Skip to content

Kata Containers

Kata Containers#

  • Virtual machines - bigger, slower with more isolation. Boot time of minutes.
  • Containers - quick, simple but less isolation

kata containers is a blending of the two. Security and isolation of virtual machines without the bloat of virtual machines.

It is a vm

How do you use Kata Containers#

It looks like runc an OCI compatible runtime

runC is the universal container runtime

History#

Before kata there was Intel Clear Containers

Traditional Containers#

You have a host and a kernel. Your containers run within namespaces on that host. All good until somebody breaks out of a container - which you have access to all containers.

Traditional containers in a VM#

A Vm within a VM with containers in it

Kata has lightweight virtual machines

Kata Containers#

  • Every Container / Pod gets it’s own virtual machine
  • Container doesn’t know it is a vm - same deployment process

Architeture and Integrations#

Components:

  • QEMU - KVM
  • Runtime (Kata Runtime)
  • Kernel - to sit inside VM
  • rootfs image - vm has to boot something (rootfs sets up the container)
  • Agent - work to happen: mount points, networks, memory resources and cgroups
  • Shim
  • Proxy

CRI-O and Kata#

Kubelet -> CRI (Container Runtime Interface) -> CRI-o/ContainerD -> runc -> kata-runtime -> vm

Can choose which you want - if you don’t trust it run it in a kata container

Fairly seemless

Networking and Storage#

Containers run at layer 3 but vm’s run at layer 2

Storage:

  • 9pfs (overlay) - easy to use network based file system with plan9 (just works) - default. Not a full POSIX unix filesystem.
  • Block devices (device mapper) - vm can find block device , map and mount to virtual machine (not going over network connection)
  • Network (ceph, gluster) - Network storage works as you expect

Overhead#

  • 50MB per container
  • Boot in less than a second

Road Map#

Primarily around security and isolation

Security, in container: seccomp Security, on host: cgroup isolation, more namespace isolation, root-less QEMU and SELinux policy

Source#