Dan Everton bio photo

Dan Everton

Tinkerer-in-Chief

Email Keybase Twitter Google+ Github Stackoverflow

I was lucky enough to be able to attend DockerCon 2015 in San Francisco this year. With over 2100 attendees it’s one of the biggest tech conferences I’ve been to and the first one outside of Australia. I learned a lot while I was there so I wanted to share my notes from the conference.

NOTE: As of 2015-05-26 This is still somewhat of a draft and may be updated

Resources

If you want to watch the talks yourself, here’s some links to the videos and slides.

Day 1

General Session

The key things out of the general session that started the day were:

  • Docker itself is going through rapid growth in adoption and use. There were some very large growth numbers for things like the number of jobs mentioning Docker and the use of Docker hub.
  • The Docker team have big ambitions around using containers and Docker to make development easier. Broadly this meant making building systems that could be distributed (in both “deploying on multiple nodes” and “packaging in an easy to use manner” senses) much easier.
  • There’s now experimental releases of Docker available to download for those that want to be on the “bleeding edge”.
  • A new networking system has been added to Docker. This system allows for multi-host networking using VXLAN tunnels to link containers on different machines. It also enables service discovery and advertising using DNS. Note that this feature requires a backing KV store (like Consul] to be deployed.
  • A plugin system for Docker that initially allows for the network and volume systems to be extended. The plugin interface is fairly limited so far, but already plugins can be added to enable some neat tricks like live container migration including volumes and data.
  • Deepak Singh from Amazon announced that the EC2 Container Service (ECS) will support Docker Swarm and Docker Compose later this year. This is a big announcement for me as I’m spending a lot of time with ECS lately and native Docker tools support would be a great help. It will also make migrating between cluster technologies (e.g. ECS to Mesos) much easier.
  • Signed data support with Docker Notary. The Notary system provides a way to sign and verify content including Docker images and updates. It differs from PGP/GPG in that it is more optimised for handling updates and provides simpler handling of key revocation in case of compromise.
  • Probably the biggest news of the conference was that large parts of Docker are to become an open standard. The standard, called the Open Container Format, will be maintained by the Linux Foundation under the Open Container Project banner. Docker Inc has donated a container runtime called runC which will be the reference implementation. There are a huge number of companies which are all participating in this standards effort so it seems the possible standards war with CoreOS and rkt has been avoided.

Faster, Cheaper and Safer: Secure Microservice Architectures using Docker

Adrian Cockcroft’s presentation covered how best to securely deploy containers. The first thing would be to read Docker’s security page which covers best practices when securing Docker deployments. Other things of note:

  • Keep your containers thin and single purpose. This reduces the attack surface and makes auditing easier.
  • Add automated pen testing to your build process
  • Aim for immutable services with state stored elsewhere.
  • For really secure systems, deploy a single container on a virtual machine to get the isolation and security benefits of both.

Reliably shipping containers in a resource rich world using Titan

Diptanu Choudry presented a description of how Netflix is deploying Docker containers on a system they call Titan. It’s a PaaS system built using Mesos and other Netflix libraries. It started out just running cron jobs and grew from there. They expect to make it open source later this year. Other interesting bits:

  • Volume management makes use of ZFS snapshots and replication
  • Each container in a VPC gets a unique IP address. An ENI is created and moved into the container namespace.
  • Container (cgroup) metrics are published to Atlas.
  • Uses Eureka for service discovery and advertising.
  • For container IAM roles, each AWS API call goes through a security proxy that verifies the source IP (since each container has an IP) and container is authorized to use that role. Uses STS to generate temporary credentials.

Docker Plugins

This was a combination presentation from some of the key people leading the plugin implementation in Docker: Jeff Lindsay, Luke Marsden, and Alexis Richardson. The title was originally a secret as the plugin feature was only announced at the keynote.

  • Major functions of Docker will become pluggable though currently only networking and storage are completed. Logging coming soon.
  • Plugins are separate processes that are talked to using JSON over HTTP via a domain socket. This enables plugins to be written in any language.
  • Since plugins are processes, they can also be containers which simplifies deployment.
  • Demonstrated using network (Weave) and volume (Flocker) plugins to enable live container migration. You can try the demo on your machine.

Enabling Microservices @ Orbitz

Another combo presentation from Steve Hoffman and Rick Fast of Orbitz. They talked about Orbitz journey to enable fast, simple deployment of microservices while keeping the existing process as far as possible. The end result is a mostly automated platform using Jenkins to coordinate actions and deployments to a Mesos cluster in their datacenter.

  • Use Chef to build the base machines including Docker, Mesos, and Consul.
  • Uses Consul for service discovery and advertising.
  • Marathon is used to deploy Docker container applications on Mesos.
  • HAProxy cluster used to provide access to the applications deployed on Mesos. Uses Bamboo to keep the HAProxy configuration in sync with Mesos.
  • Before deploying to Mesos, change notification tickets are automatically created in ServiceNow by Jenkins. Once the deploy completes the ticket is updated with the outcome.

