Slice SLA Observe-Only Assurance (Design Note)

Non-normative prototype. Slice management & assurance align with TS 28.530 / TS 28.541 and the underlying admission state is TS 29.536 (NSAC), but this loop is a research shape for education — observe-only, no remediation, and no frozen closed-loop-assurance Stage-3 applies. Tracking issue: #27.

What it does

A feature-gated (sla-observe, off by default) observer task inside nextgcore-nsacfd evaluates a configured per-slice SLA descriptor against the live per-slice KPIs the NSACF already maintains — the admitted-UE and PDU-session utilization percentages derived from its TS 29.536 admission sets — and emits an observable signal on breach:

• a log::warn! breach transition line (and info on clear) — transitions only, no per-tick spam;
• Prometheus-text metrics (three gauges plus one counter): nextgcore_slice_sla_ue_utilization_pct, nextgcore_slice_sla_pdu_utilization_pct, nextgcore_slice_sla_breached (0/1), nextgcore_slice_sla_breach_transitions_total — served on /metrics only if metrics_port opts in (reusing the generic nextgcore_metrics::nes_energy::serve_metrics helper; by default no listener is bound).

Admission behavior is untouched: the observer only reads quotas_snapshot() copies under short locks.

Config (nsacf.yaml, off by default)

nsacf:
  sla:
    enabled: true              # default false
    poll_interval_secs: 30
    metrics_port: 9090         # 0/absent = no listener (default)
    slices:
      - sst: 1
        sd: "000001"           # optional, hex like slice_quotas.sd
        max_ue_utilization_pct: 80.0
        max_pdu_utilization_pct: 80.0

The sla block is captured as a raw YAML value in the typed config (same hardening as pcfd's intent block): a malformed block warns and disables the observer but can never break the slice_quotas provisioning parse — which would otherwise reject every admission. Invalid slice entries (missing sst, bad hex sd, no positive threshold) are skipped with a warning. Slices without a configured SLA are not evaluated (opt-in per slice); a configured SLA without a provisioned quota is skipped at debug.

Design choices

• KPI source: NSACF admission counters. Per the issue, the first KPI is what the NSACF already measures (ue_utilization/pdu_utilization). Slices whose quota is uncapped (max_ues absent → u64::MAX) evaluate to ~0% utilization — a threshold on an uncapped count is meaningless, so configure SLAs only for capped slices.
• Why not nextgcore_metrics::ai_native::SlaMonitor: its SlaKpi set (latency/loss/throughput/availability…) contains nothing nsacfd can measure today; mapping admission utilization onto one of them would mislabel the signal. When NWDAF-fed per-slice KPIs exist (NWDAF currently serves NF_LOAD only), this observer is the adoption point.
• Observe-only: breach → signal, nothing else. Automatic remediation (EAC activation, quota tuning) is explicitly out of scope here.

Verification

cargo build -p nextgcore-nsacfd                           # default unchanged
cargo test  -p nextgcore-nsacfd --features sla-observe    # incl. breach/clear + lenient-config tests