The platform is ensuring zero downtime deployments: when a new version of your application is deployed, the system is waiting for the new version of the container to be online, before stopping the old version.
This is the same principle when you’re restarting one or several containers of your application, new containers have to be available at first, then the order to stop old containers is done.
With this feature, it let you achieve rolling release/continuous delivery easily, there is no friction when a new version of your software is deployed, your users won’t even notice it.
Starting new containers
When a new
tcp container is started, its environment contains the
environment variable. Your application must listen on this port in order to
receive requests. Once the application has been started, the container scheduler
will try to connect to your application on the port defined previously during
60 seconds. Once a TCP connection is accepted by your container, we consider
the container ready to get requests. Otherwise, the operation is aborted with a
Boot Timeout Error:
In the case of a deployment, it will be stopped with the status ‘timeout-error’ and you’ll have to fix why your application has not been listening in time on the corresponding port.
In a restart operation, we’ll retry to start the container 20 times with exponential backoff, then you’ll receive a notification email with a notice telling you that the scheduler didn’t manage to restart your container.
Once the application is ready, the traffic is diverted to the newly booted container(s), and the order to stop old containers is passed.
When deploying an application using multiple web containers, we’re waiting for all web containers to be available before making the switch to this new release and stopping old containers.
Short period with double containers
For a short period both old and new containers are running. If your application is based on schedule tasks, or asynchronous workers, you need to take care of developing them idempotent in the potential case where a scheduled task is handled twice.
Shutdown of old containers
We have adopted a standard process:
- Send SIGTERM to the process started in the container
- Wait 30 seconds
- Send SIGKILL if the process is still alive
In other words, your application has 30 seconds to catch the SIGTERM signal and clean its state before we force it to stop.