Docker Networking

Since Docker acquired Socketplane they’ve been working hard on refactoring the networking component of Docker. Madhu Venugopal and Jana Radhakrishnan presented a breakout session on the new networking stack.

  • Implemented as a separate library called libnetwork which can be used by other projects needing network functions
  • Models a Container Networking Model (CNM) consisting of Endpoint, Network, and Sandbox.
  • Endpoints are service connections. They’re separate from containers and have a life cycle of their own.
  • Networks are collections of endpoints that can communicate.
  • A sandbox is a group of endpoints that can be attached to a container. They provide an abstraction over the concept of a container.
  • Drivers are used to actually talk to the operating system network stack. Multiple implementations available with Docker 1.8 including a default “bridged network” driver for backwards compatibility.
  • Overlay network driver (using VXLAN tunnels) will also be provided for multi-host networking.
  • Service discovery and advertising is built in to libnetwork as well. It uses DNS to provide this feature in Docker 1.8. Future releases will make this a feature that plugins can implement to provide alternatives
  • Requires a shared datastore (for example Consul) that all Docker hosts can see to coordinate networks.

Day 2

General Session

  • Started with a walk through of how Business Insider used Docker to decrease their release cycle time. They used their existing Puppet infrastructure to deploy and manage Docker containers as services. Docker allowed the dev and ops teams to work together better without having to radically change their processes.
  • Docker Hub has seen huge growth in use over the last year. Docker is working on improving the experience, increasing the security and performance at the same time. A beta of a new UI is available.
  • One of the most popular images on Docker Hub is the Docker registry itself. The Docker team is now making a commercial, supported version of the registry called Docker Trusted Registry. This commercial version of the registry supports enterprise features like LDAP authentication, role based access control, and auditing.
  • Docker Inc. now offers commercial support for Docker deployments. Prices start from $150/month and include a license for the Trusted Registry.
  • Microsoft is integrating Docker support in to its development tools, operating system, and cloud platform. Windows applications will be able to be containerized and deployed to Windows Server hosts as well.
  • Finally a demonstration of a future product labelled Project Orca. This is a web UI that brings together all of the Docker tools like machine, compose, and swarm. You can administer clusters of Docker engines and the applications running on them from a single UI. It’s still mostly a concept at this point, early signup is available at http://bit.ly/project-orca

The distributed system toolkit: Container patterns for modular distributed system design

Brendan Burns from Google’s Cloud Platform team presented on different ways to think about system design with containers.

  • Treat containers like classes in an object oriented system.
  • Compose systems out of multiple, reusable, single purpose containers.
  • The container is natural boundary for teams and allows you to use expert knowledge everywhere and only update the bits that need updating.
  • Ensure containers are parameterized and that they are documented.
  • Sidecar containers let you compose multiple functions together to form a new application. The sidecar can extend and enhance existing functions.
  • Ambassador containers act as an intermediary between an application container and another service. This extra level of indirection allows for more features to be added without the application container being aware.
  • Adapter containers interact with the application container to adapt it to its deployed environment. For example an adapter container might extract metrics from the application container and publish them.

Scaling New Services: From Container Creation to Automated Deployments

Presented by Brian Scott and Patrick O’Connor from Disney this presentation largely covered similar ground to the Orbitz one. Again Mesos, Marathon, Chef, Docker, Jenkins, HAProxy, and Consul work together to provide a PaaS like system for automated deployments.

Container Hacks and Fun Images

Jessie Frazelle from Docker Inc. demonstrated some atypical uses of Docker containers. With some care you can deploy GUI/TUI applications as containers on the desktop. Even ones you wouldn’t think of like Skype or Chrome.

Running Aground: Debugging Docker in Production

Brian Cantrell was probably the most engaging presenter there. His presentation on debugging Docker in production covered a lot of the potential pitfalls and possible workarounds.

  • Docker and microservices are an embodiment of the Unix Philosophy
  • Microservices means that you are developing a distributed system even if you think it isn’t.
  • Distributed systems are hard to debug, you need to think of debugging systems not instances.
  • Failures can be broadly categorized in to Non-fatal/Fatal and Implicit/Explicit.
  • Fatal failures are easier to debug especially if you can capture the in-memory state at the time of failure, e.g. a core dump. Docker lacks a core dump feature but it is coming.
  • Non-fatal explicit failures require more runtime tools like logs or instrumentation to diagnose. Debugging these typically requires things like log analysis.
  • When deploying with containers, try very hard not to manipulate the Docker host in any way. As soon as you do, those Docker hosts become compromised and many assumptions about immutability and security are broken. If you need something from the host, make sure it’s shipped somewhere else for safety and easy of access.
  • The ability to debug and diagnose production problems is essential for Docker deployments to succeed.

Final General Session

I don’t have too many notes from this, but I do remember one of the more impressive final demonstrations was the live migration of a Quake 3 server while players were connected. You can watch it here https://www.youtube.com/watch?v=mL9AFkJJAq0