Container and Serverless Deployments

What it is

A deployment strategy decides how new code replaces old code and how much risk users face while it happens. The main strategies are all at once, rolling, blue green, canary and linear, and each has a different cost and rollback speed.

Containers on Amazon Elastic Container Service (ECS) and functions on AWS Lambda both support gradual traffic shifting. AWS Elastic Beanstalk packages several strategies behind one setting.

How it works

Strategy comparison

ECS rolling updates

The ECS deployment controller replaces tasks gradually. Two settings bound the process: minimumHealthyPercent is the lowest share of desired tasks that must stay running, and maximumPercent is the most that may run at once, including new tasks. The deployment circuit breaker detects tasks that keep failing and can roll the service back to the last completed deployment.

{
  "deploymentConfiguration": {
    "minimumHealthyPercent": 100,
    "maximumPercent": 200,
    "deploymentCircuitBreaker": {"enable": true, "rollback": true}
  }
}

ECS blue green with CodeDeploy

For a controlled traffic shift, set the service to use the CodeDeploy deployment controller. You need an Application Load Balancer or Network Load Balancer with two target groups. CodeDeploy starts the new task set (green) in one target group, optionally exposes it on a test listener, then moves the production listener to it.

The appspec for ECS names the task definition and the container that receives traffic.

version: 0.0
Resources:
  - TargetService:
      Type: AWS::ECS::Service
      Properties:
        TaskDefinition: "arn:aws:ecs:us-east-1:111122223333:task-definition/web:7"
        LoadBalancerInfo:
          ContainerName: "web"
          ContainerPort: 80
Hooks:
  - AfterAllowTestTraffic: "RunSmokeTests"
  - BeforeAllowTraffic: "CheckDatabase"

ECS hooks are BeforeInstall, AfterInstall, AfterAllowTestTraffic, BeforeAllowTraffic and AfterAllowTraffic. Each names a Lambda function that must report success or failure back to CodeDeploy.

Built in ECS configurations include all at once, linear and canary types. Amazon ECS also offers a built in blue green option managed by the ECS service itself, so check which deployment controller a question names.

Lambda versions and aliases

A published Lambda version is immutable. An alias is a pointer to a version, and it can split traffic between two versions by weight. Clients call the alias, so you shift traffic without changing callers.

aws lambda update-alias --function-name orders --name live \
  --function-version 2 \
  --routing-config '{"AdditionalVersionWeights":{"3":0.1}}'

That command keeps most traffic on version 2 and sends 10 percent to version 3.

Lambda with CodeDeploy and SAM

CodeDeploy automates the shifting for Lambda, using canary or linear configurations such as LambdaCanary10Percent5Minutes and LambdaLinear10PercentEvery1Minute. The AWS Serverless Application Model (SAM) wires this up in a few lines.

Resources:
  OrderFunction:
    Type: AWS::Serverless::Function
    Properties:
      Handler: app.handler
      Runtime: python3.12
      AutoPublishAlias: live
      DeploymentPreference:
        Type: Canary10Percent5Minutes
        Alarms:
          - !Ref ErrorAlarm
        Hooks:
          PreTraffic: !Ref PreTrafficFunction

Lambda deployments have two hooks: BeforeAllowTraffic and AfterAllowTraffic. In SAM they are named PreTraffic and PostTraffic. If the listed alarm fires during the shift, CodeDeploy moves the alias back to the old version.

Elastic Beanstalk

Beanstalk offers deployment policies as a setting on the environment.

Blue green in Beanstalk means a second environment and a swap of the environment URLs (CNAME swap). The database should live outside the environment so a swap does not lose data.

Storage in deployments

Deployments replace compute, so keep state off the compute. Pick the store by how the data is used.

EBS ties data to a single instance, which fights immutable and blue green releases because the replacement instances start with new volumes. Store durable state in S3, EFS or a managed database, and treat the instance as disposable. Take EBS snapshots when a volume must be backed up.