An open source system for automating deployment, scaling, and operations of applications.

Thursday, January 14, 2016

Why Kubernetes doesn’t use libnetwork

Kubernetes has had a very basic form of network plugins since before version 1.0 was released — around the same time as Docker's libnetwork and Container Network Model (CNM) was introduced. Unlike libnetwork, the Kubernetes plugin system still retains its "alpha" designation. Now that Docker's network plugin support is released and supported, an obvious question we get is why Kubernetes has not adopted it yet. After all, vendors will almost certainly be writing plugins for Docker — we would all be better off using the same drivers, right?

Before going further, it's important to remember that Kubernetes is a system that supports multiple container runtimes, of which Docker is just one. Configuring networking is a facet of each runtime, so when people ask "will Kubernetes support CNM?" what they really mean is "will kubernetes support CNM drivers with the Docker runtime?" It would be great if we could achieve common network support across runtimes, but that’s not an explicit goal.

Indeed, Kubernetes has not adopted CNM/libnetwork for the Docker runtime. In fact, we’ve been investigating the alternative Container Network Interface (CNI) model put forth by CoreOS and part of the App Container (appc) specification. Why? There are a number of reasons, both technical and non-technical.

First and foremost, there are some fundamental assumptions in the design of Docker's network drivers that cause problems for us.

Docker has a concept of "local" and "global" drivers. Local drivers (such as "bridge") are machine-centric and don’t do any cross-node coordination. Global drivers (such as "overlay") rely on libkv (a key-value store abstraction) to coordinate across machines. This key-value store is a another plugin interface, and is very low-level (keys and values, no semantic meaning). To run something like Docker's overlay driver in a Kubernetes cluster, we would either need cluster admins to run a whole different instance of consul, etcd or zookeeper (see multi-host networking), or else we would have to provide our own libkv implementation that was backed by Kubernetes.

The latter sounds attractive, and we tried to implement it, but the libkv interface is very low-level, and the schema is defined internally to Docker. We would have to either directly expose our underlying key-value store or else offer key-value semantics (on top of our structured API which is itself implemented on a key-value system). Neither of those are very attractive for performance, scalability and security reasons. The net result is that the whole system would significantly be more complicated, when the goal of using Docker networking is to simplify things.

For users that are willing and able to run the requisite infrastructure to satisfy Docker global drivers and to configure Docker themselves, Docker networking should "just work." Kubernetes will not get in the way of such a setup, and no matter what direction the project goes, that option should be available. For default installations, though, the practical conclusion is that this is an undue burden on users and we therefore cannot use Docker's global drivers (including "overlay"), which eliminates a lot of the value of using Docker's plugins at all.

Docker's networking model makes a lot of assumptions that aren’t valid for Kubernetes. In docker versions 1.8 and 1.9, it includes a fundamentally flawed implementation of "discovery" that results in corrupted /etc/hosts files in containers (docker #17190) — and this cannot be easily turned off. In version 1.10 Docker is planning to bundle a new DNS server, and it’s unclear whether this will be able to be turned off. Container-level naming is not the right abstraction for Kubernetes — we already have our own concepts of service naming, discovery, and binding, and we already have our own DNS schema and server (based on the well-established SkyDNS). The bundled solutions are not sufficient for our needs but are not disableable.

Orthogonal to the local/global split, Docker has both in-process and out-of-process ("remote") plugins. We investigated whether we could bypass libnetwork (and thereby skip the issues above) and drive Docker remote plugins directly. Unfortunately, this would mean that we could not use any of the Docker in-process plugins, "bridge" and "overlay" in particular, which again eliminates much of the utility of libnetwork.

On the other hand, CNI is more philosophically aligned with Kubernetes. It's far simpler than CNM, doesn't require daemons, and is at least plausibly cross-platform (CoreOS’s rkt container runtime supports it). Being cross-platform means that there is a chance to enable network configurations which will work the same across runtimes (e.g. Docker, Rocket, Hyper). It follows the UNIX philosophy of doing one thing well.

Additionally, it's trivial to wrap a CNI plugin and produce a more customized CNI plugin — it can be done with a simple shell script. CNM is much more complex in this regard. This makes CNI an attractive option for rapid development and iteration. Early prototypes have proven that it's possible to eject almost 100% of the currently hard-coded network logic in kubelet into a plugin.

We investigated writing a "bridge" CNM driver for Docker that ran CNI drivers. This turned out to be very complicated. First, the CNM and CNI models are very different, so none of the "methods" lined up. We still have the global vs. local and key-value issues discussed above. Assuming this driver would declare itself local, we have to get info about logical networks from Kubernetes.

Unfortunately, Docker drivers are hard to map to other control planes like Kubernetes. Specifically, drivers are not told the name of the network to which a container is being attached — just an ID that Docker allocates internally. This makes it hard for a driver to map back to any concept of network that exists in another system.

This and other issues have been brought up to Docker developers by network vendors, and are usually closed as "working as intended" (libnetwork #139, libnetwork #486, libnetwork #514, libnetwork #865, docker #18864), even though they make non-Docker third-party systems more difficult to integrate with. Throughout this investigation Docker has made it clear that they’re not very open to ideas that deviate from their current course or that delegate control. This is very worrisome to us, since Kubernetes complements Docker and adds so much functionality, but exists outside of Docker itself.

For all of these reasons we have chosen to invest in CNI as the Kubernetes plugin model. There will be some unfortunate side-effects of this. Most of them are relatively minor (for example, docker inspect will not show an IP address), but some are significant. In particular, containers started by docker run might not be able to communicate with containers started by Kubernetes, and network integrators will have to provide CNI drivers if they want to fully integrate with Kubernetes. On the other hand, Kubernetes will get simpler and more flexible, and a lot of the ugliness of early bootstrapping (such as configuring Docker to use our bridge) will go away.

As we proceed down this path, we’ll certainly keep our eyes and ears open for better ways to integrate and simplify. If you have thoughts on how we can do that, we really would like to hear them — find us on slack or on our network SIG mailing-list.

Tim Hockin, Software Engineer, Google

No comments:

Post a Comment