Brief thoughts on Docker | Hacker News Reader

Brief thoughts on Docker

gorsuch.github.io

There are a bunch of cool and exciting things about Docker and containers in general. Docker just automates many manual tasks, essentially creating an abstraction layer above many manual and tricky steps. Similar to how Vagrant abstracted away many tasks with automating virtual machine deployment.

There are a couple really cool things though, such as the Docker push/pull functionality, giving you access to Docker images, Vagrant does something similar to Vagrant boxes, but dockers goes a little further. But the killer idea, is that you do not need an OS in these Docker containers. You can basically have a statically linked binary, or a rails app, or sshd daemon, and that's it. You can then use cgroups to finely tune your resource levels (cpu, memory, bandwidth, etc) on this container.

These concepts have been around since at least 2008 (LXC [1], was released into 2.6.24 [2]), Docker just streamlines the process greatly. As a side note, it looks like Google has been heavily using Linux containers for years! There is a Wired article [3], which talks about the orchestration engine they use to deploy containers across their cells. Google's John Wilkes even does a talk, entitle "Cluster management at Google" [4], where he discusses the next generation of the orchestration engine.

p.s. It looks like there might even be "live container migration" in the pipeline [5]. This would allow you to move a container from machineA to machineB without any downtime, similar to Xen migrate or VMware live motion.

[1] http://en.wikipedia.org/wiki/LXC

[2] http://kernelnewbies.org/Linux_2_6_24

[3] http://www.wired.com/wiredenterprise/2013/03/google-borg-twi...

[4] http://www.youtube.com/watch?v=0ZFMlO98Jkc

[5] https://www.youtube.com/watch?feature=player_detailpage&v=LD...

  But the killer idea, is that you do not need an OS in these  
  Docker containers. You can basically have a statically 
  linked binary, or a rails app, or sshd daemon, and that's it.

Is there documentation of this somewhere? I thought docker relied on having an operating system installed, even if just a minimal busybox based system.

I'm not sure if there is documentation at this time, but there was a Google Hangout, entitled "Docker: the Linux container runtime" [1] (minute 24), where Solomon Hykes and Jérôme Petazzoni (from dotCloud, makers of Docker) talk about this very thing.

[1] https://www.youtube.com/watch?feature=player_detailpage&v=LD...

A docker container runs in an operating system, but doesn't contain one.

Well, that depends on your definition of "operating system".

Is an OS just the kernel? Then yes, Docker doesn't contain an OS.

Is the OS the userland? E.g. the init process + the shell + the C library + other runtimes? Then most Docker containers actually do contain an OS. Yes you can get rid of most of them by making static binaries, but why would you want to? Making static binaries with anything besides Go is a pain, and good luck trying to statically link MySQL or Redis into your binary.

I don't think you need the init process even without a statically linked binary, and I don't see why would having a bunch of libraries suddenly make it an OS. The linking happens "outside" the container - in the shared kernel.

I think docker takes the place of init in the container, so no.

You are totally correct.

Lets take sshd for example, you would need to create a skeleton directory structure, /bin, /dev, /lib, etc in your rootfs container, have /bin/bash, a whack of /libs, some /dev devices, and then your sshd application. So an extremely minimalistic rootfs with only the sshd requirements. You could find these my running 'lsof -p `pidof sshd`' , maybe 'ldd /usr/sbin/sshd' (and all their dependencies) too, which quickly snowballs.

Hopefully this explains where I was going with that. Also, this is not a simple task to try and strip off these services into self contained entities. There are lots of hidden issues, like how do we handle logging? Should the container have syslog too? So, there is work that needs to happen, I just like the idea of not running a full fledged OS in a container.

This is solved by the union-based filesystem (AUFS) that Docker uses. You start with one minimal rootfs like you describe, given in the "base" image, then when you install sshd you get copy-on-write semantics. So the sshd container gets its own syslog files separate from any other container.

Your original comment talks about how long LXC has existed but AUFS is one key component of Docker that became part of mainline Linux much more recently.

> AUFS ... became part of mainline Linux much more recently

AUFS is not part of mainline kernel. Many distros include it (Debian, Ubuntu). But some don't (Fedora).

http://en.wikipedia.org/wiki/Aufs

Taking that idea even further, how cool would it be to have one binary that ran a pre-built docker container?

FreeBSD jails already provided those containers for over 10 years. Even chroot is some kind of container.

http://en.m.wikipedia.org/wiki/Operating_system-level_virtua...

Docker is great, but there is one security issue that it should solve. Right now, any user can manipulate any Docker container. If you create a container as root, then www-data can kill that container. The reason for this is because container orchestration is done through the Docker daemon, which listens on a TCP socket. Anybody on the local host can access that socket. They should, at the very least, implement password protection.

