Limitations, Migration Triggers & Future Adapters

§16.8 Known Limitations (MVP)

Limitation	Impact	Mitigation
Depot builds run in AWS us-east-1 / eu-central-1 only	No non-US/EU data residency	Flag for Enterprise deals requiring data residency; revisit with self-hosted BuildKit migration
Default 16 CPU / 32 GB shared builder per project	Concurrent builds share resources	Enable per-project autoscaling from day 1 (Depot setting, not plan-gated)
Railpack framework coverage evolves with Railway	New frameworks supported when Railway adds them; edge cases require Dockerfile fallback	Dockerfile fallback path always available; contribute upstream when a gap blocks a Starform customer
Cache scoped per project, not global	Slightly lower hit rates than global shared cache	Acceptable at early scale; revisit post-MVP
Depot Docker build pricing at $0.04/min (public)	Build-minute gross margin compressed to ~20% at $0.05/min sell price	Negotiate Business-plan rate before launch; port abstraction enables self-hosted migration if negotiation fails
Depot cache storage ($0.20/GB/mo beyond plan allocation)	Infrastructure overhead scales with active customer count	Monitor cache utilization per workspace; garbage-collect stale caches; factor into capacity planning
Railpack maintained by Railway (direct competitor)	License or feature-gating changes could force migration	MIT license enables forking; vendor pinned version; treat as insurance-backed dependency

§16.9 Migration Triggers (to Self-Hosted BuildKit)

Any one of the following justifies writing the second BuildService adapter:

1. Depot monthly spend exceeds ~$5,000 — at this scale, 3–4 months of engineering time to migrate pays back within 6 months
2. Depot Business-plan negotiation does not yield rates below $0.02/min at Starform's projected volumes
3. Enterprise customer requires data residency Depot cannot provide
4. Reliability incidents on Depot's side damage Starform's SLA
5. A required feature (GPU builds, custom frontends beyond what LLB allows, etc.) is not on Depot's roadmap

Until one of these triggers fires, running our own build infrastructure is net-negative. The port abstraction ensures this remains a reversible decision. Trigger #1 is the most likely to fire first — at 500+ active customers averaging 80 build-minutes/month, Depot spend reaches ~$2,000/mo at list rates. Negotiated rates delay this trigger.

§16.10 Future Adapters

• adapter/buildkit/ — Self-hosted BuildKit on DOKS with rootless multi-tenant configuration
• adapter/codebuild/ — AWS CodeBuild for Starform-operated AWS regions (multi-cloud geographic expansion, §12 — not customer BYOC, which is out of scope)
• adapter/gcb/ — Google Cloud Build for Starform-operated GCP regions (multi-cloud geographic expansion, §12 — not customer BYOC)

---

Cross-references

The BuildService port that keeps the swap reversible → §16.3 · build-minute economics → §16.7 · region registry where a future adapter is wired → Starbase §12 · BYOC's out-of-scope status → Foundations §37.