Skip to main content

FCM Phase 4 Readiness Runbook

🟢 CLOSED 2026-05-19 — Phase 4 final GO confirmed, 24h soak passed tanpa rollback. Service renamed fcm_service → firebase_service di repo dan merchant_fcm → merchant_firebase (binary firebase-service, systemd unit firebase-service.service) di VM. Runbook ini kept untuk reference historical — referensi fcm_service / fcm-service di bawah harus dibaca sebagai nama lama service. Untuk runbook current state lihat services/firebase_service/README.md (di repo, di luar Docusaurus).

Tujuan (historical): Memvalidasi services/fcm_service (sekarang services/firebase_service) stabil sebelum Phase 4 cleanup (hapus merchant_core_api/internal/push/ + fallback path di dispatchPush).

Prasyarat aktif:

  • Phase 1 (token store ke service): aktif 2026-05-11. Flag FCM_TOKEN_STORE sudah dihapus per Phase 4 step 4 (2026-05-14) — single path tanpa flag.
  • Phase 2 (send transport ke service): aktif sebelum 2026-05-11. Flag FCM_TRANSPORT sudah dihapus per Phase 4 step 4 (2026-05-14).
  • Phase 4 step 1+2+4+5+6+7 DONE (2026-05-14). Folder merchant_core_api/internal/push/ dihapus seluruhnya — sumber FCM Firebase SDK + copy resolver tunggal di services/firebase_service/. Sisa hanya step 3 (MarkTokenInvalid di service postgres.go).
  • Mirror obligation feedback_fcm_service_mirror.md masih aktif — EXPIRED per Phase 4 step 7 (2026-05-14). Memory di-rewrite ke single-source mode.

Observation window: 14 hari berturut tanpa red flag.

Owner: Backend Lead Plan terkait:

1. Quick health snapshot​

Jalankan setiap kali mulai sesi monitoring:

# 1. Service alive + Firebase + Postgres OK
curl -sS http://127.0.0.1:8093/health | jq

# expected (post Phase 4 final rename 2026-05-19):
# {
#   "service": "firebase-service",
#   "status": "ok",
#   "firebase_configured": true,
#   "postgres_configured": true,
#   "app_env": "production"
# }

# 2. Tidak ada fallback warning di 24 jam terakhir
journalctl -u core-api --since="24 hours ago" 2>/dev/null \
  | grep -cE "fcm_service send failed; falling back to internal|fcm_service register failed; falling back to db"
# expected: 0

2. Monitoring categories​

2.1 Fallback rate (CRITICAL)​

Wrapper dispatchPush fallback ke pushClient internal kalau HTTP service gagal. Kalau fallback sering jalan = service belum stabil. Phase 4 wajib 0 fallback selama 48 jam terakhir.

Daily check:

journalctl -u core-api --since="24 hours ago" 2>/dev/null \
  | grep -E "falling back" \
  | awk '{print $1, $2, $3}' \
  | wc -l

Weekly aggregate:

journalctl -u core-api --since="7 days ago" 2>/dev/null \
  | grep "falling back" \
  | awk '{print $1, $2, $3}' | cut -d: -f1 | sort | uniq -c

Threshold:

Daily countStatus
0✅ Healthy
1–10⚠️ Edge case (network blip) — catat penyebab
>10❌ Service tidak stabil — investigate sebelum lanjut

2.2 Latency send-side​

HTTP loopback (127.0.0.1:8093) target add < 5ms vs internal call. Plan §4 Phase 2 acceptance: P99 < 500ms (termasuk hop HTTP).

Benchmark manual (mingguan):

KEY=$(grep ^INTERNAL_NOTIFICATION_API_KEY services/firebase_service/.env.production | cut -d= -f2-)
for i in {1..20}; do
  curl -sS -o /dev/null -w "%{time_total}\n" \
    -X POST -H "X-Internal-API-Key: $KEY" -H "Content-Type: application/json" \
    -d '{"push_token":"dummy_for_bench","type":"news","title":"bench","body":"bench"}' \
    http://127.0.0.1:8093/internal/fcm/send
done | awk '{s+=$1; n++; if($1>max)max=$1} END {print "avg=" s/n "s, max=" max "s"}'

Note: token "dummy_for_bench" akan return 410 invalidated — itu OK, ini hanya untuk ukur roundtrip HTTP + Firebase auth latency.

