SRCco.de

Docker with IPv6 and Resource Isolation

Posted:   |  More posts about docker pequod
../galleries/docker-logo.png

To experiment with Docker and IPv6 in "production", I decided to migrate this tiny blog to our (unreleased) Docker-based Pequod Cloud Solution. I will briefly describe how the setup looks like.

As of today this blog runs in its own small Docker container which regularly rebuilds static HTML from github source and serves it with CherryPy.

What did I do?

We are currently building our own Docker-based Pequod Cloud Solution at Zalando Technology and thus I decided to use the same technology to drive my own tiny blog (even if it looks over-the-top for this use case).

First I had to upgrade my root server to Ubuntu 14.04 to get an appropriate Kernel to run Docker.

Next I installed the required packages to run the Pequod Cluster Agent, these are:

  • Docker 1.3 (following the Ubuntu installation steps)
  • HAProxy 1.5
  • bridge-utils (apt-get install bridge-utils)
  • Tayga for NAT64 (apt-get install tayga)
  • Bind9 (was already running)

Obviously I have to install the Pequod Cluster Agent itself, too:

$ sudo pip3 install pequod-agent

In order to get my application running, I had to:

Every application must have an application manifest (pequod.xml) in order to run on Pequod, The manifest defines resource limits, filesystem mounts and services the application provides/requires.

The application is a very simple bash script which starts a CherryPy server to serve static HTML and at the same time loops endlessly over git pull and nikola build (Nikola is the static blog generator I'm using).

As the git pull requires access to github.com, I also defined a helper bootstrap YAML file to define github as an external service (usually this would be done in a specific Pequod cluster component, but I'm using pequod-agent in "standalone mode")

Building the necessary Docker image is easy:

$ docker build -t srcco-de:1 .

I created a small upstart init file to start the application with the Pequod Cluster Agent:

$ cat /etc/init/pequod-agent.conf

respawn
exec pequod-agent --bootstrap srcco-de:1 /root/srcco.de/pequod-bootstrap.yaml

After starting (service pequod-agent start), the Docker container will automatically get connected to github. The agent's --inspect shows this:

$ pequod-agent -i services
Instance Name              Repo Application Name Ver Type    Service      Ver Local Hostname                       Local Port
bootstrap--srcco-de-1-9641      srcco-de         1   provide srcco-de     1   2a01:4f8:190:314e:aacc:6f04:509:944e       8000
bootstrap--srcco-de-1-9641      srcco-de         1   require github-https 1   github.com                                  443

We can peek into the container using nsenter:

$ pequod-agent -i
Instance Name              Repo Application Name Ver Zone      PID   IPv6                                 Started
bootstrap--srcco-de-1-9641      srcco-de         1   bootstrap 14695 2a01:4f8:190:314e:aacc:6f04:509:944e 57m ago

$ nsenter --net -t 14695 /bin/ip a show scope global
40: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP group default qlen 1000
    link/ether 76:82:4b:8b:b1:a4 brd ff:ff:ff:ff:ff:ff
    inet6 2a01:4f8:190:314e:ff11:c3b3:509:944e/64 scope global
       valid_lft forever preferred_lft forever
    inet6 2a01:4f8:190:314e:aacc:6f04:509:944e/64 scope global
       valid_lft forever preferred_lft forever

As you can see from the ip a output, the Docker container has IPv6 connectivity only. If you have IPv6 connectivity on your local machine, you can access the Docker container directly.

Summary

This little experiment worked out so far (it seems, at least you can apparently read this page ;-)). The setup is quite complex compared to my former setup with simple static HTML served by Apache.

What did I get in the end?

  • This site is now automatically updated whenever I push to github.
  • I proved that it's feasible to run the Pequod Cluster Agent in a small "standalone" environment.
  • Packaging this site with Docker makes it easy to run it elsewhere.
  • The Pequod agent gives me better resource isolation than the default docker run.
  • I finally configured my public /64 IPv6 subnet :-)

What are the downsides?

  • Using CherryPy was only a "quick-and-dirty" choice to serve static HTML, the performance is by far not optimal.
  • Running the Pequod Cluster Agent without the Pequod core components provides reduced availability. E.g. the Service Scheduler would restart instances in a regular Pequod cluster as soon as they die.
  • The setup is overly complex for this tiny static blog site.

You can find all source code of this site in my github repository.

You can find some more information about Pequod in the Pequod Documentation.