OpenShell Drivers

[linear]

OpenShell drivers will enable first and third parties to support arbitrary deployment scenarios for OpenShell. The intent is that the core OpenShell team maintains a small deployment footprint and third parties may develop their own drivers that at allow OpenShell deployments across other container and native OS environments. See attached RFC for further details.

What's left

Goal: in-tree and out-of-tree drivers should be interacted w/ from the gateway the same over a grpc api

Remaining to consolidate

• Readiness
• Driver ability to configure gateway endpoint
• Shutdown and restart mechanics (probably need a new grpc method for this)

Other items

• Conformance tests should assert that drivers are only using public api contracts
• Land the PR to support compiling out drivers: feat(server): conditional compilation of compute drivers #2163
• Add an out-of-tree driver in examples/.
• Add docs to extensibility section.

Exit conditions: each driver is built and tested as it's own binary the gateway is built without any drivers and conformance tests pass.

[blkerby]

I'm curious about one thing that I didn't see mentioned in the RFC. Thinking about the CLI commands for managing a gateway's lifecycle (openshell gateway start/stop/destroy), it seems like these would need to do different things depending on how the gateway is deployed. I'm wondering if it means there also needs to be some sort of "gateway management" driver inside the CLI, conceptually similar to the various drivers inside the gateway (compute, credentials, etc.)?

For some example deployment scenarios it might roughly look like this:

Gateway deployment	gateway start	gateway stop	gateway destroy
Openshell-managed Kubernetes cluster	Create and populate cluster	Stop cluster nodes	Destroy cluster
Existing/shared Kubernetes cluster	Create and populate openshell namespace	Scale Deployments, StatefulSets to zero	Delete openshell namespace
Snap (local)	Configure snap to enable gateway daemon	Configure snap to disable gateway daemon	Configure snap to disable gateway daemon, and trigger snap hook to clean-up associated data

In this scenario, I imagine you would want to be able to specify the default "gateway management driver" as a CLI argument or environment variable. This can help with keeping the CLI ergonomic so that, for example, if running the CLI through the Snap, it could know to use the Snap driver to start the gateway by default; this way the first invocation of openshell sandbox create can just work, automatically starting the appropriate type of gateway without any extra setup.

[drew]

We've removed the start/stop/destroy commands from the openshell cli. Instead we rely on package managers to setup the gateway itself as a service on the host. Once the service has been registered, a driver can perform whatever setup is necessary.

[zyga]

@drew ideally openshell would have a concept of both user config and system config. Package mamabver could easily influence system config at install time.

[drew]

Definitely. We have a proposal to add a proper gateway config.toml file here, #951. As part of that we should decouple a gateway and a user config. Feel free to leave feedback on that RFC. I expect we start implementing that soon.