Studies about ECS deployment types using Terraform

Index

• 1. Deployment types
  • 1.1. Rolling Update
  • 1.2. Blue Green
• 2. How it really works?

1. Deployment types

There are three types of deployment, docs:

• External: it uses a third-party deployment controller, out of scope; docs
• Rolling Update: it uses ECS itself; docs
• Blue Green: it uses Code Deploy; docs

1.1. Rolling Update

With Rolling Update, ECS will replace running tasks with new tasks and it can be controlled with these parameters:

• minimumHealthyPercent: the lower bound of instances that must be running during deployment, rounded up;
• maximumPercent: the upper bound of instances that should be running during deployment, rounded down;

If minimumHealthyPercent is 50%, maximumPercent is 150% and there are 5 instances running, 3 (5 * 0.5, rounded up) instances is the lower bound and 7 (5 * 1.5, rounded down) instances is the upper bound. I did the experiment and these are the results:

Old Instances Running	Old Instances Deprovisioning	New Instances Running	New Instances Provisioning
5	0	-	-
3	2	-	2
1	2	2	2
0	1	4	1
0	0	5	0

If minimumHealthyPercent and maximumPercent combinations makes deployment impossible, it will throw an error!

1.2. Blue Green

With Blue Green, Code Deploy shifts traffic between Blue and Green ELB Target Groups and it can be controlled with these configurations:

• Linear: CodeDeployDefault.ECSLinear10PercentEvery1Minutes or CodeDeployDefault.ECSLinear10PercentEvery1Minutes;
• Canary: CodeDeployDefault.ECSCanary10percent5Minutes or CodeDeployDefault.ECSCanary10percent15Minutes;
• All at once: CodeDeployDefault.ECSAllAtOnce;

When using NLB, only CodeDeployDefault.ECSAllAtOnce is supported!

There are very good images at AWS Docs explaining that this is the initial state:

And this is the final state:

Basically, Code Deploys creates a new ECS Task Set (a collection os ECS tasks!), attach it to the Target Group 2 and shifts traffic from Target Group 1 to Target Group 2. At this time, there are two ECS Task Sets. After shifting all traffic, it destroys the old ECS Task Set after 5 minutes.

We can add a Test Listener and perform tests using a Lambda to stop deployment on failure using a custom logic:

2. How it really works?

I've created a simple application and a simple platform to study it, so you will find these main directories:

• platform: has Terraform code that deploys shared resources like ECS Cluster, ELB etc needed by on or more apps;
• apps/node-app: has a src directory with the node application and infra directory with Terraform code that deploys ECS Service, Code Deploy etc needed only by the app;
• apps/spring-app: it's the same node app, but written with Spring and Kotlin. I've created it to test if Code Deploy can work with multiple rules inside a ALB Listener. It can!
• .github: has Github Actions code that deploys all these items to an AWS account;

I recommend to see every file because I added comments that shows the mistakes I've made.