mirror of https://github.com/k3s-io/k3s
54 lines
3.3 KiB
Markdown
54 lines
3.3 KiB
Markdown
# Package cri-dockerd to replace dockershim
|
||
|
||
Date: 2022-07-29
|
||
|
||
## Status
|
||
|
||
Accepted
|
||
|
||
## Context
|
||
|
||
Upstream Kubernetes dropped dockershim from the kubelet in the 1.24 branch: https://kubernetes.io/blog/2022/02/17/dockershim-faq/
|
||
|
||
This means that the docker container runtime is no longer directly supported by the Kubelet; continuing to use it
|
||
requires installing cri-dockerd to translate the CRI API to the Docker API. After some internal discussion, it was
|
||
decided that we did not wish to include dockershim, in favor of requiring users to deploy cri-dockerd themselves.
|
||
|
||
```
|
||
<BD> what’s our roadmap for dockershim/cri-dockerd migration? Kubernetes 1.24 finally drops dockershim.
|
||
<BD> RKE will clearly need to migrate over to cri-dockerd, but for products like K3s where docker is supported but not the default, do we want to keep it around? Seems to work OK in K3s with some slight modifications.
|
||
<CJ> If we don't do this work, can a user configure k3s manually to use the shim?
|
||
<CJ> I'd rather reduce the surface area if possible
|
||
<BD> the work is already done, the question is do we want to include it lol
|
||
<CJ> Understood, the second half of the question is the part I care about
|
||
<BD> kk
|
||
<CJ> Can they manually get there without it
|
||
<CJ> If they can, then I don't want to include it. Less surface area for bug fixes and CVEs. And i don't see a need to make it easy for users to user docker with k3s.
|
||
<BD> users should be able to install and start cri-dockerd and then run k3s agent --container-runtime-endpoint=unix:///var/run/cri-dockerd.sock and it’ll work.
|
||
<BD> But we would need to drop the --docker flag since that explicitly uses the kubelet’s dockershim
|
||
<BD> Well, we historically did make it easy. If we want to stop, that’s fine.
|
||
<CK> Isn’t K3s supposed to be “lightweight” anyway? Who’s to say we didn’t put it on a diet for 1.24 and made it get smaller (especially cause the containerd split made it bigger)
|
||
<CJ> Ok, I THINK bill will be ok with this. CW, can you pitch to bill next week (um he might be out too) that we are dropping the docker flag with 1.24. I guess we have a little time, so no rush right?
|
||
<CJ> with the way the ecosystem has evolved, I think we can easily justify the diet
|
||
<CJ> does k3d use/need --docker for anything?
|
||
<CJ> Ask thorsten to be sure but I didn't think so.
|
||
<CJ> But we do need product to agree, fyi
|
||
```
|
||
|
||
The initial releases of 1.24 shipped without docker support; use of the `--docker` flag returns an error indicating that
|
||
the user should install and use cri-dockerd instead. This has been somewhat disruptive to the community; K3s is often
|
||
used in CI or dev environments where it is useful to use docker as the container runtime, but users are not equipped to
|
||
install and manage cri-dockerd due to its relative inaccessibility compared to docker itself. Ref:
|
||
https://github.com/k3s-io/k3s/issues/5741
|
||
|
||
## Decision
|
||
|
||
* We will embed cri-dockerd, and start it when the `--docker` flag is used. This is a drop-in replacement for K3s's
|
||
previous behavior when the Kubelet included dockershim. This meets user expectations around K3s support for the Docker
|
||
container runtime, and eases user adoption of Kubernetes 1.24.
|
||
|
||
## Consequences
|
||
|
||
* The size of our self-extracting binary and Docker images increase by several megabytes.
|
||
* We take on the support burden of keeping cri-dockerd up to date, and supporting docker as a container runtime.
|