Pipeline tooling evaluation
Evaluated 2026-11-20. Recommendation: keep native GitHub Actions scheduling, Docker BuildKit execution and the small standard-library Go CLI for now. This recommendation can change when measured bottlenecks justify a new engine.
| Option | Fit | Tradeoff |
|---|---|---|
| Actions + BuildKit + Go | Keeps individual checks, target dependency jobs, existing runner setup and the same local Make commands | Must generate native Actions job definitions ahead of submission |
| Dagger Go SDK | Typed container operations, caching, and local/CI execution in Go | Adds its engine and SDK; GitHub check aggregation, changed-target selection, registry retention and trust policy still need integration |
| Earthly | BuildKit-based target builds and caching | Adds an Earthfile/runner model alongside the already implemented Dockerfile/Bake path |
| Buildkite dynamic pipelines | Adds steps at runtime with native dependency scheduling | A CI platform change rather than a small SDK adoption |
Dagger would be the first SDK to prototype if container orchestration becomes the main source of duplicated logic. Prototype one target and measure cold/warm build latency, source-build reuse between publication/tests, multiarchitecture runners, and per-target GitHub checks before committing to a migration. Do not wrap the existing pipeline in a second scheduler simply to use an SDK.
Buildkite becomes relevant if Actions' static job graph is the limiting factor. Earthly deserves consideration if replacing Bake/recipe build orchestration rather than adding a new layer. None of these tools creates upstream source checksums, eliminates native packaging work, or supplies provider-independent registry GC.