KNI [Kubernetes Networking Interface] Initial Draft KEP #4477
Conversation
MikeZappa87
commented
Feb 2, 2024
•
MikeZappa87
commented
- One-line PR description: This is the first draft of the KNI KEP, user stories and additions to be discussed as a community
- Issue link: [KNI] Kubernetes Network reImagined (Interface) #4410
- Other comments:
Since I'm going to be helping co-author this KEP I shouldn't be an approver, but I can still be a reviewer.
Signed-off-by: Shane Utt <[email protected]>
Signed-off-by: Shane Utt <[email protected]>
KEP updates
Co-authored-by: Shane Utt <[email protected]>
Update Summary and Goals
chore: cleanup template text and blank space
Another pass at the goals
Signed-off-by: Shane Utt <[email protected]>
Signed-off-by: Shane Utt <[email protected]>
docs: kni add goal for container network namespaces
|
Skipping CI for Draft Pull Request. |
k8s-ci-robot
added
do-not-merge/work-in-progress
This is excellent! This is what I have been wanting :-) |
|
Can you squash the commits? I find it hard to comment lines in the KEP, because I don't know which commit to use? So I include quotes instead below.
Can you give examples? I am pretty well aware of how things work (or so I thought), and my interpretation is:
Is that correct? And what is "kernel isolated pods"? That I don't know. Is it kata containers for instance (VMs)? |
I guess "the pod" refers to a CNI-plugin agent POD, e.g.
And, the files are not "usually downloaded", they are copied from the init-container to the host fs. However, I agree that files in the host filesystem should be avoided. |
|
|
||
| ## Motivation | ||
|
|
||
| Kubernetes networking is an area of complexity and multiple layers which has created several challenges and areas of improvement. These challenges include deployment of the CNI plugins, troubleshooting networking issues and development of new functionality. |
There was a problem hiding this comment.
(linewrap! it's too hard to comment on lines that are too long)
There was a problem hiding this comment.
Yes, please. The current formatting has too many discussion points per line, it's difficult to follow and comment on.
For most cases, the init container downloads the primary plugin aka flannel, cilium however when they leverage the community plugins they are downloaded. This usually happens at a different time though usually container runtime install. containerd install script does this. I want to avoid this and have all dependencies inside of the container image and no longer needing to be in the host file system |
No, they copy their primary plugin and the community plugins they need from their init-container (which usually uses the same image as the "main" one, but with different start-up). The only community plugins that is required on start is "loopback". Examples what cni-plugins install beside their primary plugins:
But none of them downloads anything on installation. I set But in any case, this is misleading:
It should be something like:
|
This adds a section to the KNI KEP for "ongoing considerations" where we can put ideas and concerns that don't need to be resolved just yet, but are good to come back to as we progress. It then adds an idea for using Kubernetes controllers as an alternative to gRPC APIs for some of the KNI implementation to the ongoing considerations. Signed-off-by: Shane Utt <[email protected]>
|
I discovered that you can add sequence diagrams in github with mermaid, and I really like sequence diagrams 😄 I have created some really simplified ones that describe network setup with and without KNI. And, I know that KNI can do more than call a CNI-plugin, so please don't start a discussion on that. It's just an example and it must be there for backward compatibility. I have not worked with a CRI-plugin, so I may have got the interaction with the OCI runtime (runc, crun) all wrong. Current network setupsequenceDiagram
participant api as API server
participant kubelet as Kubelet
participant cri as CRI plugin
participant oci as OCI runtime
participant cni as CNI plugin
api->>kubelet: Notify POD update
kubelet->>cri: Create POD
activate cri
cri->>oci: Create sandbox
activate oci
oci-->>cri: response
deactivate oci
cri->>cni: ADD cmd (exec with files)
activate cni
cni-->>cri: response
deactivate cni
cri->>oci: Create containers
activate oci
oci-->>cri: response
deactivate oci
cri-->>kubelet: response
deactivate cri
kubelet->>api: Update status
Network setup with KNIsequenceDiagram
participant api as API server
participant kubelet as Kubelet
participant cri as CRI plugin
participant oci as OCI runtime
participant kni as KNI agent
