A kernel may or may not exist in a separate address space depending on hardware and software features (e.g. the arm split paging regime or the meltdown mitigations). It's just that the kernel memory is not accessible to userspace code.
You can run multiple VMs. That's probably easier to use too, since you tooling for VMs is based around spinning up more machines, not processes.
> not if you trust your kernel code more than the code of the service you run on top of it
I don't really see how that applies. The application code is still the same. It's just that the OS doesn't exist any more. If it doesn't exist, it can't have bugs.
> and not if you want the ability to look inside your containers (e.g. by ssh-ing into it)
True. But I can't easily ssh inside a process either, which is what a unikernel is more comparable to. The only reason I ssh into my VMs is because the application process is wrapped up in an OS. If that OS did not exist, there would be less reason to peek into it.
If you run five containers on a host, you're not running five kernels. You're running five instances of a Linux userspace/application, sharing the same kernel, but isolated from each other.
In a unikernel, that overhead is removed by having only a single address space. You can see it as taking the full source of the kernel and the database server, linking it into a single executable, and stripping everything you don’t need, including everything the kernel would need if it supported multiple processes.
Is that a good idea? Not if the service you’re running wants to be multi-process, not if you trust your kernel code more than the code of the service you run on top of it, and not if you want the ability to look inside your containers (e.g. by ssh-ing into it)
What you do get in exchange is smaller containers (does anybody know why the image in this example is so enormous? I expected something that easily fits in 64 kB), a smaller attack surface, and faster boot times.
(See also https://en.m.wikipedia.org/wiki/Unikernel)