When does kubernetes helm trigger a pod recreate?

The helm documentation suggests to recreate a pod by setting variable metadata values.

For example:

kind: Deployment
spec:
  template:
    metadata:
      annotations:
        checksum/config: {{ include (print $.Template.BasePath "/configmap.yaml") . | sha256sum }}
[...]

But there a situations, when a pod is not recreated:

A pod is erroneous in state CrashLoopBackOff
Only Deployment Metadata has changed

I would like to know what events do trigger a pod recreate:

Why is the pod in state CrashLoopBackOff not restarted?
Why are not all parts of the spec considered to recreate the pod?

Edit

The CrashLookBackOff is an application problem. But if a new image (containing the bugfix) is provided, the pod should be restarted without the need to kill it explicitly.

Is there a cause not to restart the CrashLookBackOff pod?

Solution

The template in a Deployment is a PodTemplate. Every time the PodTemplate is changed, a new ReplicaSet is created, and it creates new Pods according to the number of replicas using the PodTemplate.

kind: Deployment
spec:
  template:
    # any change here will lead to new Pods

Everytime a new Pod is created from a template, it will be identical as the previous pods.

A CrashLoopBackOff is a Pod-level problem, e.g. it may be a problem with the application.

But if a new image (containing the bugfix) is provided, the pod should be restarted without the need to kill it explicitly.

If a new image is provided, it should have its own unique name. That means that whenever you change the image, you have to change the image name. A change of the image name is a change in the PodTemplate, so it will always create new Pods - and delete but not reuse old Pods.