participant cni as CNI plugin
api->>kubelet: Notify POD update
kubelet->>cri: Create POD
activate cri
cri->>oci: Create sandbox
activate oci
oci-->>cri: response
deactivate oci
cri-->>kubelet: response
deactivate cri
kubelet->>kni: Setup network (gRPC)
activate kni
kni-->>api: read objects
kni->>cni: ADD cmd (exec w stdin)
activate cni
cni-->>kni: response
deactivate cni
kni-->>kubelet: response
deactivate kni
kubelet->>cri: Create containers
activate cri
cri->>oci: Create containers
activate oci
oci-->>cri: response
deactivate oci
cri-->>kubelet: response
deactivate cri
kubelet->>api: Update status
|
|
|
||
| Currently networking happens in three layers of the stack, Kubernetes itself by means of kube-proxy or another controller based solution, the container runtime with network namespace creation and CNI plugins and the OCI runtime which does additional network setup for kernel isolated pods. All of this communication happens through non network specific APIs which to the reader of the code makes it hard to determine where ‘networking’ is happening. Having networking in several layers presents an issue when needing to troubleshoot issues as one needs to check several areas and some cannot be done via kubectl logs such as the CNI execution logs. This becomes more of an effort as multiple uncoordinated processes are making changes to the same resource, the network namespace of either the root or pod. The KNI aims at reducing the complexity by consolidating the networking into a single layer and having a uniform process for both namespaced and kernel isolated pods through a gRPC API. Leveraging gRPC will allow users the ability to migrate away from the current execution model that the CNI currently leverages. | ||
|
|
||
| The next challenge is the deployment of the CNI plugins that provide the network setup and teardown of the pod. The idiomatic way to deploy a workload in Kubernetes is that everything should be in a pod; however the current approach leaves files in the host filesystem such as CNI binaries and CNI configuration. These files are usually downloaded via the init container of the pod after binding, which increases the time for the pod to get to a running state. Since all existing K8s network plugins are running as daemonsets we will take this approach as well, where all the dependencies are packaged into the container image thus adopting a well known approach. This will have added benefits of the network pod startup being much faster as nothing should need to be downloaded. |
There was a problem hiding this comment.
A key part of what current CNI provides is a framework for different vendors to independently add extensions (CNI plugins) to a node, and have guarantees about when in the lifecycle of pod creation those are invoked, and what context they have from other plugins.
There may be multiple daemonsets from multiple vendors installing multiple CNI plugins on the same node or the node may come from a cloud provider with a CNI already installed and some other vendor might want to chain another extension onto that - any model we adopt must reckon with this as a first-class concern.
That, for me, is critical to retain for this to be a 1-1 replacement to the existing CNI - we can probably do something simpler than the current model of "kubelet talks to CRI, CRI impl reads shared node config with plugin list, serially execs the plugins as arbitrary privileged binaries", as well.
At the very least, moving that "list of extensions + CNI config" to an etcd-managed kube resource would fix currently non-fixable TOCTOU bugs we have in Istio, for instance: istio/istio#46848
At a minimum it's important for for me that any CNI/KNI extension support meets the following basic criteria:
- I am able to validate that KNI is "ready" on a node
- I am able to subscribe or register an extension with KNI from an in-cluster daemonset, and have guarantees TOCTOU errors will not silently unregister my extension.
- I have a guarantee that if the above things are true, my extension will be invoked for every pod creation after that point, and that if my extension (or any other extension) fails during invocation, the pod will not be scheduled on the node.
- I am able to get things like interfaces, netns paths, and assigned IPs for that pod, as context for my invocation.
Ideally as a "this is a good enough replacement" test, I would want to see current CNI plugins from a few different projects implemented as KNI extensions, and installed/managed on the same node. If we can do that, we are effectively proving this can be a well-scoped replacement for the status quo.