Threshold:

MetricTargetInvestigate if
Avg< 200ms> 500ms
Max< 500ms> 2s
P99 (rough)< 500ms> 1s

Penyebab umum P99 lonjak: Firebase JWT cache expired bersamaan (re-auth ke Google OAuth), DNS resolution slow, GC pause di Go process.

2.3 Token invalidation rate​

FCM API return UNREGISTERED → token auto-deactivate. Rate baru harus konsisten dengan baseline pre-Phase-2.

Cek harian:

PGPASSWORD='<password>' psql -h 10.8.0.1 -U kesles -d db_kesles_merchant -c "
SELECT date_trunc('day', updated_at) AS day,
       COUNT(*) FILTER (WHERE is_active = FALSE) AS deactivated,
       COUNT(*) FILTER (WHERE is_active = TRUE)  AS active_touched
  FROM notification.fcm_push_tokens
 WHERE updated_at >= now() - interval '14 days'
 GROUP BY day
 ORDER BY day DESC;
"

Red flag:

  • Mendadak deactivate 100% token dalam 1 hari → bug error translation di dispatchPush (false-positive IsInvalidTokenError)
  • 0 deactivate >7 hari padahal merchant aktif → bug, tidak ada turnover natural dari reinstall/uninstall

2.4 Service uptime + resource​

firebase_service tidak boleh crash atau memory leak. Plan §9 Phase 3 target: CPU < 5%, RAM < 100MB idle.

Cek harian:

# Status systemd (kalau dideploy via systemd)
systemctl status firebase-service
journalctl -u firebase-service --since="7 days ago" 2>/dev/null \
  | grep -iE "panic|fatal|crash|killed|oom" | head

# Resource (mac dev / linux server)
ps aux | grep firebase-service | grep -v grep
# RSS column = memory KB. Target < 100MB = 102400 KB.

# Restart count
systemctl show firebase-service -p NRestarts
# expected: 0 (untuk window observation)

Red flag: restart count > 0, RSS naik linear (memory leak >50MB/hari).

2.5 End-to-end delivery + multi-staff (Phase B.1 verify)​

Verifikasi FCM benar-benar sampai device + fan-out ke owner/admin/staff.

Mingguan manual test:

  1. Login mobile app sebagai 2-3 user dari merchant sama (mis. Makan Kini MRC-0174105647947242) dengan role berbeda: owner, admin, staff.
  2. Dashboard PSP tester → klik "Push transaction event → FCM + TTS (Voice Announcement)" 1x.
  3. Verify dalam 5 detik:
    • FCM tray muncul di semua 3 device.
    • TTS announce "Transaksi sukses senilai Rp 19.590" di semua device.
    • Inbox in-app menerima entry.
  4. Tap notif → app open ke tab Transactions, transaksi tampil di list.

Cek event log untuk audit kuantitatif:

PGPASSWORD='<password>' psql -h 10.8.0.1 -U kesles -d db_kesles_merchant -c "
SELECT processing_status, COUNT(*) AS events_24h
  FROM psp.event_log
 WHERE created_at >= now() - interval '24 hours'
 GROUP BY processing_status
 ORDER BY events_24h DESC;
"

Expected: 'success' jadi mayoritas. 'logged_only_*' boleh ada (merchant inactive / no active device / not found) untuk traffic tester, bukan transaksi production real.

3. Daily log template​

Salin baris baru tiap hari selama observation window:

2026-05-__ │ fallback=__ │ latency_avg=__ms │ deactivate=__ │ uptime=OK/CRASH │ smoke=PASS/FAIL │ note=__

