Because running it as docker compose would imply running it on a VM. Running production stuff on a VM is not acceptable since we don't have the capacity (nor desire) to keeps VMs patched, manage redundancy (while patching and rebooting), manage secure remote access, etc.
Hence we only deploy to native managed Azure services (App service, ACA, etc.).

Plus following the principle of dependency injection I'd rather make external service dependencies explicit. For example we already know how to run redis instances, manage them, monitor them, have appropriate policies in place etc. so hiding those services somewhere in a docker-compose deployment is just not beneficial at all.

I'm aware that ClearML also comes with a helm chart to deploy on k8s but that also tries to spin up it's own dependent backend services and as such doesn't solve the problem.