I believe flannel requires bridge plugin. I just installed that to make sure I wasn't crazy. From the performance testing I have done, KNI is faster taking a consistent 1 second vs 9-23 seconds for the network pod setup. The pod network setup is faster as well. |
We actually have a backwards compatible model with a roadmap to deploy side by side, then with libkni. However your diagram is pretty much spot on. If you want, I can set up some time with you to go over the migration stories with demos? |
|
|
||
| Kubernetes networking is an area of complexity and multiple layers which has created several challenges and areas of improvement. These challenges include deployment of the CNI plugins, troubleshooting networking issues and development of new functionality. | ||
|
|
||
| Currently networking happens in three layers of the stack, Kubernetes itself by means of kube-proxy or another controller based solution, the container runtime with network namespace creation and CNI plugins and the OCI runtime which does additional network setup for kernel isolated pods. All of this communication happens through non network specific APIs which to the reader of the code makes it hard to determine where ‘networking’ is happening. Having networking in several layers presents an issue when needing to troubleshoot issues as one needs to check several areas and some cannot be done via kubectl logs such as the CNI execution logs. This becomes more of an effort as multiple uncoordinated processes are making changes to the same resource, the network namespace of either the root or pod. The KNI aims at reducing the complexity by consolidating the networking into a single layer and having a uniform process for both namespaced and kernel isolated pods through a gRPC API. Leveraging gRPC will allow users the ability to migrate away from the current execution model that the CNI currently leverages. |
There was a problem hiding this comment.
What is meant by namespaced AND kernel-isolated pods? isn't that the same thing?
There was a problem hiding this comment.
What I was trying to do here is separate the virtualized oci runtimes from the non-virtualized oci runtimes. Aka kata vs runc. However we just got off a call with Kubevirt where you could use either. Both will leverage network namespaces however the virtualized cases have the additional kernel isolation.
There was a problem hiding this comment.
If both will use (at least) network namespacing it's probably less confusing to just say that (and provide more detail later in the doc for other cases that use addl isolation, if need be).
|
|
||
| Currently networking happens in three layers of the stack, Kubernetes itself by means of kube-proxy or another controller based solution, the container runtime with network namespace creation and CNI plugins and the OCI runtime which does additional network setup for kernel isolated pods. All of this communication happens through non network specific APIs which to the reader of the code makes it hard to determine where ‘networking’ is happening. Having networking in several layers presents an issue when needing to troubleshoot issues as one needs to check several areas and some cannot be done via kubectl logs such as the CNI execution logs. This becomes more of an effort as multiple uncoordinated processes are making changes to the same resource, the network namespace of either the root or pod. The KNI aims at reducing the complexity by consolidating the networking into a single layer and having a uniform process for both namespaced and kernel isolated pods through a gRPC API. Leveraging gRPC will allow users the ability to migrate away from the current execution model that the CNI currently leverages. | ||
|
|
||
| The next challenge is the deployment of the CNI plugins that provide the network setup and teardown of the pod. The idiomatic way to deploy a workload in Kubernetes is that everything should be in a pod; however the current approach leaves files in the host filesystem such as CNI binaries and CNI configuration. These files are usually downloaded via the init container of the pod after binding, which increases the time for the pod to get to a running state. Since all existing K8s network plugins are running as daemonsets we will take this approach as well, where all the dependencies are packaged into the container image thus adopting a well known approach. This will have added benefits of the network pod startup being much faster as nothing should need to be downloaded. |
There was a problem hiding this comment.
These files are usually downloaded via the init container of the pod after binding, which increases the time for the pod to get to a running state.
This is confusing me a bit - typically the CNI config is something workload pods are not aware of at all, via init containers or otherwise - only the CRI implementation handles them?
There was a problem hiding this comment.
The pod I was referring to was the network plugin daemonset pods (flannel, calico, ...). I can try and clean this up to be more clear.
There was a problem hiding this comment.
Ah ok thanks, yeah - "node agent daemonset" or "privileged node-bound pod" or whatever. Something general and consistent, since I don't think it has to be an init container specifically, in the Pod sense.
There was a problem hiding this comment.
