I've been building a Firecracker orchestrator because I wanted branch previews that behave more like real deployments than a single throwaway container. For a configured repository, a push or pull request starts the provider's webhook and pipeline flow. CI builds a compatible OCI image, converts it to a pinned root filesystem, and a separate publisher stores that artifact in S3-compatible storage.
The controller places each service in its own microVM, potentially across several workers. Web, PostgreSQL, Redis, and search communicate over a private network according to the manifest's allowed connections. Caddy exposes the web service through a public HTTPS URL. A Web update can switch to a new release while the database VM keeps running.
The preview path works end to end. I'm now testing what longer-lived stacks need for backup and recovery after a worker fails.
commit c092be6