Log entries (isi setiap hari)​

DateFallback (24h)Latency avgToken deactivated (24h)UptimeSmoke testNote
2026-05-11TBDTBDTBDOK(start of obs)observation kickoff, FCM_TRANSPORT=service + FCM_TOKEN_STORE=service di production
2026-05-12TBD~300msTBDOKPASSsmoke test post-deploy: sent_count: 1 lewat /internal/notifications/push/test
2026-05-130TBD1OKPASSregresi 8 guard pushClient.IsConfigured() di-fix via canDispatchPush() helper. Re-smoke 3 event (login_alert / order_paid / merchant_status_active) all sent_count:1. 1 token stale auto-deactivated
2026-05-140TBDTBDOKPASSPhase 4 step 1+2+4+5+6+7 DONE (code) + Step 3 deploy + smoke verified malam: (1) delete internal/push/fcm.go + migrate 6 caller ke fcmservice.SendInput, dispatchPush jadi single-path. (2) Drop FirebaseProjectID/ClientEmail/PrivateKey field + loader dari internal/config/config.go. (4) Drop FCMTokenStore/FCMTransport flag dari config + 5 env file. (5) Drop DB fallback di registerPushTokenViaConfiguredStore. (6) Migrate LoginNewDeviceCopy ke services/firebase_service/internal/push/copy.go::ResolveCopy + auto-resolve di handleSend saat caller pass Type=login_alert + DeviceName. HAPUS folder internal/push/ di core_api seluruhnya. (7) Mirror obligation memory di-rewrite ke single-source. go build/vet/test clean kedua repo. Rollback path berubah dari flag-flip ke git revert+ rebuild. Phase 4 Step 3 (MarkTokenInvalid single-writer) — service v2 deploy 19:53 WIB + smoke verified 21:28 WIB (fake-token push returns {invalidated:true, deactivated_at:"2026-05-14T14:28:40Z", status:"invalidated"}). Production incident: gateway 192.168.62.253 lost internet upstream ~18:09-21:28 WIB (~3 jam, full FCM push delivery DOWN). Network self-recovered (atau operator DC fix). Action item filed: fix static default route metric supaya kernel auto-fallback ke dhcp default. Soak window aktif T+0 = 21:28:40 WIB → T+24h decision 2026-05-15 21:28 WIB
2026-05-150TBD10OKPASST+24h backend soak PASSED. fcm-service.service uptime 16h post manual restart 06:34:57 WIB (clean stop+start systemd, no signal/crash, 56s ops gap). Step 3 (MarkTokenInvalid single-writer auto-deactivate) verified pada production traffic: 10 fcm_send: invalidated events (FCM HTTP v1 return errorCode: UNREGISTERED), rate ~1.6/h. Error-level log 24h = 1 entry (very clean). Token masking benar (dIUtxpuH...), tidak ada deactivate failed log → SQL UPDATE sukses tiap kali. Opsi C smart fallback di core_api 4 callsite tidak perlu trigger backup path (DB sudah konsisten via service-side write). Next: monitor AAB v10 mobile soak Day 1-7 (T+0 = 15 Mei dini hari → T+7 promote decision = 22 Mei)
2026-05-160223ms (smoke)0 (prod) + 1 (smoke)OKPASSDay 1 soak PASSED. Service uptime 29h32m sejak restart 06:34:57 WIB Jum, no new restart. Journal 13.5h window (22:36→11:56) total 1 entry (persistent error baseline kemarin, no growth, 0 new invalidated production). Day 1 smoke test 12:09:49 WIB return {invalidated:true, deactivated_at:"2026-05-16T05:09:49Z", status:"invalidated"}, HTTP 410, latency 223ms → Step 3 path live + DB UPDATE sukses. Interpretasi 10→0 invalidated Day 0→Day 1: pool stale token dibersihkan via Step 3 Day 0; Day 1 tokens yang is_active=true semuanya valid → no UNREGISTERED triggers. FCM error class shift di smoke: fake token format invalid → INVALID_ARGUMENT (vs prod stale → UNREGISTERED); keduanya match push.IsInvalidTokenError() → path Step 3 identik
2026-05-170250ms (smoke)10 (prod) + 1 (smoke)OKPASSDay 2-3 soak PASSED. Service uptime 48h+ tanpa restart sejak 15 Mei 06:34:57 WIB. Memory 4.5 MB (stabil, no leak vs 4.3 MB Day 0), CPU 7.336s total 48h (~0.0043% avg). Production window 16 Mei 12:09 → 17 Mei 19:45 (~31.5h): 10 invalidated events (rate ~0.32/h ≈ 7.7/hari = natural turnover steady-state). Error-level log tetap 1 entry baseline (no growth Day 0→1→2). Total journal 139 lines / 31.5h = ~4.4 lines/h (activity normal). Day 2 smoke 17 Mei 19:52 WIB: {invalidated:true, deactivated_at:"2026-05-17T12:52:27Z", status:"invalidated"} HTTP 410, latency 250ms → Step 3 path live + DB UPDATE sukses. Steady-state pattern Day 0 surge (10 cleanup) → Day 1 quiet (0 new) → Day 2-3 natural turnover (10 in 31.5h ≈ ekspektasi long-term). Cross-stack: merchant-whatsapp.service active 3+ hari (since 2026-05-13 21:08) memory 5.1 MB — stack overall sehat untuk publish-readiness