Please see also #4477 (comment), and the followup comments
I was referring to kernel isolated pods as pods that leverage kata or kubevirt. The additional network setup happens for both kata/kubevirt use cases after the CNI add. This is done in a couple ways, for Kata the setup happens through the execution of the kata-runtime via containerd/cri-o. In containerd its via the StartTask thus not clear that additional networking is happening. I can squash the commits as well. It might make sense to move the contents to a google doc as well? |
|
|
||
| Currently networking happens in three layers of the stack, Kubernetes itself by means of kube-proxy or another controller based solution, the container runtime with network namespace creation and CNI plugins and the OCI runtime which does additional network setup for kernel isolated pods. All of this communication happens through non network specific APIs which to the reader of the code makes it hard to determine where ‘networking’ is happening. Having networking in several layers presents an issue when needing to troubleshoot issues as one needs to check several areas and some cannot be done via kubectl logs such as the CNI execution logs. This becomes more of an effort as multiple uncoordinated processes are making changes to the same resource, the network namespace of either the root or pod. The KNI aims at reducing the complexity by consolidating the networking into a single layer and having a uniform process for both namespaced and kernel isolated pods through a gRPC API. Leveraging gRPC will allow users the ability to migrate away from the current execution model that the CNI currently leverages. | ||
|
|
||
| The next challenge is the deployment of the CNI plugins that provide the network setup and teardown of the pod. The idiomatic way to deploy a workload in Kubernetes is that everything should be in a pod; however the current approach leaves files in the host filesystem such as CNI binaries and CNI configuration. These files are usually downloaded via the init container of the pod after binding, which increases the time for the pod to get to a running state. Since all existing K8s network plugins are running as daemonsets we will take this approach as well, where all the dependencies are packaged into the container image thus adopting a well known approach. This will have added benefits of the network pod startup being much faster as nothing should need to be downloaded. |
There was a problem hiding this comment.
Since all existing K8s network plugins are running as daemonsets we will take this approach as well [...]
This is definitely not true ...?
CNI plugin binaries can be pre-installed to the host and run entirely on the host without a daemonset and there are major cluster operators working this way.
Through CRI, even the plumbing node IPAM can be done without any pod.
There was a problem hiding this comment.
CNI plugin binaries can be pre-installed to the host and run entirely on the host without a daemonset and there are major cluster operators working this way.
Yep, some environments bake the nodes and don't let you change them (and the corollary of that is you can't use other CNIs or meshes). Some do.
Either way, the current gate is "you need node privilege to install CNI plugins" - some envs have K8S policies that allow pods to assume node-level privileges, some do not and prebake everything. That k8s-based gate should be maintained for KNI, so either option works, depending on operator/environment need.
I don't think there's a world in which we can effectively separate "able to add new networking extensions" from "require privileged access to the node". That's worth considering tho.
If extensions are driven by K8S resources it makes policy enforcement via operator-managed admission hooks potentially a bit easier, I guess.
There was a problem hiding this comment.
Either way, the current gate is "you need node privilege to install CNI plugins" [...]
I would expect that to be true, given the position they fill in the stack (!)
There was a problem hiding this comment.
How you deploy the CNI plugins can be opinionated. However, the piece that is important is that you no longer need to have any files in the host filesystem such as CNI binaries or CNI configuration files.
There was a problem hiding this comment.
However, the piece that is important is that you no longer need to have any files in the host filesystem such as CNI binaries or CNI configuration files.
But why is that important? I mean taken to an extreme ... You cannot run Kubernetes without files on the host filesystem anyhow, and configuring the pod network is an incredibly privileged place to operate.
There was a problem hiding this comment.
But why is that important? I mean taken to an extreme ... You cannot run Kubernetes without files on the host filesystem anyhow, and configuring the pod network is an incredibly privileged place to operate.
It's important because allowing privileged pods to run on a node (where privileged necessarily means "can mutate node state", networking or otherwise) is an operator choice today, and it seems wrong to take that choice away from operators.
This is how, for instance, it is possible to install Cilium in AWS today. Or Calico in GKE. Or Openshift, etc etc.
Anyone that doesn't want to allow privileged pods on a node can already choose to leverage existing K8S APIs to preclude that as a matter of operational policy - it's rather orthogonal to CNI.
There was a problem hiding this comment.
Wouldn't it be "less layers" to just put opinionated networking in containerd/cri-o ?
CNI is enables extensibility, and we've got a large and healthy ecosystem around it.
This feels like a big technical-debt inducing, disruptive undertaking in search of a concrete need.
I don't see any discussion of goals that cannot be accomplished by improving the current tools, except some non-specific hand-waving about reducing layers.
