Practicing Deployment

Deployment talk from OSCON 2013

Transcript

1. Practicing Deployment Laura Thomson laura@mozilla.com @lxt 1
2. Disclaimers Not about tools Not prescriptive Recognize where you are, and where you want to be 2
3. Models 3
4. Maturity model (after Capability Maturity Model, CMU) 1. Initial: “chaotic”, “individual heroics” 2. Repeatable: documented 3. Defined: standard process, some tuning 4. Managed: measured 5. Optimizing: continual improvement, innovation 4
5. Initial Startup phase of many projects Long term Push code whenever you feel like it Devs push code Not a lot of tests, automation, or verification 5
6. Repeatable Often after a 1.0, first non-beta ship, or first ship with a significant number of users Some kind of documented/known process Push when a feature is done: less often than initially, typically 6
7. Defined Procedural documentation Start of automation Often done by a sysadmin 7
8. Managed Automation Tools: packaging Verification post-push Measurement: How often do we push? How long does it take? How did that push affect performance? 8
9. Optimized Take the drama out of deployment Often - not essentially - continuous deployment Typically a lot of test automation Lightweight 9
10. How much do we ship? (Size of a release) Start with per-patch pushes Move to features Then to releases Then back to features The back to per-patch pushes 10
11. Per-patch Features Releases Features 11
12. Velocity models (Frequency of a release) Critical mass Single hard deadline Train model Continuous deployment 12
13. Critical mass “enough stuff to release” MVP smallest quantum with user value 13
14. Single hard deadline Support for X by date Y Shipping to a marketing plan Hard deadlines are hard 14
15. Train model Release e.g. every Wednesday Whatever’s ready to ship, ships Anything else catches the next train 15
16. Continuous deployment Ship each change as soon as it’s done Continuous is kind of a misnomer; deployment is discrete 16
17. Tools and practices 17
18. Source control Stable vs unstable Branch per bug, branch per feature “git flow” is overkill, but you need a process If it’s not per-patch-push, tag what you push Open source needs ESRs even if you’re high velocity 18
19. Dev Envs Dev’s laptop is a horrible environment VMs can be hard to maintain Development databases are hard: fake data, minidbs Development API sandbox Lightweight set up and tear down VMs “Development” staging server (unstable) “Try” servers for branches 19
20. Staging Staging environment MUST REFLECT PRODUCTION Same versions, same proportions: a scale model Realistic traffic and load (scale) Staging must be monitored Staging must have managed configuration 20
21. One Box Fail Staging needs to be more than one box If you have multiple databases or webheads or whatever in prod...you need that in staging 21
22. Continuous Integration Build-on-commit VM-per-build Leeroy/Travis (PR automation) Run all unit tests (Auto) push build to staging Run more tests (acceptance/UI) 22
23. Testing Unit tests: run locally, run on build Acceptance/User tests: run against browser (Selenium, humans) Load test: how does it perform under prod load? Smoke test: what’s the maximum load we can support with this build? 23
24. Deployment tools It doesn’t really matter what you use Automate it Do it the same way in staging and production Use configuration management to deploy config changes and manage your platform...the same way in staging and production 24
25. QA Feature tests on unstable Full tests on stage Full tests on production (verification) 25
26. Measurement Monitoring Performance testing Instrument, instrument, instrument Is it actually possible to have too much data? (Hint: yes. But only if no insight) 26
27. Postmortems What went right What went wrong Blameless: scapegoats only hurt you 27
28. When things go wrong 28
30. Quantum of deployment (via Erik Kastner) “What’s the smallest number of steps, with the smallest number of people and the smallest amount of ceremony required to get new code running on your servers?” http://codeascraft.etsy.com/2010/05/20/quantum-of-deployment/, 30
31. Chemspills Even if you have heavyweight/non-automated deployments, what does a chemspill look like? 31
32. THIS IS NOT A DRILL 32
33. Fail forward Fail forward: the premise that Mean Time To Repair is the key measure, not MTBF 33
34. Fail Sometimes you can’t fail forward Example: intractable/unforeseen performance problem, hardware failures, datacenter migrations Hit upper time limit (failing forward is taking too long) 34
35. Rollback Going back to the last known good Having a known process for rollback is just as important as having a known process for deployment Practice rollbacks 35
36. Decision points When shipping something new, define some rules and decision points If it passes this test/performance criteria we’ll ship it If these things go wrong we’ll roll back Make these rules beforehand, while heads are calm 36
37. Feature switches A nicer alternative to rollback Turn a feature on for a subset of users: beta users, developers, n% of users Turn it on for everybody Turn things off if you’re having problems or unexpected load: “load shedding” 37
38. Continuous Deployment 38
39. What is CD? Total misnomer Not continuous, discrete Automated not automatic, generally Intention is push-per-change Usually driven by a Big Red Button 39
40. Technical recommendations Continuous integration with build-on-commit Tests with good coverage, and a good feel for the holes in coverage A staging environment that reflects production Managed configuration Scripted single button deployment to a large number of machines 40
41. People and process High levels of trust Realistic risk assessment and tolerance Excellent code review Excellent source code management Tracking, trending, monitoring 41
42. Testing vs monitoring Run tests against production Continuous testing = one kind of monitoring Testing is an important monitor You need other monitors You need tests too 42
43. You should build the capability for continuous deployment even if you never intend to do continuous deployment. 43
44. The only way to get good at deployment is to deploy a lot. 44
45. Questions? laura@mozilla.com 45