quick breakdown of containers vs gvisor vs kata vs full vms (took me a while to get this straight)
retained text
kept mixing these up so i wrote it down, figured it might help someone. from weakest to strongest isolation: * containers (runc/crun): namespaces for isolation, cgroups for resource limits. everything shares the host kernel * gvisor (runsc): a userspace kernel sits in between and handles most syscalls itself, so only a small subset ever reaches the host kernel * kata containers: each container (or pod) runs in a micro-vm with its own kernel, via kvm. you still use it like a normal container * full vms (qemu/kvm): a whole os, windows included, with its own disk, init, ssh etc the thing that clicked for me: kata and a full vm give you basicaly the same isolation, both are real vms on kvm. the difference is what's inside and how you use it. kata hides a tiny vm behind a container workflow, with a full vm the vm is the workload for kata you can pick the hypervisor: * firecracker: fastest boot, smallest footprint, but minimal devices (no gpu passthrough, at least for now) * cloud hypervisor: good middle ground * qemu: most complete, also the heaviest on kubernetes, gvisor and kata plug in through a runtimeclass, and full vms go through kubevirt. also worth knowing kubevirt isnt a hypervisor itself, it's a management layer on top of qemu/kvm, same way k8s manages containers without being a container runtime outside k8s all of these work fine too: docker/podman/containerd can use runsc or kata as runtimes, and full vms are managed with libvirt or proxmox hope this saves someone a few hours of reading docs or asking an llm
document fetched 2026-10-29 00:39 · canonical https://reddit.com/r/kubernetes/comments/1xho98v/quick_breakdown_of_containers_vs_gvisor_vs_kata