v0.8 — Adoption Track

[korbonits]

Context

Through v0.7, sheaf's focus has been capabilities: 20+ model types, Ray Serve + Modal + offline batch paths, batching/caching/streaming/Feast, logging/metrics/tracing. The remaining gap is friction — installing, deploying, and evaluating sheaf is harder than it needs to be for new users. v0.8 shifts from capabilities to adoption.

Scope

1. Prebuilt Docker images

• Multi-stage uv sync Dockerfile (core + per-extra variants)
• GHA workflow to build + push to ghcr.io/korbonits/sheaf on tag
• Variants: :0.8.0, :0.8.0-time-series, :0.8.0-audio, :0.8.0-vision, :0.8.0-all
• One-liner quickstart in README

2. Docs site (mkdocs-material → GitHub Pages)

• Scaffolding + theme + GHA deploy
• Quickstart (lifted from examples/)
• Model types page (per-type quickstart + extras flag table)
• Concepts: batching, caching, Feast, streaming, offline batch
• Deployment guides: Docker, Modal, k8s
• Auto-generated API reference from Pydantic contracts

3. Benchmarks (bench/)

• Reproducibility script + hardware spec committed
• Throughput + p50/p99 latency for Chronos2, OpenCLIP, Whisper
• Baselines: raw Ray Serve, BentoML, TorchServe
• Results table in docs

4. Helm chart (charts/sheaf/)

• Scaffolding + values (model specs, resources, Feast config, Prometheus scrape annotations)
• Publish to a gh-pages Helm repo
• Example: Chronos2 deployment with HPA

5. CLI DX (stretch)

• sheaf serve <backend> --model <id> — run a ModelServer without writing a spec file
• Start with time-series + tabular; expand per demand

Non-goals

• New model backends (capabilities track)
• New feature store integrations (separate issue)
• ONNX / Triton interop

Risks

• CI bandwidth — GPU-extra Docker builds are large; consider nightly rather than per-PR
• Docs drift — auto-generation covers API reference; prose sections will rot without discipline
• Benchmark controversy — framework comparisons invite pushback; lean on reproducibility + honesty about where sheaf loses

Success criteria

• New user gets Chronos2 running end-to-end in <5 minutes from `docker run`
• Docs site live at a stable URL
• ≥1 published benchmark comparing against ≥2 alternatives
• Helm chart installable via `helm install`