substrate: provisioners beyond docker and podman

[ajbouh]

This is a place to take notes on what we need from a provisioner driver implementation.

Substrate uses provisioners to:

• launch services from a string identifier (e.g. container image name)
• bind mount additional folders at launch
• check and watch service status (e.g. starting, running, exited)
• stop services
• configure network access for services (we don't yet have a very strong policy on this)

The string identifiers we use for services need to fully specify all files we need for a service. If at all possible, it should pin to a specific version.

Podman on Linux is our best-supported provisioner. Docker is almost as good, though we only test it on macOS.

Provisioner Backend	OS	Tested	GPU access	Service Primitive	Image Definition	Example Image Reference
Docker	Linux	❌	✅	container	Dockerfile	ghcr.io/ajbouh/substrate:substrate-faster-whisper
Podman	Linux	✅	✅	container	Dockerfile	ghcr.io/ajbouh/substrate:substrate-faster-whisper
Docker Desktop	macOS	✅	❌	container	Dockerfile	ghcr.io/ajbouh/substrate:substrate-faster-whisper

Unfortunately, Docker on macOS does not provide access to any accelerators. Given that Macs are some of the fastest machines available for machine learning inference, if we can easily support them then we'd like to.

These are the questions we have:

Provisioner Backend	OS	Tested	GPU access	Service Primitive	Image Definition	Example Image Reference
custom	macOS	✅	✅	?	?	?

One possible option would be host processes running natively built and packaged with nixpkgs

Provisioner Backend	OS	Tested	GPU access	Service Primitive	Image Definition	Example Image Reference
custom	macOS	✅	✅	OS process	nixexpr	/nix/store/8m3wjb23sfbjpjsj4l82b4zh9xnw62hh-faster-whisper

There are still other questions to answer, but writing this up now as a starting point for if/when it matters