Today I also found out that it's not possible to run Docker inside Docker. :( I'm working on a Docker-based continuous integration system similar to Travis, and it would be great if I can distribute the CI system as a Docker container.

You can actually configure Docker to use an Unix socket instead of TCP[1]. It seems it still opens it up to every user (it runs a chmod 777 on it), but you can always enforce stricter controls (e.g. with SELinux).

[1] https://github.com/dotcloud/docker/pull/938

I came across this wiki page last night. It doesn't support this out of the box, but it looks like there is a way to run docker inside of a docker container. It seems to take a good bit of customization though.

https://github.com/dotcloud/docker/wiki/Docker-in-Docker

Docker will be changing to just listen on a unix domain socket by default.

I've been reading about Docker, Flynn and CoreOS on HN for the last several days. Would anyone care to venture how these technologies might work together? That is, could Flynn load CoreOS VMs/AMIs on a cloud service which hosts Docker containers? Does CoreOS etcd allow these VMs/containers to self-configure?

CoreOS is an operating system designed to run docker containers. You would use it as lighter weight replacement for something like Ubuntu as the base os that Docker runs on. When you install CoreOS you basically just get a kernel, Docker and etcd plus a minimal number of other processes. Etcd is designed as a way to allow a distributed set of containers to self-configure.

Flynn is an open source project to build a "platform as a service" platform on top of Docker. In theory Flynn could run on top of CoreOS.

Docker is the core of everything.

Anyone looked at combining something like ansible and docker for building images?

It's pretty straightforward to use ansible to build the host. This is one step forward but I wonder if anyone's used it to create images and if so how?

Would you build on top a base that includes the ansible libraries, or somehow run the ansible commands against a docker instance?

Could someone explain to me the difference between Docker and Vagrant ?

Vagrant manages virtual machines. Docker uses linux containers, which is essentially "chroot on steroids" [1]. A linux container is just an isolated process (from which you run your code) on the host machine, like a FreeBSD jail. The host machine can peer into the container, but the container cannot see the host machine. Docker just helps to automate and manages the linux container life-cycle.

p.s. I'm planning on doing a couple screencast episodes about linux containers and docker on my website, which should be online in the next week or so @ http://sysadmincasts.com/

[1] http://lxc.sourceforge.net/

Please post your screencasts here.

Seconded. I'm looking for as much learning material on Docker as possible.

They're actually quite similar. Docker is to lightweight containers as Vagrant is to virtual machines (more or less anyway; Vagrant lets you configure a lot more at this point). The nice thing is that the Dockerfile makes the whole thing a lot more transparent; VMs are sort of nebulous black boxes, whereas you can follow the construction of the container through the Dockerfile.

I'll be interested in seeing how the ecosystem evolves when Vagrant becomes better able to manage containers.

There is also a lot of talk about Docker and Vagrant working together. Like using Vagrant to manage Docker containers..

The typical use case is to use Vagrant to spin up the machine which then runs docker. I wouldn't recommend managing Docker containers with Vagrant unless you're really heavily invested in a pure vagrant workflow and for some reason cannot afford to call docker commands directly.

Fundamentally, Vagrant builds VirtualBox VMs, while Docker builds containers (e.g. LXC containers). Other than that, feature sets and communities are different.

And-- Vagrant has expanded a bit and now can build out a few different types of VMs, including VMWare images and AMIs for Amazon EC2.

I too, would really like to see a comparison between Docker and Vagrant with an LXC provider (https://github.com/fgrehm/vagrant-lxc).

I don't know for memory usage (although I'd bet containers uses less) but a VM take time to boot, with docker it's literally up in a second.

Are there any PaaS or VPS providers that are using Docker (other than DotCloud) ? I'm actually eager to give openstack-docker a testdrive.

If a PaaS/VPS provider is actually USING Docker to provide their service, it probably wouldn't be the best platform to actually TRY Docker in. It's easy enough to get it set up on DigitalOcean; I created a tutorial at https://www.digitalocean.com/community/articles/how-to-insta...

Actually my second sentence was not connected to the first. I was curious if there are any PaaS providers using Docker as the underlying infrastructure. And I also would like to setup openstack-docker and play around with it (not inside a PaaS or VPS of course). I did see your tutorials though, definitely another great way to test out Docker quickly.

Hello Relaxitup,

There are several projects to build a PaaS around Docker: - Deis: https://pypi.python.org/pypi/deis/0.0.4 - Dokku: https://github.com/progrium/dokku - Flynn: https://flynn.io/

Also, have a look at this blog post for playing with openstack + docker: http://blog.docker.io/2013/06/openstack-docker-manage-linux-...

Thanks julien421 will check these out.

Reply on news.ycombinator.com