We're very glad to see containerd rapidly grow to this big milestone. Alibaba Cloud started to use containerd actively since its first day, and thanks to the simplicity and robustness emphasise, make it a perfect container engine running in our Serverless Kubernetes product, which has high qualification on performance and stability. No doubt, containerd will be a core engine of container era, and continue to driving innovation forward.

— Xinwei, Staff Engineer in Alibaba Cloud

Architecture Improvements

The Kubernetes containerd integration architecture has evolved twice. Each evolution has made the stack more stable and efficient.

Containerd 1.0 - CRI-Containerd (end of life)

Containerd 1.1 - CRI Plugin (current)

In containerd 1.1, the cri-containerd daemon is now refactored to be a containerd CRI plugin. The CRI plugin is built into containerd 1.1, and enabled by default. Unlike cri-containerd, the CRI plugin interacts with containerd through direct function calls. This new architecture makes the integration more stable and efficient, and eliminates another grpc hop in the stack. Users can now use Kubernetes with containerd 1.1 directly. The cri-containerd daemon is no longer needed.

Performance

Improving performance was one of the major focus items for the containerd 1.1 release. Performance was optimized in terms of pod startup latency and daemon resource usage.

The following results are a comparison between containerd 1.1 and Docker 18.03 CE. The containerd 1.1 integration uses the CRI plugin built into containerd; and the Docker 18.03 CE integration uses the dockershim.

Pod Startup Latency

The "105 pod batch startup benchmark" results show that the containerd 1.1 integration has lower pod startup latency than Docker 18.03 CE integration with dockershim (lower is better).

CPU and Memory

At the steady state, with 105 pods, the containerd 1.1 integration consumes less CPU and memory overall compared to Docker 18.03 CE integration with dockershim. The results vary with the number of pods running on the node, 105 is chosen because it is the current default for the maximum number of user pods per node.

As shown in the figures below, compared to Docker 18.03 CE integration with dockershim, the containerd 1.1 integration has 30.89% lower kubelet cpu usage, 68.13% lower container runtime cpu usage, 11.30% lower kubelet resident set size (RSS) memory usage, 12.78% lower container runtime RSS memory usage.

crictl

Container runtime command-line interface (CLI) is a useful tool for system and application troubleshooting. When using Docker as the container runtime for Kubernetes, system administrators sometimes login to the Kubernetes node to run Docker commands for collecting system and/or application information. For example, one may use docker ps and docker inspect to check application process status, docker images to list images on the node, and docker info to identify container runtime configuration, etc.

For containerd and all other CRI-compatible container runtimes, e.g. dockershim, we recommend using crictl as a replacement CLI over the Docker CLI for troubleshooting pods, containers, and container images on Kubernetes nodes.

The scope of crictl is limited to troubleshooting, it is not a replacement to docker or kubectl. Docker's CLI provides a rich set of commands, making it a very useful development tool. But it is not the best fit for troubleshooting on Kubernetes nodes. Some Docker commands are not useful to Kubernetes, such as docker network and docker build; and some may even break the system, such as docker rename. crictl provides just enough commands for node troubleshooting, which is arguably safer to use on production nodes.

Kubernetes Oriented

As another example, crictl pods includes a --namespace option for filtering pods by the namespaces specified in Kubernetes.

For more details about how to use crictl with containerd:

What about Docker Engine?

"Does switching to containerd mean I can't use Docker Engine anymore?" We hear this question a lot, the short answer is NO.

Since containerd is being used by both Kubelet and Docker Engine, this means users who choose the containerd integration will not just get new Kubernetes features, performance, and stability improvements, they will also have the option of keeping Docker Engine around for other use cases.