In the beginning

There was chaos

In the beginning

• Applications deployed on physical servers
• Managed by circus
• Packages installed in virtualenv
• Under a single user

Pros

• Simple implementation and deployment

Cons

• Package versioning hell
• No failover

A wild blue whale appeared

The promise of a brighter future

• Unified environment
• Simple deployment
• Simple migrations
• Faster Continuous Integration (CI)
• Atomic releases

Migration challenges

• Image storage - GitLab registry
• Image caching
• Dedicated building environment

  
  build_job:
  	script:
  		- docker build -t $CI_IMAGE_NAME .
  

  
  build_job:
  	script:
  		- 'docker run -d -p 9000:9000 -v "/:/data" sleep 1y'
  

• CI pipeline design
• Cleaning up old images

Migration highlights

• Unified, stable environment
• Fast builds
• Isolated environments
• Faster CI pipeline

Cons of plain Docker

• Known bugs
• No failover
• dockerd is a single point of failure

Docker gotchas

• PID 1 pitfall
• User permissions within containers

Ahoy!

Warmly welcomed features

• Failover when a server fails
• Configuration stored in namespaces
• Service discovery

  my-service.my-namespace.svc.cluster.local

• Ingress controller
• Deployment history

  $ kubectl rollout undo deployment my-app

Where we are now

• Migration to Kubernetes is in progress
• Environment is configured by namespace variables
• Deployments are described in Jinja2 templates

  
  {% set env_type, profile_name = PROFILE.split('_', 1) %}
  {% set namespace = "my-ns-" + env_type %}
  {% set data_directory_mount = '/data' %}
  
  {%  if profile_name == 'cthulu' %}
          - name: API_KEY
            valueFrom:
              secretKeyRef:
                name: secret
                key: api_key
          - name: SUBSCRIPTION_INSTRUMENT_FILTER
            value: 'False'
  {% endif %}

Notable features

• Probes
• Update strategies

Update strategies - Rolling update

Update strategies - Recreate