There also seem to be some bad assumptions, e.g. that "all plugins are populated by daemonset" (and therefore network readiness is blocked by pulling / starting this pod) which is not true, e.g. off the top of my head: GKE does not do this, kind does not do this (you can pre-populate plugin binaries / config or containerd CNI config template with very fast node network readiness), and then KNI is suggested to ... run as a container image, which will ... block network readiness on pulling a container...?
At the very least, I'd suggest spelling out the use-cases and why they cannot be accomplished without a new RPC service.
|
|
||
| Another area that KNI will improve is ‘network readiness’. Currently the container runtime is involved with providing both network and runtime readiness with the Status CRI RPC. The container runtime defines the network as ‘ready’ by the presence of the CNI network configuration in the host file system. The more recent CNI specification does include a status verb, however this is still bound by the current limitations, files on disk and execution model. The KNI will provide an RPC that can be implemented so the kubelet will call the KNI via gRPC. | ||
|
|
||
| KNI aims to help the community and other proposals in the Kubernetes ecosystem. We will do this by providing necessary information via the gRPC service. We should be the API that provides the “what networks are available on this node” so that another effort can make the kube-scheduler aware of networks. We should also provide IPAM status as a common issue, is that the IPAM runs out of assignable IP addresses and pods are no longer able to be scheduled on that node until intervention. We should provide visibility into this so that we can indicate “no more pods” as setting the node to not ready will evict the healthy pods. While the future state of KNI could aim to propose changes to the kube-scheduler, it's not a part of our initial work and instead should try to assist other efforts such as DRA/device plugin to provide the information they need. |
There was a problem hiding this comment.
We should also provide IPAM status as a common issue, is that the IPAM runs out of assignable IP addresses and pods are no longer able to be scheduled on that node until intervention.
Typically this is handled by coordinating the assigned IP range and the max pods setting in the kubelet.
Cluster operators already have the tools to prevent this issue, why would you allow a node to be configured to have more pods than IPs?
|
|
||
| KNI aims to help the community and other proposals in the Kubernetes ecosystem. We will do this by providing necessary information via the gRPC service. We should be the API that provides the “what networks are available on this node” so that another effort can make the kube-scheduler aware of networks. We should also provide IPAM status as a common issue, is that the IPAM runs out of assignable IP addresses and pods are no longer able to be scheduled on that node until intervention. We should provide visibility into this so that we can indicate “no more pods” as setting the node to not ready will evict the healthy pods. While the future state of KNI could aim to propose changes to the kube-scheduler, it's not a part of our initial work and instead should try to assist other efforts such as DRA/device plugin to provide the information they need. | ||
|
|
||
| The community may ask for more features, as we are taking a bold approach to reimagining Kubernetes networking by reducing the amount of layers involved in networking. We should prioritize feature parity with the current CNI model and then capture future work. KNI aims to be the foundational network api that is specific for Kubernetes and should make troubleshooting easier, deploying more friendly and innovate faster while reducing the need to make changes to core Kubernetes. |
There was a problem hiding this comment.
The community may ask for more features, as we are taking a bold approach to reimagining Kubernetes networking by reducing the amount of layers involved in networking.
Reducing the layers ... by adding a new GRPC service? Why isn't this CRI? (An existing Kubernetes-Specific gRPC service for Pods, which includes some networking related bits today ...)
There was a problem hiding this comment.
Adding this to the CRI-API is a design detail. The only piece of information relevant to networking is the PodIP coming back in PodSandboxStatus. This talks about eliminating the network setup/teardown and netns creation from the container/oci runtimes.
There was a problem hiding this comment.
This talks about eliminating the network setup/teardown and netns creation from the container/oci runtimes.
Has this been a significant obstacle for implementers? Examples?
What blocks a CNI revision from handling netns/teardown?
|
|
||
| Since the network runtime can be run separated from the container runtime, you can package everything into a pod and not need to have binaries on disk. This allows the CNI plugins to be isolated in the pod and the pod will never need to mount /opt/cni/bin or /etc/cni/net.d. This offers a potentially more ability to control execution. Keep in mind CNI is the implementation however when this is used chaining is still available. | ||
|
|
||
| ## Ongoing Considerations |
There was a problem hiding this comment.
