The kubelet is the primary "node agent" that runs on each node and provides you a lot of configuration options, but sometimes it still not enough, and you still want to run an additional configuration on the node before starting or stopping the container.

For example, when you want to get isolated CPUs, but you do not want to exclude CPUs from the load balancing forever, you can disable CPU load balancing by demand for pinned CPUs via sched domains. Once the container will start, our code will fetch the container pinned CPUs and will disable the CPU load balancing for specific CPUs via sched domain, and once container removed, our code will restore back the CPU load balancing.

For such use cases, the CRI can help, it provides different mechanisms to inject the desired configuration via a custom script once a container starts or stops.

Important to notice that currently, we have two main CRI implementations, containerd the default CRI implementation used by the vanilla Kubernetes and CRI-O used by default under the OpenShift. Depends on the specific implementation, availability and the configuration can be different.

OCI hooks

OCI hooks mechanism defines several entry points to inject your code. To be more specific runtime-spec version 1.0.0 supports prestart, poststart, and poststop entry points.

Pros

  • have a number of entrypoints
  • have a filtering mechanism that will prevent to run the hook code on every container
  • easy to configure

Cons

  • not supported by containerd
  • impossible to prevent use of hooks via RBAC

Configuration

Because currently, only CRI-O supports it, I will concentrate on the CRI-O configuration.

To enable hooks under the CRI-O you should specify hooks directory under the /etc/crio/crio.conf and add the hook JSON specification under hooks directory.


NOTE: Default search paths for hooks are /etc/containers/oci/hooks.d/ and /usr/share/containers/oci/hooks.d/

Example:

# cat  /etc/crio/crio.conf
...
hooks_dir=[]
...
# cat /etc/containers/oci/hooks.d/test-hook.json
{
"version": "1.0.0",
"hook": {
"path": "/usr/libexec/oci/hooks.d/oci-systemd-hook" // path to the the hook binary
},
"when": {
"commands": [".*/init$" , ".*/systemd$"] // run this hook only when the cmd of the container ends with the init or systemd
},
"stages": ["prestart", "poststop"] // run this hook before the container starts and after the container stops
}

You can get the container state under the hook script from the stdin, you should get the JSON format input that will allow to parse it and fetch all relevant information regarding the container state.

You can check the link for a better explanation about the hook schema - https://www.mankier.com/5/oci-hooks.

Runtime wrapper

The second option is to create a wrapper around runc runtime handler that used by containerd and CRI-O.

Pros

  • provide more flexible approach to run the code on every life phase of the container

Cons

  • requires to provide a wrapper around the runc
  • more complex configuration
  • impossible to prevent use of RuntimeClass via RBAC

Configuration

Under the node, you should specify an additional runtime handler under the CRI implementation configuration.

CRI-O

# cat /etc/crio/crio.conf
...
[crio.runtime.runtimes.wrapper]
runtime_path=/usr/bin/wrapper
...

Containerd

# cat /etc/containerd/config.toml
...
[plugins.cri.containerd.runtimes.wrapper]
runtime_type = "io.containerd.runc.v1"
pod_annotations = ["*"] // run for every pod
container_annotations = ["*"] // run for every container

[plugins.cri.containerd.runtimes.wrapper.options]
BinaryName="/usr/bin/wrapper"
...

The wrapper script runs each time when the CRI implementation calls to the runtime. It is important to say that it can be any executable file, but it should call the runc binary with all relevant parameters before the exit.

Example:

#!/bin/bash
echo "$@"
exec /usr/bin/runc "$@"

Under the cluster, you should create a new RuntimeClass resource, that will use the new runtime handler.

Example:

apiVersion: node.k8s.io/v1beta1
kind: RuntimeClass
metadata:
name: wrapper # The name the RuntimeClass will be referenced by
handler: wrapper # The name of the corresponding CRI configuration
scheduling:
node-selector: [...] # The node selector defines where will run pods that are using this runtimeclass

In the end you should specify the runtime class under the pod specification.

Example:

apiVersion: v1
kind: Pod
metadata:
name: mypod
spec:
runtimeClassName: wrapper
...

Now when you will start the pod, the CRI implementation will use our wrapper as a runtime handler and execute our script code.

Conclusion

As you can see we have some good options that allow us to run any additional operating system configurations before the container start or on any other life phase of the container. But you should be aware that currently it impossible to limit usage of hooks or runtime classes via standard Kubernetes mechanisms, so do not overuse them.