15.266 14.9.26 Pipeline de despliegue
DPL-01. Resolve source commit and locked dependencies. DPL-02. Run unit, schema and property tests. DPL-03. Build application artifacts and OCI images. DPL-04. Generate SBOM and provenance. DPL-05. Run vulnerability and secret scans. DPL-06. Build/package model manifests and portable model artifacts. DPL-07. Run deterministic fixture analysis and golden-output comparison. DPL-08. Run benchmark subset on compatible hardware. DPL-09. Sign all release artifacts. DPL-10. Assemble offline update bundle and manifest. DPL-11. Verify bundle on a clean verifier. DPL-12. Stage to inactive slot/cache on the edge. DPL-13. Check disk, power, runtime, model and schema compatibility. DPL-14. Backup metadata and current deployment receipt. DPL-15. Activate services in dependency order. DPL-16. Run health, capture, inference and sample-result smoke tests. DPL-17. Observe canary window and compare SLI/profile baseline. DPL-18. Finalize activation or rollback and persist receipt.
La matriz de compatibilidad debe fijar, como mínimo: Jetson Linux/JetPack, CUDA, TensorRT, DepthAI, driver/container runtime, Python/C++, modelos, plugins y schema release. Actualizar sólo un componente fuera de la matriz crea un perfil no validado.