Confusing Terms In Container Ecosystem

Containers are great, in fact, software components which enable them are some of my favorite projects. But the more you try to know about how they work, the more terminology gets confusing. Docker, Docker Engine, containerd, runC, rkt, cri-o – all describing themselves are container runtime, are they alternatives or complementary components? Based on my understanding while trying to contribute to one of these projects, I’ve summarized these terms below. This assumes you’ve used containers before. If you have any suggestions, please let me know.

Going bottom-up, Open Container Initiative or OCI is a project by Linux Foundation which standardizes the container runtime, image and distribution specifications, these specifications are released as runtime-spec, image-spec and distribution-spec respectively.

runC is one of the many implementations of runtime-spec, other implementations can be found here. Now given the specs, these implementations have very limited and defined scope in container world i.e. runc can create, start and delete a container. Generally, a more comprehensive runtime will be required in real systems and can be implemented on top of runC. This runtime would be responsible for managing multiple containers at a time and things like downloading container images, managing storage and network interfaces, etc. Widely used example of this runtime is containerd which can manage multiple containers using runC or any other OCI implementation. Check this post by Michael Crosby to know more about how containerd integrates runC.

Another runtime is cri-o by Kubernetes which is OCI implementation with just enough functionality that is required by Kubernetes CRI. CRI or Container Runtime Interface is a Kubernetes interface for using any runtime with its container manager called kubelet. That means, containerd and cri-o are alternative runtimes for using with kubelet.

Now, kubelet and Docker Engine are a higher level abstractions that use and depend on above-mentioned runtimes. Docker Engine uses containerd and adds things like networking, volumes and security to the containers.

Following image shows how these pieces fit together visually.

Container Ecosystem Source: https://blog.docker.com/2017/08/what-is-containerd-runtime/

Docker, the company, also provides an end-user CLI client docker to interact with Docker Engine and what we use directly on our local systems to manage containers. Don’t be confused if Docker, containerd and runC all are described as runtimes, they are classified as runtimes but with different and very defined scopes.

Another interesting project is the Moby Project, which you might have come across if you’ve been reading about these components. So, earlier Docker used to be a huge piece of software by Docker, the company. With time, they separated out the major components like containerd which can be used and developed independently. All these components that Docker now uses as upstream along with some tools and framework to assemble them, comes under this Moby Project. This introductory post talks about the same in detail.

Another parallel container project is called rkt, which implements a whole different approach to manage containers and developed by CoreOS. It has a runtime which supports OCI images but as of now, it does not follow exact runtime-spec. Think of rkt as an alternative to Docker Engine along with containerd and runC. rkt also has CRI implementation for using with Kubernetes, called rktlet. Following image captures this comparison along with other details.

rkt vs Docker

Source: https://coreos.com/rkt/docs/latest/rkt-vs-other-projects.html#rkt-vs-docker

To summarize, runC is OCI spec implementation for creating, starting and deleting containers. containerd uses runC and provides a broader runtime for containers. Docker client uses Docker Engine on local systems, which is built on top of containerd. Docker uses Moby Project as upstream which includes all these components. Kubernetes can use containerd, cri-o or rktlet as it’s container runtime. rkt is an alternative to Docker and has all components of its own.

[Edit]

Adding about LXC and libcontainer, they both use kernel features like namespaces and cgroups to provide the virtualization and have different motivations. Docker Engine used to use LXC before they created their own libcontainer. So, runC is mostly just a wrapper around libcontainer.

Reddit discussion link.

PS: Thanks Fakabbir Amin for reading the draft of this post.