This is a Proof of Concept of how to run a local container registry in Kind (Kubernetes in Docker).
When developing locally, you want to push images to the cluster as fast as possible.
Pushing to an in-cluster container registry skips a lot of overhead:
-
Unlike with a remote registry, the image stays local to your machine, with no network traffic
-
Unlike with
kind load, docker will skip pushing any layers that already exist in the registry
This makes it a great solution for iterative local development. But setting it up is awkward and fiddly.
This repo demonstrates how you might build a more robust solution.
- Create a cluster
kind create cluster
- Start the registry. Currently it creates the registry at port 32001
./kind-registry.sh
- Try pushing an image.
docker tag alpine localhost:32001/alpine
docker push localhost:32001/alpine
You can now use the image name localhost:32001/alpine in any resources you deploy to the Kind cluster
kind-registry.sh has three major jobs:
- Deploy a container registry in the cluster that's available on each node at 32001.
Currently we do this with registry.yaml.
This runs a single pod and exposes it with a NodePort, which means it only works if your Kind cluster is a single-node cluster.
TODO: make this work with a multi-node cluster.
- Configure containerd to trust the registry at 32001, even though it's on http (instead of https).
This is trickier. See config.toml, which configures this as
[plugins.cri.registry.mirrors."local.insecure-registry.io"].
- Expose the registry as 32001 on the host.
We do this with kubectl port-forward. A more robust implementation would have
to manage the port-forwarding.
Inspired by MicroK8s's private registry feature.
Much of the code in this repo is inspired by the equivalent registry setup in MicroK8s.
The KIND private registries doc had some helpful sample code in how to set up registry configuration.
We're hoping that all local Kubernetes tooling will eventually support this workflow natively. See this issue in the Kind repo for more technical discussion.
Copyright 2019 Windmill Engineering
Licensed under the Apache License, Version 2.0