ecs-deploy

This script uses the Task Definition and Service entities in Amazon's ECS to instigate an automatic blue/green deployment.

Usage

Required arguments:
    -k | --aws-access-key         AWS Access Key ID. May also be set as environment variable AWS_ACCESS_KEY_ID
    -s | --aws-secret-key         AWS Secret Access Key. May also be set as environment variable AWS_SECRET_ACCESS_KEY
    -r | --region                 AWS Region Name. May also be set as environment variable AWS_DEFAULT_REGION
    -p | --profile                AWS Profile to use - If you set this aws-access-key, aws-secret-key and region are not needed
       | --aws-instance-profile   Use the IAM role associated with the current AWS instance. Can only be used from within a running AWS instance. If you set this, aws-access-key and aws-secret-key are not needed
    -c | --cluster                Name of ECS cluster
    -n | --service-name           Name of service to deploy
    -i | --image                  Name of Docker image to run, ex: repo/image:latest
                                  Format: [domain][:port][/repo][/][image][:tag]
                                  Examples: mariadb, mariadb:latest, silintl/mariadb,
                                            silintl/mariadb:latest, private.registry.com:8000/repo/image:tag

Optional arguments:
    -D | --desired-count          The number of instantiations of the task to place and keep running in your service.
    -m | --min                    minumumHealthyPercent: The lower limit on the number of running tasks during a deployment. (default: 100)
    -M | --max                    maximumPercent: The upper limit on the number of running tasks during a deployment. (default: 200)
    -t | --timeout                Default is 90s. Script monitors ECS Service for new task definition to be running.
    -e | --tag-env-var            Get image tag name from environment variable. If provided this will override value specified in image name argument.
    --max-definitions             Number of Task Definition Revisions to persist before deregistering oldest revisions.
                                  Note: This number must be 1 or higher (i.e. keep only the current revision ACTIVE).
                                        Max definitions causes all task revisions not matching criteria to be deregistered, even if they're created manually.
                                        Script will only perform deregistration if deployment succeeds.
    --enable-rollback             Rollback task definition if new version is not running before TIMEOUT
    -v | --verbose                Verbose output
         --version                Display the version

Examples:
  Simple (Using env vars for AWS settings):

    ecs-deploy -c production1 -n doorman-service -i docker.repo.com/doorman:latest

  All options:

    ecs-deploy -k ABC123 -s SECRETKEY -r us-east-1 -c production1 -n doorman-service -i docker.repo.com/doorman -m 50 -M 100 -t 240 -D 2 -e CI_TIMESTAMP -v

    Using profiles (for STS delegated credentials, for instance):

    ecs-deploy -p PROFILE -c production1 -n doorman-service -i docker.repo.com/doorman -m 50 -M 100 -t 240 -e CI_TIMESTAMP -v

Notes:
  - If a tag is not found in image and an ENV var is not used via -e, it will default the tag to "latest"

Installation

• Install and configure aws-cli
• Install jq
• Install ecs-deploy:

curl https://raw.githubusercontent.com/silinternational/ecs-deploy/master/ecs-deploy | sudo tee -a /usr/bin/ecs-deploy
sudo chmod +x /usr/bin/ecs-deploy

How it works

Note: Some nouns in the next paragraphs are capitalized to indicate that they are words which have specific meanings in AWS

Remember that in the EC2 Container Service, the relationship between the group of containers which together provide a useful application (e.g. a database, web frontend, and perhaps some for maintenance/cron) is specified in a Task Definition. The Task Definition then acts a sort of template for actually running the containers in that group. That resulting group of containers is known as a Task. Due to the way docker implements networking, generally you can only run one Task per Task Definition per Container Instance (the virtual machines providing the cluster infrastructure).

Task Definitions are automatically version controlled---the actual name of a Task Definition is composed of two parts, the Family name, and a version number, like so: phpMyAdmin:3

Since a Task is supposed to be a fully self-contained "worker unit" of a broader application, Amazon uses another configuration entity, Services, to manage the number of Tasks running at any given time. As Tasks are just instantiations of Task Definitions, a Service is just a binding between a specified revision of a Task Definition, and the number of Tasks which should be run from it.

