epic: v0.6 cluster-scope NTN orchestration — consume OCUDU runtime push (!798), fan-out, lifecycle (ADR 0005)

[thc1006]

Context

Upstream OCUDU has absorbed the per-cell NTN configuration surface. Verified against ocudu/ocudu dev HEAD on 2026-06-06: polarization, cell_specific_koffset, ta_info, ta_report, ntn_ul_sync_validity_dur, epoch_time, ephemeris (ECEF / orbital / Keplerian), Doppler, feeder_link, gateway_location, sat_switch_with_resync, t_service are all exposed via du_high_ntn_config_cli11_schema.cpp + du_high_ntn_config_yaml_writer.cpp (piotr.gawlowicz's ntn_config wave: MR !658/!660/!661/!662; Keplerian propagator !836). The runtime push interface ADR 0002 anticipated has shipped upstream as MR !798 ntn_config_update_remote_command (merged 2026-05-29).

Consequence: "expose NTN config to a gNB" is now commoditized by upstream. The operator's durable value moves up a layer — from per-cell config plumbing to cluster-scope orchestration.

Scope (proposed; gated on ADR 0005 promotion Draft → Proposed)

1. Refresh ADR 0002's runtime-update path to target the merged !798 instead of the speculative WebSocket. Implement PushRuntimeUpdate against ntn_config_update_remote_command; retire replaceEphemeris for covered fields.
2. SatelliteEphemeris → multi-NTNCellConfig fan-out — one ephemeris source drives N cells.
3. Fleet lifecycle / reconciliation — NTNSlice failover across the fleet; GroundStationLifecycle binding.

Out of scope (tracked elsewhere)

• L2 orbit-aware generalization / Crossplane-style provider pattern — strategic: 2026 H2 roadmap — Crossplane-style provider pattern for NTN ecosystem #137.
• Onboard / orbital compute orchestration — strategic: OnboardWorkload CRD — vendor-neutral satellite onboard compute orchestration (Rel-19 UPF-on-satellite) #138, strategic: HostedPayload CRD — Crossplane-style provider pattern across Loft / Spire / D-Orbit / ArkEdge #139.

Notes

• Captured in ADR 0005 (Draft) "Cluster-Scope NTN Orchestration" (skeleton; the decision is deferred). Do not start implementation until ADR 0005 is promoted to Proposed.
• The first merged NTN MR that strengthens our OCUDU standing should fall out of step 1 (a real improvement found while integrating !798), not a manufactured plumbing MR.

Related: ADR 0002, ADR 0005 (draft), #49. Upstream: ocudu/ocudu !798.

[thc1006]

Upstream deep re-verification 2026-07-06 (OCUDU dev HEAD 90191bd6) — grounds this epic

A full month of upstream NTN work since the 2026-06-06 baseline (almost all piotr.gawlowicz, plus yagoda1, naposto, michael-cns). Concrete facts for the three sub-tracks:

Upstream NTN state (all merged unless noted):

• Config-manager rewrite: ntn_orbital_compute_module + ntn_orbital_state interface (!951); per-satellite context (!944, renames ntn_assistance_info→ntn_serving_cell_config); global ntn.satellites + per-cell satellite_idx (!975 — multi-satellite config is now node-local); cleanup (!943/!971); backward-epoch tolerance (!1019).
• MAC-layer NTN: DL HARQ buffer fix (!904, fixes #510); HARQ soft-buffer lifetime scaled by ntn_cs_koffset (!929).
• Bands: Rel-19 NTN bands n247/248/250/251/252/253 (!1009, TS 38.108 v19.4.0); SSB n254/30kHz fix (!997).

Sub-track 1 — runtime push target is confirmed and stable. ntn_config_update (MR !798, merged 2026-05-29). Transport = uWS server (apps/services/remote_control/remote_server.cpp), YAML key remote_control: { enabled(default false), bind_addr(127.0.0.1), port(8001) }, plaintext / no auth, envelope { "cmd": "ntn_config_update", "cells": [ … ], "common_ntn_config"?: { … } }, per-cell keyed by plmn+nci. ADR 0002's PushRuntimeUpdate should target this, not the speculative MR!411 WS it currently names. Persistence gap (reviewer fmaroszek on !862): WS commands don't survive pod restart — the gNB reloads from its config file. → CRD-as-source-of-truth that re-applies desired state on every (re)start is our value prop; captured in ADR 0005.

