Developers tend to work with multiple environments; production, staging, UAT, regression testing and so forth. Ideally there should be a Continuous Delivery pipeline setup to move containers and configurations from
dev -> test -> UAT -> staging -> production. In an ideal world we'd reuse as much as possible (container images and kubernetes manifests between environments.
So service discovery takes care of finding the right database, message brokers, remote caches and other services for the environment (namespace) a pod is provisioned in.
If you really have to, you can change the kubernetes JSON / YAML on a per environment bases to pass in different environment variables (or even config files mounted into volumes in pods); but the more things are the same between environments, the more likely you really are testing the real stuff in test before production. So if possible; just use service discovery in each environment and leave everything else the same.
You can always use a different, say, database pod in testing containing test data; but ideally you’d reuse the same JSON/YAML and images for your application code.
Incidentally we have an Arquillian integration testing framework for running complete environments as a test case; it spins up a new namespace, creates however many services/pods/replication controllers, uses Arquillian / JUnit to assert things startup and then performs the actual tests (e.g. querying MBeans and stuff) and then tears everything down at the end.
e.g. here’s an example integration test that runs a message broker, producer and consumer and then asserts that the ActiveMQ consumer starts up properly and the ActiveMQ queues enqueue/dequeue correctly. Pretty neat for about a page of code!
So we'd recommend you to try to keep the JSON/YAML the same between environments. One thing that does tend to change between environments is scale. Ideally things would automatically scale; so you use modest sizes in Replication Controllers (e.g. replica size 1); then use per environment auto-scaling to scale things up based on load.
Then your JSON/YAML and docker images can work in all environments without having to change them per environment and folks can have a completely automated Continuous Delivery mechanism setup between however many environments folks need; safe in the knowledge you are testing the same metadata, configuration and images in each environment.
If you do find you need to add some kind of per-environment configuration change, the other option is to put the JSON/YAML in source control and have a branch per environment so you can make environment-specific changes when you absolutely must do (and can then easily diff between environments)