Conveniently, Amazon allows this binding to be updated, either to change the number of Tasks running or to change the Task Definition they are built from. In the former case, the Service will respond by building or killing Tasks to bring the count to specifications. In the latter case, however, it will do a blue/green deployment, that is, before killing any of the old Tasks, it will first ensure that a new Task is brought up and ready to use, so that there is no loss of service.

Naturally, enough computing resources must be available in the ECS cluster for any of this to work.

Consequently, all that is needed to deploy a new version of an application is to update the Service which is running its Tasks to point at a new version of the Task Definition. ecs-deploy uses the python aws utility to do this. It,

• Pulls the JSON representation of the in-use Task Definition
• Edits it
• Defines a new version, with the changes
• Updates the Service to use the new version
• Waits, querying Amazon's API to make sure that the Service has been able to create a new Task

The second step merits more explanation: since a Task Definition [may] define multiple containers, the question arises, "what must be changed to create a new revision?" Empirically, the surprising answer is nothing; Amazon allows you to create a new but identical version of a Task Definition, and the Service will still do a blue/green deployment of identical tasks.

Nevertheless, since the system uses docker, the assumption is that improvements to the application are built into its container images, which are then pushed into a repository (public or private), to then be pulled down for use by ECS. This script therefore uses the specified image parameter as a modification key to change the tag used by a container's image. It looks for images with the same repository name as the specified parameter, and updates its tag to the one in the specified parameter.

A direct consequence of this is that if you define more than one container in your Task Definition to use the same image, all of them will be updated to the specified tag, even if you set them to use different tags initially. But this is considered to be an unlikely use case.

This behavior allows two possible process to specify which images, and therefore which configurations, to deploy. First, you may set the tag to always be latest (or some other static value), like so:

ecs-deploy -c CLUSTERNAME -n SERVICENAME -i my.private.repo.com/frontend_container:latest

This will result in identical new versions of the Task Definition being created, but the Service will still do a blue/green deployment, and will so will pull down the latest version (if you previously pushed it into the registry).

Alternatively, you may specify some other means of obtaining the tag, since the script evals the image string. You could use git tags as a map to docker tags:

ecs-deploy -c CLUSTERNAME -n SERVICENAME -i 'my.private.repo.com/frontend_container:`git describe`'

Or perhaps just obtain read the docker tag from another file in your development:

ecs-deploy -c CLUSTERNAME -n SERVICENAME -i 'my.private.repo.com/frontend_container:$(< VERSION)'

In any case, just make sure your process builds, tags, and pushes the docker image you use to the repository before running this script.

Use Environment Variable for tag name value

In some cases you may want to use an environment variable for the tag name of your image. For instance, we use Codeship for continous integration and deployment. In their Docker environment they can build images and tag them with different variables, such as the current unix timestamp. We want to use these unique and changing values for image tags so that each task definition refers to a unique docker image/tag. This gives us the ability to revert/rollback changes by just selecting a previous task definition and updating the service. We plan to add a revert command/option to ecs-deploy to simplify this further.

Using the -e argument you can provide the name of an environment variable that holds the value you wish to use for the tag. On Codeship they set an env var named CI_TIMESTAMP.

So we use ecs-deploy like this:

ecs-deploy -c production1 -n doorman-api -i my.private.repo/doorman-api -e CI_TIMESTAMP

AWS IAM Policy Configuration

Here's an example of a suitable custom policy for AWS IAM:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "ecs:DeregisterTaskDefinition",
        "ecs:DescribeServices",
        "ecs:DescribeTaskDefinition",
        "ecs:DescribeTasks",
        "ecs:ListTasks",
        "ecs:ListTaskDefinitions",
        "ecs:RegisterTaskDefinition",
        "ecs:StartTask",
        "ecs:StopTask",
        "ecs:UpdateService",
        "iam:PassRole"
      ],
      "Resource": "*"
    }
  ]
}

Troubleshooting

• You must provide AWS credentials in one of the supported formats. If you do not, you'll see some error output from the AWS CLI, something like:

   You must specify a region. You can also configure your region by running "aws configure".
  

Testing

Automated tests are performed using bats. The goal of testing is to ensure that updates/changes do not break core functionality. Unfortunately not all of ecs-deploy is testable since portions interact with AWS APIs to perform actions. So for now any parsing/processing of data locally is tested.

Any new functionality and pull requests should come with tests as well (if possible).