Skip to content

Latest commit

 

History

History
89 lines (68 loc) · 5 KB

roadmap.md

File metadata and controls

89 lines (68 loc) · 5 KB

Deferred fleet capabilities

This file records deliberate V1 boundaries. An item here is not an accepted workaround and does not weaken the current worker contract. Each capability requires a separate design review, threat model, compatibility proof, and a first rolling-host rollout under production traffic with the documented health checklist before a workflow may depend on it.

Standing non-goals

The two-host fleet does not need Kubernetes or an Actions Runner Controller deployment. It also does not accept privileged workers, host Docker-socket passthrough, a persistent native Windows runner on a daily-use host, Docker Desktop engine switching, mutable latest deployment, or an external watchdog that cancels and replays workflow side effects. Reconsidering one of these is an architecture change, not routine configuration.

Isolated Linux VM backend

Testcontainers, service containers, job containers, Docker actions, and other Docker-daemon workloads remain GitHub-hosted. A future backend may run them in a disposable Linux VM whose Docker daemon, filesystem, and runner identity are destroyed after one job. The host Docker socket must never be mounted or proxied into a worker. The backend must preserve JIT registration, external diagnostics, resource admission, and graceful drain. The audited hosted-only cutoff it used to have to preserve is gone: ci-perf Phase 7 retired the selector, so there is no eligibility to recompute on a rerun and no routing variable to cut over with. A backend that cannot serve a job leaves that job queued on the fleet label until it can. No hostedExceptionReasons member expresses a fleet outage, so there is no consumer-side escape; a backend whose unavailability must not block CI needs that case designed in, not waived per repository.

This work is motivated by GitHub's warning that self-hosted runners can be persistently compromised by workflow code and Docker's warning that daemon access is equivalent to highly privileged host access:

Ephemeral Windows VM backend

Windows-only tests remain on windows-2025. A future Hyper-V adapter may create a Windows VM and disposable disks from a verified patched base, inject one JIT runner configuration, execute one job, export diagnostics, and destroy the VM and its disposable disks. Image licensing, Windows Update, base-image integrity, host/guest boundaries, and unattended failure recovery must be proven before adoption. Any alternative reset mechanism remains out of scope until Microsoft documents it and a separate threat model proves equivalent isolation.

V1 will not register a persistent native Windows runner on a daily-use host and will not switch Docker Desktop between Linux and Windows container engines.

Stronger worker network isolation

V1 workers require outbound network access for GitHub Actions, source checkout, tool setup, packages, caches, and artifacts. A future network adapter may add audited DNS/egress policy or a proxy, but it must measure action compatibility, fail visibly, keep controller credentials unreachable, and avoid claiming an isolation guarantee the Docker network does not enforce.

No future network layer may justify privileged workers, host devices, host paths, or Docker-socket access.

Optional OTLP export

The controller now provides optional OTLP trace and metric export, disabled by default, alongside JSON Lines and archived runner diagnostics. Configuration, redaction, failure isolation, lifecycle behavior, and collector guidance are documented in Observability.

Cost evidence

Cost reporting may automate the same GitHub billing-usage summary used for the rollout baseline. It must keep private billing data out of this public repository and normalize by eligible completed jobs before claiming savings. There is no longer any selector spend to separate from workload spend: the selector's two-minute ubuntu-slim control job is retired with the selector itself.