS ASG resources do not natively support rolling update policies. The rolling update policy, monitoring, and execution require life-cycle awareness within the platform to enact changes on ASGs. The most basic example of this is rolling new AMIs via Launch Configuration changes across a managed set of machines. This requires that the ASG be provisioned via a Cloudformation template. Cloudformation is the conduit allowing for the attachment to resources and monitoring of qualifying life-cycle events and prescribed actions. Terraform, Ansible, and other infrastructure provisioning tools, or even more actively managed tools such as Chef and Puppet lack the life-cycle awareness necessary to react.

That said, achieving rolling updates within an ASG with Terraform is easily with a few additional resources over a traditional ASG.

First let’s review a very basic example of an ASG without monitoring, auto scaling, or rollout policy. Note: These examples reference a statically defined VPC, AV Zones, and ELB.

Basic ASG Monitoring and Scaling

For the most basic example we have the following resources and data in Terraform:

  • Data querying for the most recent Ubuntu AMI
  • A launch config (without showing all the networking resources)
  • An ASG
data "aws_ami" "most_recent_ubuntu_xenial" {

  most_recent = true

  filter {
    name = "name"
    values = ["ubuntu/images/hvm-ssd/ubuntu-xenial-16.04-amd64-server-*"]
  }

  filter {
    name = "virtualization-type"
    values = ["hvm"]
  }

  owners = ["099720109477"] # Canonical
}

resource "aws_launch_configuration" "sample_lc" {

  image_id = "${data.aws_ami.most_recent_ubuntu_xenial.id}"
  instance_type = "t2.medium"
}

resource "aws_autoscaling_group" "sample_asg" {

    availability_zones = ["us-west-2a", "us-west-2b"]
    load_balancers = ["elb-123456"]
    vpc_zone_identifier = ["vpc-abc1234"]
    launch_configuration = "${aws_launch_configuration.sample_lc.name}"
    max_size = 5
    min_size = 2
}

Note that the min, max size of the ASG are defined, but have no monitoring or event triggers to scale up from 2 or down from 5, etc… Manipulating the current count in this case requires manual action, typically API calls from custom code.

Monitoring

There are many different metrics available for scrutiny within the ASG resources. The one we are going to focus on is:

  • CPUUtilization

The following block defines a Cloudwatch Metric Alarm that will average the CPU utilization for the machines in an ASG over a 60 second window and trigger an alarm state when that average is >= the treshhold of 75. This functions as our upper bound…

resource "aws_cloudwatch_metric_alarm" "scaling_app_high" {

    alarm_name = "sample-cpu-utilization-exceeds-normal"
    comparison_operator = "GreaterThanOrEqualToThreshold"
    evaluation_periods = "1"
    metric_name = "CPUUtilization"
    namespace = "AWS/EC2"
    period = "60"
    statistic = "Average"
    threshold = "75"

    dimensions {
        AutoScalingGroupName = "${aws_autoscaling_group.sample_asg.name}"
    }
}

The next block is similar, but with different comparison operators and thresholds. It also averages over a longer period of time because scale in events should only occur after whatever event triggering scale out due to load is over.

resource "aws_cloudwatch_metric_alarm" "scaling_app_low" {

    alarm_name = "sample-cpu-utilization-normal"
    comparison_operator = "LessThanThreshold" #  <------- CHANGED
    evaluation_periods = "1"
    metric_name = "CPUUtilization"
    namespace = "AWS/EC2"
    period = "300" #  <---------------------------------- CHANGED
    statistic = "Average"
    threshold = "60" #  <-------------------------------- CHANGED

    dimensions {
        AutoScalingGroupName = "${aws_autoscaling_group.sample_asg.name}"
    }
}

These Cloudwatch Metric Alarm resources in their current state are only alarms and are in capable of influencing scale events on the ASG. In order for the Cloudwatch Metric Alarms to affect the ASG we have to define Autoscaling Policies and bind the alarms to said policies.

Scaling Policies

The scaling policies are the magic bridge between our aforecreated alarms and the changes to our ASGs (within the ASG pre-defined lower and upper bounds).

The following two policies achieve scale up by one and down by one via the ChangeInCapacity adjustment type:

resource "aws_autoscaling_policy" "scale_out_scaling_app" {

    name = "scale-out-cpu-high"
    scaling_adjustment = 1
    adjustment_type = "ChangeInCapacity"
    cooldown = 300
    autoscaling_group_name = "${aws_autoscaling_group.sample_asg.name}"
}

resource "aws_autoscaling_policy" "scale_in_scaling_app" {

    name = "scale-in-cpu-normal"
    scaling_adjustment = -1
    adjustment_type = "ChangeInCapacity"
    cooldown = 300
    autoscaling_group_name = "${aws_autoscaling_group.sample_asg.name}"
}

Tying the knot

Now that we have our Auto Scaling policies defined, we need to alter our Cloudwatch Metric Alarms to invoke these policies when an alarm condition is raised. Achieving this union is simple with the definition of the alarm_action array of ARNs on the target aws_cloudwatch_metric_alarm resources.

