Today I would like to announce a new open source project called RancherOS – the smallest, easiest way to run Docker in production and at scale. RancherOS is the first operating system to fully embrace Docker, and to run all system services as Docker containers. At Rancher Labs we focus on building tools that help customers run Docker in production, and we think RancherOS will be an excellent choice for anyone who wants a lightweight version of Linux ideal for running containers.
How RancherOS began
The first question that arises when you are thinking about putting Docker in production is which OS to use. The simplest answer is to run Docker on your favorite Linux distribution. However, it turns out the real answer is a bit more nuanced. After running Docker on just about every distro over the last year, I eventually decided to create a minimalist Linux distribution that focuses explicitly on running Docker from the very beginning.
Docker is a fast-moving target. With a constant drum beat of releases, it is sometimes difficult for Linux distributions to keep up. In October 2013, I started working very actively with Docker, eventually leading to an open source project called Stampede.io. At that time I decided to target one Linux distribution that I thought to be the best for Docker since it was included by default. With Stampede.io, I was pushing the boundaries of what was possible with Docker and able to do some fun things like run libvirt and KVM in Docker containers. Consequentially I always needed the latest version of Docker, which was problematic.
At the time, Docker was releasing new versions each month (currently Docker is on a two month release cycle). It would often take over a month for new Docker versions to make it to the stable version of the Linux distro. Initially, this didn’t seem like a bad proposition because undoubtedly, a new Docker release couldn’t be considered “stable” on day one, and I could always use alpha releases of my distribution of choice. However, alpha distribution releases include other recently released software, not just Docker but alpha kernel and alpha versions of other software.
With RancherOS we addressed this by limiting the OS to just the things we need to run Docker, specifically, the Linux kernel, Docker and the bare minimum amount of code needed to join the two together. Picking which version of RancherOS to run is as easy as saying which version of Docker you wish to run. The sole purpose of RancherOS is to run Docker and therefore our release schedule is closely aligned. All other software included in the distribution is considered stable, even if you just picked up the latest and greatest Docker version.
An OS where everything is a container
When most people think of Docker they think about running applications. While Docker is excellent at that, it can also be used to run system services thanks to recently added capabilities. Since first starting Rancher (our Docker orchestration product), we’ve wanted the entire orchestration stack to be packaged by and run in Docker – not just the application we were managing. This was initially quite difficult, since the orchestration stack needed to interact with the lower level subsystems. Nevertheless, we carried on with many “hacks” to make it possible. Once we determined the features we needed within Docker, we helped and encouraged the development of those items. Finally, with Docker 1.5, we were able to remove the hacks, paving the way for RancherOS. Docker now allows sufficient control of the PID, IPC, network namespaces and capabilities.
This means it is now possible to run systems oriented processes within Docker containers. In RancherOS we run absolutely everything in a container, including system services such as udev, DHCP, ntp, syslog, and cloud-init.
A Linux distro without systemd
I have been running systemd and Docker together for a long time. When I was developing Stampede.io I initially architected the system to run on a distribution that heavily leveraged systemd. I started to notice a number of strange errors when testing real world failure scenarios. Having previously run production cloud infrastructure, I cared very much about reliably managing things at scale. Having seen odd errors with systemd and Docker I started digging into the issue. You can see most of my comments in this Docker issue and this mailing list thread.
As it turns out, systemd cannot effectively monitor Docker containers due to the incompatibility with the two architectures. While systemd monitors the Docker client used to launch the Docker container, this is not really helpful and I worked hard with both Docker and systemd communities to fix the issue. I even went so far as to create an open source project called systemd-docker. The purpose of the project was to create a wrapper for Docker that attempted to make these two systems work well together. While it fixed many of the issues, there were still some corner cases I just couldn’t address. Realizing things must change in either Docker or systemd I shifted focus to talking to both of them.
With the announcement of Rocket more effort is being put into making systemd a better container runtime. Rocket, as it stands today, is largely a wrapper around systemd. Additionally, systemd itself has since added native support for pulling and running Docker images which seems to indicate that they are more interested in subsuming container functionalities in systemd than improving interoperability with Docker. Ultimately, all signs continue to point to no quick resolution between these two projects.
When looking at our use case for RancherOS, we realized we did not need systemd to run Docker. In fact, we didn’t need any other supervisor to sit at PID 1. Docker was sufficient in itself. What we have done with RancherOS is run what we call “System Docker” as PID 1. All containers providing core system services are run from System Docker, which also launches another Docker daemon which we call “User Docker” under which we run user containers. This separation is quite practical. Imagine a user did docker rm -f $(docker ps -qa). You run the risk of them deleting the entire operating system.
Minimalist Linux distributions
As users look at shifting workloads to containers, dependencies on the host system become dramatically less. All current minimalist Linux distributions have taken advantage of this fact, allowing them to drastically slim down their footprint. I love the model distributions such as CoreOS have pioneered and we have been inspired by them.
By constraining the use case of RancherOS to running Docker, we decided only core system services (logging, device management, alerting) and access (console, ssh) were required. With the ability to run these services in containers, all we needed was the container system itself and a bit of bootstrap code (to get networking up, for example). If you take this one step further and put the server under the management of a clustering/orchestration system, you can even minimize the need to run a full console.
On March 31st, I’ll be hosting an online meet up to demonstrate RancherOS, discuss some of the features we are working on, and answer any questions you might have. If you would like to learn more, please register now:
When we looked at simplifying large scale deployments of Docker, there were no solutions available that truly embraced Docker. We started the RancherOS project because we love Docker and feel we can significantly simplify the Linux distribution necessary to run it. Hopefully, this will allow users to focus more on their container workloads and less on managing the servers running them.
If you’re primary requirement for Linux is to run Docker, we’d love for you to give RancherOS a try and let us know what you think. You can find everything at https://github.com/rancherio/os.