Any API call at any time can introduce a problem, or the system itself can simply get into a bad state, and there is no way to simply return back to the original good state. One has to jump through a lot of hoops to to to figure out how to fix the state, live, like breaking a plate and then quickly trying to glue it back together. When the state of the art's best practice should be to grab another perfect plate from the shelf.
To be sure, some cloud systems so have immutable features. Container and virtual machine technology depends on it. Yet the systems that control those containers and VMs have no immutable properties with which to build and operate the larger system-of-systems.
If you design a system architecture today, you cannot build the cloud infrastructure, get it to a working state, and then make an immutable snapshot of how the system works, and restore it later if the system changes or becomes unstable. Again, you can snapshot a VM, or a container; but you cannot snapshot the system that runs those VMs and containers, nor the system that depends on those, etc. If you make a policy change to an S3 bucket, you can not simply restore the policy (and all other aspects of a bucket) to where it was before. Almost no cloud systems have this property, because they are just a mess of APIs and mutable state.
This all becomes apparent when you operate large scale systems using tools like Terraform. Terraform is a Configuration Management tool, designed to modify the system's mutable state to return it to some desired state. It's the same as using Puppet or Ansible to modify a running VM, to "fix" the VM in place. This is a laborious and error-prone process which requires a great deal of investment in the Configuration Management - essentially writing programs just to tell the system where each file should be.
But the best practice for VMs is not to "fix" its state. It is to simply deploy a snapshot of a VM, with the exact properties of that VM that it should have, and deleting the old VM. This brings a number of desirable results to operations, and completely does away with the need to program a Configuration Management system to "fix" the VM.
All cloud systems should support immutable operations. But virtually none of them do. This is a tremendous disadvantage and gap in the cloud computing ecosystem. It will cause increased instability and require dedicated jobs just to deal with the Configuration Management problem. This creates a problem for developers, too, who will stay away from learning cloud architecture due to the complexity involved.