System · 2026
Cloning production without the blast radius
STEM stamps out production-shaped database environments — masked, permissioned, and disposable by default.
Fallstudien sind auf Englisch verfasst.
Problem
Staging databases lie. Every team says “don't test against production,” and every team eventually does — because the alternative is stale, hand-fed data that fails to reproduce anything real.
Ansatz
One pipeline, end to end: a GitHub App triggers Aurora clones, masks sensitive data in-flight, scopes IAM per environment, meters everything, and destroys the environment when its work is done.
Stand
Carried from idea to public demo — GitHub App, cloning, masking, IAM, cleanup, telemetry, and dashboard, shipped as one system.
The noticing.
“Don't test against production” is the most-repeated and least-followed rule in infrastructure. The failure isn't discipline. It's that the rule offers no artifact: teams break it because nothing gives them production-shaped data without production consequences.
The fix isn't another warning in the runbook. It's making the safe path the lazy path — an environment that looks exactly like production, appears on demand inside the pull request, and cannot leak what it holds.
Decisions.
Masking inside the pipeline, not after it
Post-hoc masking means a window where real data exists in a weaker perimeter. In STEM, an Aurora clone is never reachable before the masking pass completes — the unsafe state is unrepresentable in the pipeline's state machine.
Disposable by default
Environments get a TTL reaper at birth, not at cleanup review. Long-lived clones are the failure mode — they drift, they leak, they become the new staging. Anything worth keeping must argue for renewal; nothing is immortal by accident.
The PR is the control plane
No new portal, no new CLI to learn. STEM runs as a GitHub App — if the workflow doesn't live where the developer already is, it becomes shelfware. A scoped IAM layer and a dashboard make every clone accountable — who asked, what it cost, when it vanished.
State.
Scope
end-to-end
Shipped
public demo
Surfaces
app + dashboard
Weiter
Hardening the masking rule engine and publishing the write-up — the system taught more about IAM edge cases than any tutorial ever did.