CI CD Best Practices

CI/CD Best Practices

Runner placement

Choose the cheapest runner that matches the job's real bottleneck.

ubuntu-24.04

Use for I/O-bound or orchestration-heavy jobs:

• git / GitHub API orchestration
• AWS/Terraform apply jobs that do not compile code
• artifact upload/download
• release-please
• branch promotion and audit tagging
• Copybara sync jobs
• IAM preflight / policy simulation
• ECS force deploy
• link checking
• small lint/validation jobs that do not build or test significant code
• self-hosted runner provision / cleanup control jobs

blacksmith-4vcpu-ubuntu-2404

Use for compute-bound jobs:

• Docker builds
• Rust / Go / Python / TypeScript builds
• unit, integration, and end-to-end tests
• typechecking over real codebases
• static analysis that scans large codebases (for example Semgrep)
• Kind / KWOK / Chaos / stress workloads
• anything that spends real time compiling, bundling, or executing test suites

self-hosted AWS runners

Use only when GitHub-hosted runners cannot do the job:

• KVM / nested virtualization
• ARM64-native builds that must run on our own infrastructure
• privileged or hardware-specific workloads
• deploy jobs whose payload must run inside the target architecture/runtime

If the job is only provisioning or cleaning up the self-hosted runner, run that control job on ubuntu-24.04. Only the payload belongs on self-hosted.

Docker artifact policy

PR CI

Every deployable image must be built in CI with push: false.

Rules:

• CD must never be the first place a Dockerfile runs
• do not push unreviewed application images to ECR from pull requests
• verify the exact Dockerfile, build context, and target platform the CD workflow uses
• make CI gates and CD path filters cover every file copied into that Docker build context
• if the Dockerfile depends on private ECR base images, log in only for pull access; still keep push: false

merge / branch CD

After merge, CD may build and publish the application image for the merged commit.

Current repo rule:

• PRs verify builds
• merged branches publish images

If you later refactor a service toward true build-once/deploy-later, keep the same invariant: no unreviewed artifact is written to ECR.

Workflow structure

Prefer these shapes:

1. verify-build in CI
2. publish in CD
3. deploy/apply in CD

Split build from deploy when runner classes differ. A pure Terraform/ECS apply job should not stay on Blacksmith just because the build job in the same workflow needs it.

Use reusable workflows for generic behavior. Use service-specific wrapper workflows when a service needs custom setup such as:

• private base image bootstrap
• extra build contexts
• special build arguments
• architecture-specific setup

Self-hosted runner guidance

Prefer server-side runner registration when the platform supports it.

For GitHub Actions runners:

• prefer JIT runner configuration over registration-token plus polling loops
• avoid scanning the runner list in a retry loop when the API can return the runner identity directly
• keep provisioning and cleanup logic separate from the payload job

Validation

After editing workflows:

1. Run actionlint
2. Check that cheap jobs use ubuntu-24.04
3. Check that builds/tests remain on Blacksmith or self-hosted
4. Confirm PR image verification uses push: false
5. Confirm CD still publishes only from merged branches or explicit deploy triggers

Source: Kevin-Liu-01/Agent-Machines — distributed by TomeVault.