blog

12021-06-04

SEAN K.H. LIAO

ci/cd

Conitnuos Integration: roughly translates to build early, build often, and never deviate far from mainline, on the idea the many small changes are easier to swallow than a single big one.

Continuous Deployment: You have a steady stream of build artifacts, now what? You deploy and run them often, hopefully you spot breaking changes early and fix them sooner rather than later.

GitOps: There's desired state and reality, and gitops is about tracking desired state in git so you have a clear history and it can be reviewed before being rolled out. Usually/hopefully you have you have automated tooling to do the rollout part, so humans don't get the chance to make mistakes.

implementation questions

Theory sounds all nice and good, but how do you implement this?

linear single stage

This is easy to understand: as part of your CI pipeline, you run the tasks:

to redeploy, an existing release, either rerun part or all of the pipeline (dummy commits anyone?) or do it manually. A critical weak point when it comes to gitops with this setup is there's no good way for automation to record the deploy version into git for a single repo/branch. If you use commit hash (or any other generated identifier), modifying the deployment manifests to record it changes the identfier.

Advantages: easy to understand and implement.

Disadvantages: sometime after deploy reality may go out of sync with what you pushed, less direct control over deploy timing, gitops versioning problem.

linear 2 stage

Like the previous version but split into 2 stages:

and

This is usually done with 2 repos, it's your choice of placing the canonical version of deployment manifests with code and generating the entire thing into a separate repo or having the canonical version in the separate repo and just updating the artifact id every build.

The deploy stage is just another pipeline.

Advantages: this solves most of the problems from linear single stage: no more versioning problems and you have better control over deploy timing.

Disadvantages: desired state and reality can still go out of sync.

linear/declarative mix

This builds on the 2 stage setup, but instead of having a pipeline for the second stage, you have a controller watch both desired state and reality and take the actions (redeploying) to keep them in sync. No more configuration drift!

I've yet to see a declarative approach for the build stage, but i guess it's not really necessary as artifacts don't drift?