An Ideal CI/CD System

A good CI/CD system means developer productivity. What an ideal CI/CD system looks like today.

Serverless – A CI/CD system should be serverless. There's no reason to be managing Jenkins or individual nodes. Managing CI/CD machines makes it really easy to introduce configuration and environment drift, which causes hard to trace down bugs. Container-native is usually a good choice, especially if you often deploy software inside containers, but anything that provides a real ephemeral environment is good.

On-prem (Cloud-prem) – For small projects, it's fine to use hosted CI/CD services, but for anything bigger than a toy project, you should probably be spinning up your CI/CD runners inside your own cloud account.

Minimal permissions / native IAM – Workers should authenticate in a cloud-native way (such as OIDC) and be deployed with a minimal set of permissions. While it might seem like a hassle to expand the scope of workers while projects deploy a wider variety and number of services, limiting the blast radius and separating out environments will save you from a whole class of bugs and mistakes.

Easy to debug – Easy to debug means that it's possible to run the pipeline in a meaningful way locally. It also means that logs and artifacts are easily accessible from runs.

Triggers, but not complicated ones – GitHub Actions provides a good model for running workflows automatically (from pull requests or merge events) as well as manual triggers. However, don't build a complicated business process around the manual triggers.

Code, not YAML – Following my TypeScript for Infrastructure post, I believe it's much easier to design pipelines in a full programming language rather than a configuration one. You get reusability, control flow, strong typing, and other properties that you probably want when you're describing a dependency graph.

DAG – There's been some experimentation with having a fully event-driven CI/CD system. Generally, I think it's best to mostly have a static DAG that encodes the dependency graph. Events might trigger the initial workflow and might be emitted at the competition (or failure) of the DAG, but don't fire in between.