Making these changes yields updated Cloudwatch Metric Alarms…:

resource "aws_cloudwatch_metric_alarm" "scaling_app_high" {

    alarm_name = "sample-cpu-utilization-exceeds-normal"
    comparison_operator = "GreaterThanOrEqualToThreshold"
    evaluation_periods = "1"
    metric_name = "CPUUtilization"
    namespace = "AWS/EC2"
    period = "60"
    statistic = "Average"
    threshold = "75"

    dimensions {
        AutoScalingGroupName = "${aws_autoscaling_group.sample_asg.name}"
    }

    alarm_actions = ["${aws_autoscaling_policy.scale_out_scaling_app.arn}"]
}

resource "aws_cloudwatch_metric_alarm" "scaling_app_low" {

    alarm_name = "sample-cpu-utilization-normal"
    comparison_operator = "LessThanThreshold"
    evaluation_periods = "1"
    metric_name = "CPUUtilization"
    namespace = "AWS/EC2"
    period = "300"
    statistic = "Average"
    threshold = "60"

    dimensions {
        AutoScalingGroupName = "${aws_autoscaling_group.sample_asg.name}"
    }

    alarm_actions = ["${aws_autoscaling_policy.scale_out_scaling_app.arn}"]
}

Basic ASG rollout policy support

As previously mentioned in the beginning of this post, rollout policy support requires the definition of something that’s aware of life-cycle events happening to resources within AWS (like the Launch Config). Terraform, Ansible, and other infrastructure tools that run to completion and cloud provider/machine states match the prescribed “code” do not support this.

In order to achieve this, we have to drop the definition we presently have of our aws_autoscaling_group resource sample_asg and replace it with a Cloudformation Template that does the same thing (with the addition of the rollout policy.

The Cloudformation template creates the AWS resources (much like Terraform) and has the advantage of being aware of life-cycle events within the AWS ecosystem.

Cloudformation Template

Our template will place particular emphasis on allowing for changes to the following aspects of the ASG and the update policy:

  • Availability Zone
  • Launch Config
  • VPC Zone ID
  • Load Balancer Name
  • Min and max capacity
  • The rollout update pause time

Our terraform AWS Cloudformation Stack resource defines:

  1. The name of the stack
  2. The parameters passed to the stack
  3. The stack itself

Within the Cloudformation stack or template, there are four sections (defined as the template_body within the opening and closing multiline STACK keyword):

  1. Description (self explanatory)
  2. Parameters
  3. Resources defined
  4. Outputs
resource "aws_cloudformation_stack" "rolling_update_asg" {

    name = "asg-rolling-update"
    template_body = "${file("${path.module}/cloudformation-asg.json")}"

    parameters {

        AvailabilityZones = "us-west-2a,us-west-2b"
        LaunchConfigurationName = "${aws_launch_configuration.sample_lc.name}"
        VPCZoneIdentifier = "vpc-abc1234"
        LoadBalancerNames = "elb-123456"
        MaximumCapacity = "5"
        MinimumCapacity = "2"
        UpdatePauseTime = "60"
    }

      template_body = <<STACK
{
  "Description": "ASG cloud formation template",
  "Parameters": {
    "AvailabilityZones": {
      "Type": "CommaDelimitedList",
      "Description": "The availability zones to be used for the app"
    },
    "LaunchConfigurationName": {
      "Type": "String",
      "Description": "The launch configuration name"
    },
    "VPCZoneIdentifier": {
      "Type": "List<AWS::EC2::Subnet::Id>",
      "Description": "The VPC subnet IDs"
    },
    "LoadBalancerNames": {
      "Type": "CommaDelimitedList",
      "Description": "The load balancer names for the ASG"
    },
    "MaximumCapacity": {
      "Type": "String",
      "Description": "The maximum desired capacity size"
    },
    "MinimumCapacity": {
      "Type": "String",
      "Description": "The minimum and initial desired capacity size"
    },
    "UpdatePauseTime": {
      "Type": "String",
      "Description": "The pause time during rollout for the application"
    }
  },
  "Resources": {
    "ASG": {
      "Type": "AWS::AutoScaling::AutoScalingGroup",
      "Properties": {
        "AvailabilityZones": { "Ref": "AvailabilityZones" },
        "LaunchConfigurationName": { "Ref": "LaunchConfigurationName" },
        "MaxSize": { "Ref": "MaximumCapacity" },
        "MinSize": { "Ref": "MinimumCapacity" },
        "DesiredCapacity": { "Ref": "MinimumCapacity" },
        "VPCZoneIdentifier": { "Ref": "VPCZoneIdentifier" },
        "LoadBalancerNames": { "Ref": "LoadBalancerNames" },
        "TerminationPolicies": [ "OldestLaunchConfiguration", "OldestInstance" ],
        "HealthCheckType": "ELB",
        "MetricsCollection": [
          {
            "Granularity": "1Minute",
            "Metrics": [ ]
          }
        ],
        "HealthCheckGracePeriod": "300",
        "Tags": [ ]
      },
      "UpdatePolicy": {
        "AutoScalingRollingUpdate": {
          "MinInstancesInService": "2",
          "MaxBatchSize": "1",
          "PauseTime": { "Ref": "UpdatePauseTime" }
        }
      }
    }
  },
  "Outputs": {
    "AsgName": {
      "Description": "ASG reference ID",
      "Value": { "Ref": "ASG" }
    }
  }
}
STACK
}

