Known Limitations - rsys-bshankar/ai-ran-smo GitHub Wiki
Known limitations (0.3.0)
- Reads are open to every valid token, apart from a short deny-list of internal-only routes. Per-tenant and per-region read scoping is not built.
- The compose sample rApps are static services with module identities, not deployed through
POST /instances, so manifest limits and the per-instance kill switch do not reach them there. - Credential delivery to a workload exists only as a Kubernetes Secret (Helm,
rappCredentials.delivery=kubernetes, off by default). NFO does not yet reference that Secret when it deploys the workload, and Docker Compose has no equivalent. - The audit chain has no GUI view, and does not protect against someone who can write the database and recompute every hash: keep an exported head (
python -m smo_shared.audit export) somewhere else. - The Helm chart is proved on kind from images built from the repository. More than one replica of a module, PodDisruptionBudget and autoscaling are in the chart but off and untested under load; the bundled Postgres is one pod with no backup or HA; NetworkPolicy is rendered but its denied paths are not tested.
- The seeded KPIs use this build's own counters, not the TS 28.554 definitions. KPI and topology names are this build's own, not the TEIV model. The worker exports no metrics.
- The O-RAN WG4 YANG models are not bundled (needs the licensed files).
- During a rolling upgrade from 0.2.0, a module still on 0.2.0 does not forward the originating rApp, so the DME safeguard fix applies once every module is on 0.3.0.
The full list of open work with reasons: smo/OPEN_ITEMS.md. The record of what was decided and built: smo/HISTORY.md.