What if deployment selector labels are only subset of template(pod) labels? I've read documents about Labels and Selectors, Deployments. Also googled but not found answer.
When using helm3 to create a new chart by helm create x, the default deployment labels are like this(omitted unrelated parts):
spec:
selector:
matchLabels:
{{- include "x.selectorLabels" . | nindent 6 }}
template:
metadata:
labels:
{{- include "x.selectorLabels" . | nindent 8 }}
spec:
which means it is highly recommended, that selector labels are same as pod labels. But what if selector labels are only subset of pod labels?
I did two experiments:
- Define sigle deployment, and helm install to two different releases
spec:
replicas: 2
selector:
matchLabels:
foo: bar
template:
metadata:
labels:
foo: bar
unique: {{- uuidv4 }}
spec:
- Define two deployments in single chart, and helm install
- deployment1
metadata:
name: deployment1
spec:
replicas: 2
selector:
matchLabels:
foo: bar
template:
metadata:
labels:
foo: bar
unique: {{- uuidv4 }}
spec:
- deployment2
metadata:
name: deployment2
spec:
replicas: 2
selector:
matchLabels:
foo: bar
template:
metadata:
labels:
foo: bar
unique: {{- uuidv4 }}
spec:
My expectations is that since deployment select pods by labels, and not anything else, I think in each case final pods count should be 2, because the pods created by the two deployments share same labels.
Actual result is not. Totally 4 pods created. Seems deployment "knows" pods reference other than labels. And uninstalling one of the releases in case 1 only removed two pods, which proves the idea again.
I checked replicasets, there's a label called pod-template-hash which surely is unique and might explain. But if pod-template-hash is really implicitely used, why do we need explicitely define label selectors that aren't acutally used?
