Continuous Integration (or Continuous Deployment) concept is always amaze me, because it allows an organization to do more iteration and shipping incremental changes without breaking any part of their system. In this story I want to explain how we moving from “pull-based” deployment to complete CI/CD architecture in Backstreet Academy.
You might say “Wait, what is ‘pull-based’ deployment?”. Well, the term “pull-based deployment” is a term that I invented myself 😆. That deployment architecture is pretty simple. You just need install git on the server, and setup a cron job to trigger
git pull every minute to download the latest code base that has been pushed to git repository.
If there is any test that we need to run before the deployment, I must run the test locally in my machine. While doing the test, I cannot do anything else. It’s very unproductive for me. Once the test is success, then I push the changes to git repository, and the server will pull the changes and apply it to the code base that sit in the server.
I can say that is pretty dumb architecture. But hey, as long as it works, it should be fine right? Well, I want to eliminate my time just to run the test on my local machine and waiting until the test done. We have more than 1500 test on our current site which consist both unit test and integration test, and it takes almost 30 minutes for all the test to complete in my local machine.
So, I had an idea to implement CI/CD for our deployment system and since we are very small team engineering team, we had to make sure the system that we will implement should not require to much time to setup and less maintenance work.
Some ideas came up, either we are using third party CI service (like CircleCI or Travis CI) or self-hosted one (like Gitlab Runner, or Jenkins). When we looked up for third party service, the cost is little bit expensive than self-hosted CI service. So we decided to setup our own CI service using Gitlab Runner. We chose Gitlab Runner because in our team we are using Gitlab for our code repository and the integration with Gitlab Runner is seamless.
The flow is pretty much like this:
- Push code changes to repository.
- Gitlab create new pipeline and notify the Master Runner. (AWS EC2).
- Master Runner launch new Gitlab Runner server or use the existing one. (AWS EC2).
- Gitlab Runner run test.
- Test done and start Docker build process.
- Build process done, store Docker image to AWS Elastic Container Registry (ECR) and start deployment process.
- Deployment process will replace old container that exist on AWS Elastic Container Service (ECS) and replace with the new one.
The process above takes 15–20 minutes for most of our services.
So, how we do that?
We are using Gitlab services for our code repository and it come with Gitlab Runner feature on it. But the feature will not work unless you have the runner itself. Usually if you using Gitlab you can use either their own shared runner or our own runner.
For us, we are using our own runner because we want better performance for our CI process since we can configure the runner resources (CPU and memory) as we like. Also we want to make sure secret credentials stays on our AWS environment.
We are using Autoscale configuration on our Gitlab runner which you can see more about it here. In simple terms, Autoscale require one master server that have a task to launch new runner if there is a new process need to be handled. In our Autoscale configuration, we choose Docker as our runner options.
Autoscale also allow us to be more efficient in terms of cost, since it only launch new runner server only when it needed. We also set the timeout of the runner really short around 5 minutes which means if there is no more jobs after 5 minutes idle the server will shutdown automatically.
This configuration has a drawback though. It will increase cold-start time when there is a job available and there is no active runner server. However, We have very small engineering team and we currently don’t mind with this because our deployment rate is still small. Since we implemented this architecture 2 months ago, we only have made around 700 deployment process (~11 deployments/day).
Next is container registry. After Docker build process done, it will produce an image. We push the image to AWS Elastic Container Registry (ECR) then deploy it to AWS Elastic Container Service (ECS). The setup of ECR and ECS is pretty simple since they both are AWS services. On ECS we want simplify things, that’s why we choose Fargate instead EC2 as our container cluster. It’s higher cost but lower maintenance work required.
In Gitlab, you need configuration called
gitlab-ci.yml This file is use by Gitlab to configure each step of the Continuous Integration (CI) process. In our setup we have 3 step: test, build, and deploy. So we have configuration like this:
- ecs deploy ClusterName ServiceName --timeout=1800
Code above is the simple version of our
gitlab-ci.yml and the file usually have more configuration which I will not explain deeply since you can read the documentation by yourself. In our code-base usually we have 200–300 LoC of
Let’s break it down.
Test process is covering our unit test and integration test. Currently we don’t have E2E test but I really want to have it in the future.
As you can see in the code above, we have specific Docker image that has built as test runner. We store it on Gitlab Registry since it contains some of our code that should not available publicly. We also have test script that will run the test once the container is launched.
Build process cover the process building our code into proper Docker image.
In our repository, we have Dockerfile that will help us generate the image. But in order to have docker inside the runner, we need activate Docker-in-Docker service. This service allow us to have
docker binary available to build the image. Without this the command will not exists.
We are using image called docker-aws. This image is specifically build for our needs to have build and deployment process in our CI infrastructure. It contains toolkit such as ecs-deploy and AWS command line.
In our build script, we have something like this:
#!/bin/shecho "Login to ECR Repository"
$(aws ecr get-login --no-include-email --region $AWS_DEFAULT_REGION)echo "Preparation task"echo "Build Docker Image"
docker build -t sample-image .echo "Push to ECR Repository"
docker tag sample-image:latest $AWS_ECR_REPOSITORY:sample-tag
docker push $AWS_ECR_REPOSITORY:sample-tag
In order to have this script running we need to have environment variable implemented on our CI/CD configuration on Gitlab repository which we set to be available only on protected branches.
It means we only trigger build process for our feature, staging, and production branch. So, usually we have more than two build process configured in our
gitlab-ci.yml. Once the build process completed and image is stored in ECR, we continue to deployment process.
Deploy process cover the deployment of the image that has been generated in build process. The task will launch new image, and shutdown the old image.
- ecs deploy ClusterName ServiceName --timeout=1800
This stage still using docker-aws image and the script contains only deployment script to specific cluster and specific service. In our configuration we set the timeout to 30 minutes, because we usually have couple task running in one service and need to make sure all the network connection drained properly before killing it.
As you can see, this article is not very specific to teach you “how to implement Continuous Integration”, instead it’s more the overview that may help you to setup your own infrastructure. Because I believe every organization has their own complexity that need to be solve differently, so the specific things like preparing credentials, building binary or testing the code need to be adapted to your code.
Currently our setup is not integrated well with our project management tool, since I currently still working to make more automated. So in the future I might be will share more about this integration.