There doesn't seem to be any discussion of how we might improve CNI, CRI instead and why that isn't sufficient versus this entirely new API and RPC service.
There was a problem hiding this comment.
This could live in the CRI-API as multiple services, no one has indicated that we must use a new API. However the CNI 2.0 was talked about being closer to K8s for years now. This is that effort.
There was a problem hiding this comment.
I do agree we could probably be more clear up front in the KEP about why the current CNI model (slurping out-of-k8s config files from well-known paths, TOCTOU errors, telling CRI implementations via out-of-band mechanisms to exec random binaries on the node by-proxy) is something we could explicitly improve on with KNI, and that KNI is basically "CNI 2.0" - it is the proposed improvement to current CNI.
|
|
||
| ### User Stories | ||
|
|
||
| We are constantly adding these user stories, please join the community sync to discuss. |
There was a problem hiding this comment.
We are constantly adding these user stories [...]
Where?
|
|
||
| ## Motivation | ||
|
|
||
| Kubernetes networking is an area of complexity and multiple layers which has created several challenges and areas of improvement. These challenges include deployment of the CNI plugins, troubleshooting networking issues and development of new functionality. |
There was a problem hiding this comment.
Yes, please. The current formatting has too many discussion points per line, it's difficult to follow and comment on.
| ### Goals | ||
|
|
||
| - Design a cool looking t-shirt | ||
| - Provide a RPC for the Attachment and Detachment of interface[s] for a Pod |
There was a problem hiding this comment.
This seems like a significant departure from current expectations around pod networking.
There was a problem hiding this comment.
Can you clarify this? We already do something similar with CNI ADD/DEL with an execution model. This is levering gRPC to communicate with the gRPC server that would be flannel, calico, cilium.
| - Design a cool looking t-shirt | ||
| - Provide a RPC for the Attachment and Detachment of interface[s] for a Pod | ||
| - Provide a RPC for the Querying of Pod network information (interfaces, network namespace path, ip addresses, routes, ...) | ||
| - Provide a RPC to prevent additional scheduling of pods if IPAM is out of IP addresses without evicting running pods |
There was a problem hiding this comment.
Surely this is just reporting status up through CRI?
There was a problem hiding this comment.
Last place I mention this, that the decision around CRI or a new API is a design detail.
| approvers: | ||
|
|
||
| see-also: | ||
| - "/keps/sig-aaa/1234-we-heard-you-like-keps" |
There was a problem hiding this comment.
delete these, or update with relevant keps?
same below for replaces
| milestone: | ||
| alpha: "v1.31" | ||
| beta: "v1.32" | ||
| stable: "v1.33" |
There was a problem hiding this comment.
This seems unlikely, without even defining an API yet?
| - "@shaneutt" | ||
| owning-sig: sig-network | ||
| participating-sigs: | ||
| - sig-network |
There was a problem hiding this comment.
Surely at least SIG Node should be participating (there's no way this doesn't affect kubelet, CRI)..?
I would also tag Cluster Lifecycle at least as FYI / advisory since cluster lifecycle folks will know about and have suggestions re: node readiness and cluster configuration.
|
If use cases and design are still being collected, a working group may be more appropriate than a KEP. |
Lets sync up on slack and schedule some time to discuss. |
|
Are there any serious KNI use-cases that don't include multi-networking? KNI doesn't enable multiple interfaces, but I think that something like KNI is a prerequisite for K8s multi-net. But to independently motivate both KNI and K8s-multi-net with multi-networking use-cases is very confusing. I hope Antonio's workshop at KubeCon will sort this out (great initiative! But I can't attend myself unfortunately). But the referred comment above must be considered if multiple interfaces are handled via KNI: who cleans up when a POD is deleted? Obviously someone who knows what interfaces that are in the POD. So if something except |
|
The Kubernetes project currently lacks enough contributors to adequately respond to all PRs. This bot triages PRs according to the following rules:
You can:
Please send feedback to sig-contributor-experience at kubernetes/community. /lifecycle stale |
k8s-ci-robot
added
the
lifecycle/stale
