Deployments should be

Fast

• 10 minutes is too long

Small

• Ideally a single commit
• You must be aware of the whole change

Easy

• As little human involvement as possible
• Minimal context switching
• Simple to understand

We believe this because

• There are less things to break
• You can focus on one thing at a time
• The changes are fresh in everyone's memory
• Devs should be focused on the app
• Each project should be easy to learn
• Nobody should be afraid to deploy

We can back it up too

• We have 30 separate microservices
• 88 Docker machines running across 15 workers
• Around 7 deploys to prod every working day
• 4 emergency rollbacks in 1.5 years
• Maintained by myself and a single developer

How did we do it?

You'll need:

• A microservices architecture (kinda optional)
• A git repo that can run triggers on push
• A build agent that allows several steps
• A packaging format that tracks dependencies
• Something to manage your cluster of machines

Deployment steps

1. Write some code
2. Push to the git repo
3. Application is built
4. Automated tests are run
5. Application is packaged
6. Deployment to staging
7. Test in staging
8. Single click to approve for prod
9. Deployment to production

3. Application is built

• Happens on trigger
• Not all apps need to be built
• We have a build.sh script in each repo
• We wanted to be consistent across all our services
• You could also use Makefiles

4. Automated tests are run

• We actually do this at the end of build.sh
• Again consistency is key
• Don't do end to end testing here
• We do end to end testing continuously against staging and prod

5. Application is packaged

• We build docker containers with a Dockerfile
• Push it to an internal docker repository on GCS
• Anyone should be able to pull this for testing
• You could also use RPMs, Debs, and others
• If you do make per app repos, to make repo metadata creation time short

6. Deployment to staging

• We use Apache Mesos with Marathon
• For us we essentially assemble a JSON file from a template
• Then submit it to Marathon using curl
• You can use these with or without docker

8. Single click approval

• At this point our change is well tested
• Trigger the deployment to staging
• Most build agents have this, or let you build it easily
• Audit who clicks this!

9. Deployment to production

• Should be very similar to prod right?
• Remember, consistency!
• We use Marathon and Mesos in prod too

Thanks!

LIFX is seeking a Cloud Engineer and Firmware Engineer to join our Melbourne team.

 

 

 

 

 

 

 

Either see this link, email me at daniel.hall@lifx.com or come talk to me at LCA all week.

See: https://goo.gl/0KOX1C and https://goo.gl/YY30Gh

This Talk

Will be available at:

• https://slides.com/smarthall/from-commit-to-cloud
• I'll post this URL on my Twitter (@smarthall)