2026-05-180n/a (no smoke)1 (prod)OKDEFERREDDay 3 soak PASSED (dengan catatan deploy). Manual ops 18 Mei 00:30-00:37 WIB: operator stop fcm-service.service 00:30:53, binary rebuild/replace 00:33:31 (mtime 2026-05-18 00:33), start kembali 00:37:35 (clean systemd: dotenv loaded → postgres connected → listening :8093, no Failed, no Main process exited). PID baru 144961, uptime 20h tanpa restart. Memory 4.6 MB (stable vs 4.5 MB Day 2). Production traffic 15:35:38 WIB: 1 fcm_send: invalidated token=ebFcQnwx... UNREGISTERED event → Step 3 logic verified-by-traffic di binary baru (rate ~0.05/h, pool sudah cleaned out, natural turnover decay). Caveat: source diff untuk binary baru tidak bisa di-verify (folder deploy flat, no .git, source repo path tidak terverifikasi). Behavior Step 3 identik pattern Day 0-2 (auto-deactivate via FCM v1 UNREGISTERED), tapi continuous-uptime metric formal di-reset 00:37:35. Smoke test endpoint deferred (endpoint path /fcm/send 404, perlu discovery `strings
2026-05-190n/a (no smoke)0 (prod, 4.6h window)OKDEFERREDDay 4 soak PASSED (behavioral) — redeploy ke-2 berturut. Manual ops 19 Mei 05:05-05:08 WIB: operator stop fcm-service.service 05:05:53, binary rebuild/replace 05:06:37 (mtime 2026-05-19 05:06:37, size 7,770,274 bytes — identik size dengan deploy 18 Mei tapi mtime beda), start kembali 05:08:07 (clean systemd: dotenv loaded → postgres connected → listening :8093, no Failed, no Main process exited). PID baru 183422 (≠ 144961 kemarin), uptime 4h39min saat check 09:47 WIB. Memory 1.8 MB (fresh boot, vs 4.6 MB Day 3 yang lebih lama running). CPU 91ms (idle). Total downtime gap 2 menit 14 detik. Production events: 0 UNREGISTERED dalam window 05:08 → 09:47 (~4.6h) — bisa karena pool sudah cleaned out (mirror Day 1 quiet pattern) atau natural Poisson variance (steady-state Day 2-3 ~0.32/h ekspektasi ~1.5 events di 4.6h, observasi 0 masih dalam range). Caveat strict soak: 2 redeploy berturut-turut (18 Mei + 19 Mei) breaks "7-day continuous binary" claim untuk 22 Mei promote decision. Behavioral soak masih ✅ karena: Step 3 logic verified-by-traffic di binary 18 Mei (event 15:35:38), no crash, no auto-recover systemd, restart selalu clean operator action. Size binary identik 18→19 (7770274) suggests same source rebuild (Go embed timestamp di output → mtime beda walau source sama) atau trivial tweak yang tidak mengubah size. Action item operator: freeze fcm-service redeploy dari 19 Mei sampai 22 Mei (≥3 hari) untuk lock soak window — perubahan code urgent defer ke post-launch (25 Mei+). Kalau redeploy harian terus berlanjut, decision 22 Mei wajib pakai behavioral verdict (Step 3 verified + 0 crash) bukan strict uptime
2026-05-200409ms (smoke)1 (smoke, post-recovery)🟢 RECOVEREDPASSDay 5 — SERVICE RENAMED (Phase 4 final closeout) + VPN OUTAGE → RECOVERED + Step 3 VERIFIED. Cut-over fcm-service.service → firebase-service.service di 19 Mei 20:46:05 WIB. Description: "Merchant Firebase Service (FCM+Phone Auth)" — Phone Auth belum di-set / belum implementasi (confirmed operator 20 Mei): description aspirational, scope sebenarnya saat ini masih FCM-only. Binary effectively fcm-service yang di-rename. Path baru: /home/enalfarid/kesles_merchant/merchant_firebase/firebase-service (sebelumnya merchant_fcm/fcm-service). Repo rename: services/fcm_service → services/firebase_service + merchant_fcm → merchant_firebase (source confirmed di binary strings: git.kesles.com/merchant/services/firebase_service/internal/push/copy.go + fcm.go). Binary mtime 2026-05-19 10:41:58, size identik 7,770,274 bytes dengan fcm-service binary 18-19 Mei → confirmed same FCM code base (Phone Auth pending). Service started 19 Mei 20:46:05 WIB, uptime ~27h saat check Day 5, memory 8.6 MB, CPU 105ms (idle), Tasks 5, PID 686. Endpoints inventory (via `strings
2026-05-21
2026-05-22
2026-05-23
2026-05-24
2026-05-25(end of observation)

