Running in Kubernetes

Yes, because why not, especially if you run k3s :D

This is just an example but you should be able to adjust it to your needs.

See directory for the content you can use with a kustomize.

Overview

each camera should be a separate kubernetes deployment, easier to manage.
in configs there is a one file with env vars loaded per deployment, those env vars MUST BE changed as in env vars Also do not use double quotes the values.
deployment possible using for example kustomize
camera device - if you have more cameras you probably want to use device by-id or by-path, see tuning for more details

deployment-1.yaml is an example to fetch image from a stream using ffmpeg and with custom prusa-connect-camera.sh for easier development/iteration. This is also preferred option when you are not using cameras that are directly attached to the hosts, but rely on using ffmpeg to fetch images from remote streams or static images via curl.
deployment-2.yaml is an example how to run it on Raspberry Pi with USB camera using default parameters. You want to change .spec.nodeName and volumes to point to desired camera.
daemonset.yaml is an example how to run it on Raspberry Pi with USB camera using default parameters. You want to change .spec.nodeName and volumes to point to desired camera. The difference between the DaemonSet and Deployment is that with the Deployment kubernetes will try to spawn new pods even if node is not available. DaemonSet it is better when you have to use directly attached devices to the hosts. Of course in that case you will need a daemonset per camera per host, and separate config for each camera/host, and you would need the same with a deployment, but at least when node is gone you are not getting non-schedulable pod every few minutes. With a daemonset when node is gone, then pod is gone and k8s is not trying to spawn new pods.

If you want to add more cameras then you should:

copy example config file from config/ and adjust it
copy desired deployment.yaml and adjust it
update any cam-x references in the deployment.yaml
update .spec.nodeName in the deployment.yaml
if using direct camera update volumes - dev-video hostpath to point to the desired camera device
fine tune requests/limits to use less resources if needed

just look into the logs of the failing pods
some hosts do not have predictable camera names, will have to think about udev rules how to handle it...
not tested with any device operators
not tested with istio/traefik whatever.

Last update: 2024-10-07 22:58:35 +0200 (UTC)