Shipping Code at Saltside: 2 Years on Docker

Transcript

1. Shipping Code at Saltside: 2 Years on Docker Adam Hawkins

   - DevopsDays India - BLR - 2016-11-04

2. Hi, I’m Adam • Joined Saltside in 2013 • Lead

   SRE Team @ Saltside • @adman65 • http://blog.slashdeploy.com

3. This Talk • Saltside: A Curious Case • Our production

   infrastructure • Why we are the way we are (for better & worse) • Saltside’s future • How you can make a better future (for yourself and your team)

4. Saltside Org & Tech

5. One Product, One Code Base, Four Markets

6. None

7. Quick Stats • 2 Product Development Offices. • Hundreds global

   employees, ~50 in PD • 1 Product, 4 Markets • 2 AWS Regions • SoA; 15+ production services • 5 production Languages • ~300 production containers • Mix of cross functional teams; 4 People in SRE Team (pssst. We need SRE!)

8. Q3 2014 • Rewrite entire product and infrastructure from scratch

   starts (Never do this; no seriously, never do this!) • Docker 1.0 Released • Introduce market specific infrastructure • Shift to SoA • I am the quasi-architect for new system

9. Why Docker? • Give engineers the freedom to choose best

   stack stack for the problem • Infrastructure standardization • dev & production parity. The development environment is make, docker, docker-compose, and few others

10. 2014 != 2016 • Mature container orchestration tools … lol

    no • official images … lol no • docker-machine … lol no • established best practices … lol no

11. Oh my Times have changed

12. Rolling Our Own: Apollo • configuration file driven (similar to

    docker-compose) • Set’s horizontal & vertical scales per market/container • It’s dynamic CloudFormation all the way down • One EC2 instances runs one Docker container • Every EC2 instance is behind an ELB • Zero down time deploys via HAproxy • Instances poll S3 for which image/tag to use; change containers where appropriate

13. apollo • $ apollo deploy -e production -m bikroy -t

    823714 • $ apollo update -e production -m bikroy

14. 2 Years Later • 1 container per 1 EC2 instance

    is damn expensive • Things are stable and work • No blue/green, canary, or alternate deployment strategies • No indication of deploy status (started, failed, rollout percentage etc) • Time to bootstrap new services = Days • Time to each new employees apollo = Months • Sunset

15. 2014 != 2016 We have Kubernetes. Punt successful!

16. Kubernetes PoC • One cluster per-region; multiple markets per cluster

    • Decrease costs • Increase reliability • Move to maintained and active open source project instead of end-of-life’d private internal tools • Increase velocity

17. How can we have environments per topic branch?

18. Issues • Cannot do with apollo; only support fixed environment

    names. Also too slow and expensive • How to give client developers a functioning API platform to develop against? • How to give QA access to N number of service configurations? • How give engineers a place to experiment outside production? • How to save everyone from configuring 15+ services?

19. Sandbox Infinite Environments with Docker Compose & Docker Machine

20. • sandbox create # create a new environment • sandbox

    sync # Pull image tags from production • sandbox reset # wipe data and boot everything • sandbox logs # Get logs from containers • sandbox tag foo-service bar # change tags • sandbox dev foo-service # Build images locally

21. Take Aways • Prefer Kubernetes/Mesos/DCOS instead of rolling your own

    • Prefer docker-compose over manual docker commands • Google Container Engine is the easiest way to get production ready container infrastructure • Package up your distributed application instead of making engineers manage it themselves • Include log and metrics systems in your budgets from t-zero • Create deployment APIs instead of CLIs • Distribute internal tools as Docker images • Prefer containerized workflows over host level dependencies • Prefer one Dockerfile per project • Prefer the official Docker images over internally maintained base images

22. Adjust Madi