4. Decision matrix (Phase 4 GO / NO-GO)​

Saat hari ke-14 (2026-05-25), evaluasi:

KriteriaThresholdHasil
Fallback warning 48 jam terakhir0__
Service crash 14 hari0__
Latency P99< 500ms__
Token deactivation ratenormal__
Multi-staff smoke test (mingguan)pass × 2 minggu__
event_log processing_status='success' mayoritas> 90% non-test traffic__

Decision:

  • ✅ Semua kriteria pass → GO Phase 4 cleanup
  • ⚠️ 1-2 kriteria warning (recoverable) → WAIT, fix dulu, reset counter 48 jam
  • ❌ 1+ kriteria fail (uncoverable) → NO-GO, investigate akar masalah, pertimbangkan rollback flag ke internal/db

5. Phase 4 cleanup actions (kalau GO)​

Update 2026-05-14: step 1-3 + 7 dieksekusi lebih awal (sebelum Phase 3 14-hari observation selesai) atas keputusan team. Phase 3 day 3 zero red flag + scope step 1-3 self-contained tanpa butuh data observasi. Step 4-6 ditahan sampai 2026-05-25 GO confirmed.

  1. [DONE 2026-05-14] Hapus code embedded di merchant_core_api/:
    • internal/push/fcm.go — DELETED (305 baris)
    • internal/push/message_copy.go — DIPERTAHANKAN (masih dipakai auth_mobile_handlers.go::LoginNewDeviceCopy). Akan dihapus kalau copy helper di-migrate ke service.
  2. [DONE 2026-05-14] Hapus field & init:
    • pushClient *push.Client dari httpapi.Server struct — REMOVED
    • push.NewClient(cfg) di NewServer — REMOVED
  3. [DONE 2026-05-14] Sederhanakan dispatchPush (push_dispatcher.go):
    • Fallback ke s.pushClient.Send — REMOVED
    • Translation hack "fcm api error via service: token unregistered" — REMOVED
    • Signature jadi (*fcmservice.SendResult, error)
    • 6 caller migrate: push.SendInput → fcmservice.SendInput, push.IsInvalidTokenError(err) → result.Invalidated
  4. [DONE 2026-05-14] Hapus feature flag:
    • FCMTransport field + loader getEnv("FCM_TRANSPORT", ...) di config.go — REMOVED
    • FCMTokenStore field + loader getEnv("FCM_TOKEN_STORE", ...) di config.go — REMOVED
    • FCM_TRANSPORT + FCM_TOKEN_STORE baris dari semua env files (.env, .env.production, .env.example, .env.development.example, .env.production.example) — REMOVED
    • Bonus Step 2: field FirebaseProjectID, FirebaseClientEmail, FirebasePrivateKey + loader os.Getenv("FIREBASE_*") di config.go juga REMOVED — merchant_core_api benar-benar tidak tahu apa-apa tentang Firebase. Kredensial hanya di services/firebase_service/.env.
  5. [DONE 2026-05-14] Hapus fallback DB di auth_profile_handlers.go::registerPushTokenViaConfiguredStore:
    • Cuma path service, error = real failure. Drop DB fallback path.
    • Drop import log/slog (warning log "falling back to db" tidak ada lagi).
  6. [DONE 2026-05-14] Migrate copy resolution + delete internal/push/:
    • Decision: Option A (migrate ke service).
    • LoginNewDeviceCopy dipindah dari merchant_core_api/internal/push/message_copy.go → services/firebase_service/internal/push/copy.go::ResolveCopy + loginNewDeviceCopy.
    • Tambah field DeviceName di SendInput (core_api fcmservice + firebase_service push).
    • Service auto-resolve title/body saat Type=login_alert + caller pass DeviceName tanpa Title/Body (wire di handleSend).
    • Caller auth_mobile_handlers.go drop pre-resolve, kirim Type + DeviceName saja.
    • 4 dead code fn (EmailVerifiedCopy/WelcomeUserCopy/MerchantRegistrationPending/MerchantActivated) ikut hapus.
    • HAPUS folder merchant_core_api/internal/push/ (fcm.go + message_copy.go).
    • Mirror di repo_exports/.../internal/push/ ikut HAPUS.
    • PENDING separately: profileAuth.RegisterPushToken (DB direct write) tidak dipanggil dari auth_profile_handlers.go lagi. Cek apakah ada caller lain; kalau tidak ada → deprecate atau hapus method. Not blocking Phase 4 GO.
  7. [DONE 2026-05-14] Mirror obligation berakhir:
    • services/firebase_service/internal/push/fcm.go jadi sumber tunggal
    • Memory feedback_fcm_service_mirror.md di-rewrite untuk single source
  8. [PARTIAL] Lainnya:
    • PENDING: Implementasi MarkTokenInvalid di services/firebase_service/internal/store/postgres.go (auto-deactivate token saat FCM return UNREGISTERED — saat ini caller core_api yang s.postgresStore.DeactivatePushToken)
    • DONE 2026-05-14: Update service-topology.md reflect firebase-service sebagai standalone node
    • DONE 2026-05-14: Update architecture.md §10 caller table + §3.7 /ready endpoint + §6 schema reconciliation (migration 125)
    • PENDING: Decision retention internal/push/message_copy.go (masih dipakai auth_mobile_handlers.go::LoginNewDeviceCopy — opsi: tinggalkan permanen di core_api, atau migrate ke service supaya package internal/push/ bisa benar-benar dihapus)
  9. Update plan status:
    • extraction-plan.md Phase 4 header → "Step 1+2+4+5+7 DONE"
    • services/firebase_service/README.md status table — belum di-update