Terraform passes values to the Cloudformation template via the parameters section of template and receives values via the outputs.

The UpdatePolicy sub-document in the JSON template is the magic sauce that controls updates. Read more about it here. Our UpdatePolicy hard codes the min instances in service and max batch size and allow for dynamic configurations of the update PauseTime (for launch configs that spin up faster/slower than others and pass health checks, in this case).

"UpdatePolicy": {
  "AutoScalingRollingUpdate": {
    "MinInstancesInService": "2",
    "MaxBatchSize": "1",
    "PauseTime": {
      "Ref": "UpdatePauseTime"
    }
  }
}

Note that the PauseTime (and other Cloudformation variables) use a JSON sub-document that references a particular key and that key machines a parameter definition in the Cloudformation template… which matches a parameter passed from Terraform.

Finally, take note of the Outputs section of the Cloduformation template. The outputs are made available to the Terraform resource and are referenceable. In this case, we expose the name of the ASG as AsgName and can access this value in Terraform via: ${aws_cloudformation_stack.rolling_update_asg.outputs["AsgName"]}.

Integration with Metric, alarm, scaling policy resources

The final step in this example is to link our previously created metric/alarm and scaling policy resources with the ASG defined in our Cloudformation template. (Note: that we don’t have to link the Launch Configuration because it’s already passed to the Cloudformation template itself.)

Doing so essentially involves changing the autoscaling_group_name attribute on the aws_autoscaling_policy resource using the Cloudformation template output we just discussed.:

resource "aws_autoscaling_policy" "scale_out_scaling_app" {

    name = "scale-out-cpu-high"
    scaling_adjustment = 1
    adjustment_type = "ChangeInCapacity"
    cooldown = 300
    autoscaling_group_name = "${aws_cloudformation_stack.rolling_update_asg.outputs["AsgName"]}"
}

resource "aws_autoscaling_policy" "scale_in_scaling_app" {

    name = "scale-in-cpu-normal"
    scaling_adjustment = -1
    adjustment_type = "ChangeInCapacity"
    cooldown = 300
    autoscaling_group_name = "${aws_cloudformation_stack.rolling_update_asg.outputs["AsgName"]}"
}

…and changing the AutoScalingGroupName dimension on the aws_cloudwatch_metric_alarm resource:

resource "aws_cloudwatch_metric_alarm" "scaling_app_high" {

    alarm_name = "sample-cpu-utilization-exceeds-normal"
    comparison_operator = "GreaterThanOrEqualToThreshold"
    evaluation_periods = "1"
    metric_name = "CPUUtilization"
    namespace = "AWS/EC2"
    period = "60"
    statistic = "Average"
    threshold = "75"

    dimensions {
        AutoScalingGroupName = "${aws_cloudformation_stack.rolling_update_asg.outputs["AsgName"]}"
    }

    alarm_actions = ["${aws_autoscaling_policy.scale_out_scaling_app.arn}"]
}

resource "aws_cloudwatch_metric_alarm" "scaling_app_low" {

    alarm_name = "sample-cpu-utilization-normal"
    comparison_operator = "LessThanThreshold"
    evaluation_periods = "1"
    metric_name = "CPUUtilization"
    namespace = "AWS/EC2"
    period = "300"
    statistic = "Average"
    threshold = "60"

    dimensions {
        AutoScalingGroupName = "${aws_cloudformation_stack.rolling_update_asg.outputs["AsgName"]}"
    }

    alarm_actions = ["${aws_autoscaling_policy.scale_out_scaling_app.arn}"]
}

Tada, magic

…and there you have it! Any changes to Launch Configuration will change the input to the Cloudformation template that creates your ASG and the remaining infrastructure is directly managed by Terraform.

If, at this point, you’re asking yourself why you wouldn’t just do all of this in Cloudformation, my answer for you would be… Cloudformatio is key, again, to resources that require life cycle awareness. Other than this, though, you’re dealing with a syntax that’s more difficult to work with (JSON over HCL), and vendor lock in. Terraform shines when you link multiple cloud providers together, leveraging each for the best tasks its capable of (Kubernetes on GCP for example). Also, here is Hashicorp’s response to that question.

comments powered by Disqus