Sub-track 2 — SatelliteEphemeris fan-out. Found a concrete gap: the controller runs SGP4 (pkg/ephemeris/propagator.go) but the propagated ECEF is used only for pass prediction and discarded — ephemerisRef merely re-triggers a re-push of the NTNCellConfig's own static ephemeris. Filed as #176 (feed live ECEF into ntn_config_update instead). This is the concrete first slice of the fan-out.

Sub-track 3 — fleet lifecycle / cell deactivate is NOT near-term upstream. cell_lock/cell_unlock WS commands (!863, Draft) + per-cell activate/deactivate handler (!862, opened) — but !862 conceded to a redesign on 2026-06-30 (reuse existing cell_activation_routine, CU-driven UE drain, add a CU-CP logical-cell config); expected landing order !862 → !864 → !863. Payload { "cmd": "cell_lock", "cgi": { "plmn": "<str>", "nci": <uint> } } (mirrors sib_update). Track, do not bind to the current cu_cp_cell_command_handler signatures. Today an NTNCellConfig "deactivate" only happens via CR deletion → ConfigMap removal.

k_mac / #52: re-verified — the serving-cell k_mac our operator emits is still architecturally unreachable at dev HEAD (7/8 stands); k_mac static config newly exists only on neighbor + sat-switch blocks. Detail in #52.

Positioning / landscape (2026-07 sweep): "world's-first open-source K8s Operator suite for 3GPP NTN" holds under the strict 6-element scope (authenticated GitHub search resolves the 4 CRD kinds + ntn.operators.dev only to this repo). Closest adjacent = Aalyria SpaceTime (open API, proprietary engine, gRPC SDN — not an operator). Note: OCUDU now ships ocudu_netconf + a K8s ocudu_o1_adapter sidecar + ocudu_helm, so the moat is NTN-domain fleet orchestration above the node-local surface, not "K8s-ifying the RAN."

ADR 0005 promoted Draft → Proposed (working recommendation: Option B, cluster orchestrator on !798), pending final sign-off. Fixed the stale #60→#162 cross-reference in the ADR.

[thc1006]

v0.6 scope status — 2026-07-09 (ADR 0005 Accepted, PR #190)

ADR 0005 is now Accepted. The v0.6 Option-B scope has largely shipped:

• Step 1 — runtime push → !798: ✅ PR fix+feat(v0.6): gNB-parseable OCUDU config + runtime NTN push (#176) + k_mac (#52) #177 (PushRuntimeUpdate → ntn_config_update over the remote_control WS, pkg/provider/ocudu/wsclient.go).
• Step 2 — close SatelliteEphemeris: propagated SGP4 ECEF is discarded — feed it into OCUDU via ntn_config_update (!798) instead of re-emitting static ephemeris #176: ✅ PR fix+feat(v0.6): gNB-parseable OCUDU config + runtime NTN push (#176) + k_mac (#52) #177 (SatelliteEphemeris SGP4 ECEF flows into the push).
• Step 3 — SatelliteEphemeris → multi-NTNCellConfig fan-out: ✅ already wired — NTNCellConfigReconciler watches SatelliteEphemeris; ephemerisToNTNCellConfig enqueues every referencing cell (one source → N cells; mapper unit-tested).
• Off-cycle freshness: ✅ PR Runtime NTN push: off-cycle ephemeris epoch staleness window (decouple propagation from GP-fetch cadence) #179 — decouples propagation from the GP-fetch cadence so the fan-out delivers a fresh epoch continuously, not only every ~2h.

Remaining (v0.7, upstream-gated) — out of v0.6:

• Cell activate/deactivate (cell_lock/cell_unlock) — OCUDU !862/!863 still redesigning.
• NTNSlice fleet failover across the constellation.
• GroundStationLifecycle → cell binding.
• Open Question 3 (ADR 0005): fan-out home — extend NTNSlice vs a new cluster-scoped CRD.

Tail item #188 (pass-prediction recompute churn on the short cadence) is non-blocking.

Net: the v0.6 cluster-scope foundation (runtime push + orbit→SIB19 loop + fan-out + off-cycle freshness) is in main. Next: an end-to-end multi-cell fan-out integration test, then the v0.7 items.