6. Rollback procedure (kalau NO-GO)​

⚠️ Post 2026-05-14 (Phase 4 step 1): rollback flag-flip ke FCM_TRANSPORT=internal sudah tidak applicable karena internal/push/fcm.go + pushClient field sudah dihapus dari code base. Rollback satu-satunya = git revert <Phase 4 step 1 commit> + rebuild.

Rollback emergency post-Phase-4 (recommended path)​

ETA ~5-10 menit:

# 1. Identifikasi commit Phase 4 step 1
git log --oneline --all | grep -i "phase 4\|fcm cleanup\|delete fcm.go" | head -5

# 2. Revert commit (multi-file: push_dispatcher, server.go, 6 caller, push/fcm.go)
git revert <commit-sha>

# 3. Restore FIREBASE_* di merchant_core_api/.env.production
#    (sebelumnya di-comment out di Phase 2 — uncomment baris berikut)
#    FIREBASE_PROJECT_ID=kesles-merchant
#    FIREBASE_CLIENT_EMAIL=firebase-adminsdk-fbsvc@kesles-merchant.iam.gserviceaccount.com
#    FIREBASE_PRIVATE_KEY=<PEM>

# 4. Set FCM_TRANSPORT=internal (force pakai pushClient lokal)
echo "FCM_TRANSPORT=internal" >> merchant_core_api/.env.production

# 5. Rebuild + redeploy core_api
make build && scp dist/merchant_core_api kesles-merchant:/home/.../bin/
ssh kesles-merchant 'sudo systemctl restart merchant_core_api'

# 6. Smoke test
curl -X POST .../internal/notifications/push/test ...

Rollback FCM_TOKEN_STORE saja (PRE-Phase-4 era, sudah tidak applicable)​

# HISTORICAL — pre-2026-05-14 valid path:
# Di merchant_core_api/.env.production:
FCM_TOKEN_STORE=db          # ganti dari "service"
# Restart core_api → register pakai DB langsung lagi

Catatan: FCM_TOKEN_STORE flag masih ada di config (no-op since Phase 4 step 1 cleanup belum hapus flag handler). Tapi semantik post-cleanup adalah db-only fallback path tidak ada lagi kalau code di auth_profile_handlers.go::registerPushTokenViaConfiguredStore ikut dibersihkan di step 5 (pending).

Yang tidak perlu rollback:

  • DB migration 125 (app_version, heartbeat_at) — additive, non-breaking
  • Code di services/firebase_service/ — tetap kompil & jalan idle
  • Token data di notification.fcm_push_tokens — tetap valid

7. Signoff​

FieldValue
Observation start2026-05-11
Observation end2026-05-25
Decision date__
DecisionGO / WAIT / NO-GO
Decided by__
Notes__

8. Referensi terkait​

  • extraction-plan.md (di working notes)
  • architecture.md
  • shadow-mode-plan.md — Phase B (Polqo migration; di working notes)
  • Memory feedback_fcm_service_mirror.md — mirror obligation core_api ↔ firebase_service