ADR-014 — MLOps Governance Layer (Model ABAC + 배포 + 드리프트)
| 항목 | 값 |
|---|---|
| Status | Proposed (2026-05-25 — #995 design step) |
| Related Epic | #995 |
| Related ADR | ADR-003 (ModelDeployment 명명), ADR-005 (KAI Scheduler) |
한 줄 결정: MLflow Model Registry 위에 ABAC 거버넌스 계층 — 모델 등록/배포/드리프트 측정/사용 audit 을 모두 PG
ModelGrant+ModelDeployment(ADR-003) 로 통합.
결정
- ModelGrant:
(grantee_type, grantee_name, model_name, version, privilege). privilege =VIEW/DEPLOY/ADMIN. - Deployment gate: 배포 시
ModelGrant.DEPLOY검증. KServe RawDeployment + KAI Scheduler (ADR-005). - Drift detection: Evidently AI sidecar (옵션) — feature distribution diff > threshold 시 Slack alert.
- Usage audit: 모든 inference call 이
ModelCallLog적재 (model_name, version, caller, latency, input_features_hash). - Prerequisite: ADR-005 (KAI Scheduler), #1018 (Workspace scope).
비목표
- AutoML 자동 재학습 별도. Federated learning 별도.