OSDC 2014: Tobias Schwab - Continuous Delivery with Docker
2 likes1,212 views
The document outlines the philosophy and practical implementation of continuous delivery using Docker, highlighting concepts such as canary releases and immutable infrastructure. It discusses challenges faced with AWS and other hosting options, while detailing the benefits of Docker containerization for building and deploying applications consistently across environments. Key topics include Dockerfile usage, caching strategies, deployment pipelines, and logging/monitoring solutions within a containerized setup.
![• Problems
• unicorn: Rack HTTP server for fast clients
• static assets
• logging: default ruby syslog library uses syscall (needs local syslog daemon)
• Solution
• run 3 daemons in 1 container: unicorn, nginx and rsyslogd
• upstart
• ENTRYPOINT [“/sbin/init”]
• load ENV from /proc/1/environ
• foreman
Use Case: Ruby on Rails](https://image.slidesharecdn.com/tobiasschwab-docker-osdc2014-140425081503-phpapp02/85/OSDC-2014-Tobias-Schwab-Continuous-Delivery-with-Docker-22-320.jpg)
OSDC 2014: Tobias Schwab - Continuous Delivery with Docker
- 1. Continuous Delivery with Docker Tobias Schwab
- 2. Myself • Tobias Schwab • tobias.schwab@dynport.de • www.dynport.de • twitter.com/tobstarr • github.com/tobstarr
- 3. Philosophie • continuous delivery: deploy multiple times a day • canary releases • “never touch a running system” • “Immutable Infrastructure and Disposable Components" • don’t fix it, if it can be replaced
- 4. Theory • AWS • AMI based deployments • Elastic Load Balancer • AutoScaling Groups • S3, RDS, …
- 5. Reality • privacy concerns: AWS not an option • hoster we could not pick • first no, then proprietary and unreliable API • flash based infrastructure management • limited capacity • we were the biggest customer
- 6. Docker • build, distribute and deploy container based applications • creator: dotcloud • initial release: March 13, 2013 • license: Apache 2.0 • 11k stars on Github (top 50) • golang client/server
- 7. Container Virtualization • os level • shared kernel • cgroups: isolate CPU, Memory, Block IO, Network • lxc: cgroups + application namespaces • lightweight and fast
- 8. Images • blueprints for containers • tarball of os installation/packages • read only • stateless • layered
- 9. Containers • instances of images • copy on write / union file system • running or exited • goal: stateless and immutable • can be “saved” (docker commit) as images • created to be thrown away
- 10. Containers and images Source: http://docs.docker.io/en/latest/terms/container/
- 11. Demo
- 12. Build • manual • start and attach container • install required packages • checkout application code • run build management tool • bad: not reproducible • bad: does not utilise caching
- 13. Build • chef/puppet/… • start an attach container • run chef/puppet/… client • good: automated and documented • bad: does not utilise caching
- 14. Dockerfile • simple, plain text script to create images • commands: • FROM: base image to use • RUN: execute shell command • ENV: set environment variable • ADD: write local file to image • ENTRYPOINT: start command for containers • others: MAINTAINER, EXPOSE, CMD, USER, VOLUME, WORKDIR, ONBUILD
- 15. Dockerfile
- 16. Dockerfile
- 17. Caching • statement based: each step creates a new image • existing steps (command tree exists) are re-used • tricky: “non functional” commands (e.g. apt-get update/upgrade) • use ENV or comments to break caching of non functional commands
- 18. Caching
- 19. Configuration Management • “store config in the environment” (http://12factor.net/config) • dependency injected with start of container • same image for • development • testing • staging • production
- 20. Don’ts • full blown VMs • ssh daemon inside containers • syslog daemon inside containers (sometimes needed) • user management: everything can run as root • chef/puppet/… => makes caching useless
- 21. Build Management Tools • candidates: bundler, pip, mvn, carton, composer, … • problem with caching: bmt are slow when started with “clean slate” • option 1: add bmt manifest before code • bmt needs to run only when manifest changes • option 2: use pre-bundled base images • bmt only needs to work the delta • re-build base images from time to time • option 3: combine option 1 and option 2
- 22. • Problems • unicorn: Rack HTTP server for fast clients • static assets • logging: default ruby syslog library uses syscall (needs local syslog daemon) • Solution • run 3 daemons in 1 container: unicorn, nginx and rsyslogd • upstart • ENTRYPOINT [“/sbin/init”] • load ENV from /proc/1/environ • foreman Use Case: Ruby on Rails
- 23. Multi-Host • image distribution via docker registry • weighted load balancing via HAProxy • SSL termination via nginx in front of HAProxy
- 24. Registry • push and pull images • public • private • backends: local, S3, Elliptics, Google Cloud Storage, hosted
- 25. Load Balancing • HAProxy • license: GPL v2 • pool configuration stored in redis/etcd • config update • compile config files from stored configuration • upload via ssh • verify on remote hosts • replace current config with verified one • reload
- 26. HAProxy
- 27. HAProxy
- 28. Deployment Pipeline • commit triggers new image build • build suite executed with image • image is pushed to registry if tests passed • optional: start image with staging ENV settings for manual testing • start image with production ENV for last pre-flight tests • deploy image to more hosts • update load balancer (canary or green/blue) • monitor new containers/image
- 29. Deployment Pipeline Nginx HAProxy Nginx HAProxy Docker Container Container Container Docker Container Container Container Docker Container Container Container Docker Container Container Container Docker Registry Docker Build 2 push 3 pull + run 1 build 4 update4 update Route 53
- 30. Logging • host: docker host, container_id • code: image_id, revision • request: request_id, action, status_code, etag, times, calls • NOT inside containers • remote syslog (when possible) • alternative: local syslog relay inside container
- 31. Metrics • OpenTSDB • “distributed, scalable Time Series Database” • license: LGPLv2.1+ • HBase • Tags / Dimensions • from syslog via udp (StatsD “like”) • rickshaw.js for graphs • compare status codes, counts and times between actions of two revisions
- 32. OpenTSDB
- 33. Metrics
- 34. Metrics request counts by revision
- 35. Metrics
- 36. Metrics
- 37. Metrics
- 38. Docker reduces • external dependencies (“rubygems/github slow/unreliable/down”) after image is built • “did work on my machine/staging”: same OS package versions, configuration and code in all stages • unused CPU cycles • number of hosts • feedback times • time to get new host online • bottlenecks: hosts are more flexible
- 39. VS. AWS • HAProxy much more flexible • multiple containers per host • balancing weights • faster build process • faster deployments • instance flexibility
- 40. Resources • docker.io • opentsdb.net • haproxy.1wt.eu • continuousdelivery.com • chadfowler.com/blog/2013/06/23/immutable- deployments/ • 12factor.net
- 41. Questions?!?
- 42. Thank you!