Files
karaoke-gen/docs/ARCHITECTURE.md
T
Andrew Beveridge ☄️andClaude Opus 5.5 e9465c73d4 chore: gemini-3.8-flash everywhere + auto-correct on Opus 5.5 (cost) (#1054)
## Why
GCP credits expired 2026-09-19; Vertex Gemini 3.1 Pro spend had grown $9
→ $37 → $105/30d. `gemini-3.8-flash` is $0.75/$3.75 per 1M (vs $2/$12
for 3.1 Pro; rises to $1.50/$7.50 on 2027-01-01). Opus 5.5 is $4/$20 (vs
$5/$25 for 4.8).

## Changes
- All Gemini defaults → `gemini-3.8-flash`: credit eval, custom lyrics,
auto-correct (default + compare leg), match judge, parse titles, tenant
bulk analyze, agentic AI router, error-monitor LLM analysis (also
`us-central1` → `global`, required for Gemini 3.x), frontend
`translate.py`.
- Auto-correct Claude leg `claude-opus-4-8` → `claude-opus-5-5` (ci.yml
env), with `effort: "high"` pinned (5.5 defaults to medium). Pricing
table gains both new models.
- Docs updated; version 0.241.1.

## Evidence (12 recent prod jobs, pre-review lyrics scored against the
human-reviewed final)
| Model | Precision | Word-level recall | $/job |
|---|---|---|---|
| gemini-3.1-pro-preview | 91% | 74% | $0.067 |
| gemini-3.8-flash | 98% | 38–41% | $0.020 |
| gemini-3.8-flash @ high thinking | 97% | 60% (4/12 timed out, 504) |
$0.041 |
| claude-opus-4-8 | 94% | 33% | $0.080 |
| claude-opus-5-5 (effort high) | 93% | 58% | $0.048 |

Flash for the Gemini leg is a deliberate cost/recall trade-off (Andrew
chose it); Opus 5.5 more than compensates. Also verified:
`gemini-3.8-flash` works on Vertex `global` (rejects
`thinking_level=minimal`; nothing here uses it), match-judge latency
median 3.3s / max 6.2s (< 12s timeout), Japanese translation smoke test
preserves ICU placeholders. Translation cache is keyed on English text +
locale, so no re-translation churn.

Tests: auto_correct, credit eval, custom lyrics, parse_titles,
error_monitor, agentic router suites pass locally.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* AI-powered correction, translation, matching, title parsing, and
evaluation now use updated models, including Gemini 3.8 Flash and Claude
Opus 5.5.
* **Updates**
* AI service pricing reflects scheduled rate changes, including a Gemini
3.8 Flash rate change beginning January 1, 2027.
  * Application version updated to 0.241.1.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-26 14:53:04 -04:00

43 KiB

Architecture

System Overview

┌─────────────────────────────────────────────────────────────────┐
│  Frontend (Cloudflare Pages)                                    │
│  Next.js + TypeScript                                           │
│  https://gen.nomadkaraoke.com                                   │
└───────────────────────────┬─────────────────────────────────────┘
                            │ HTTPS
                            ▼
┌─────────────────────────────────────────────────────────────────┐
│  Cloudflare edge  (api.nomadkaraoke.com proxied — orange-cloud) │
│  WAF (exploit-path block) + rate limit + injects X-Edge-Auth    │
│  header. See § Edge Security (Cloudflare).                      │
└───────────────────────────┬─────────────────────────────────────┘
                            │ HTTPS (SSL "full")
                            ▼
┌─────────────────────────────────────────────────────────────────┐
│  Backend API (Cloud Run)                                        │
│  FastAPI + Python 3.12                                          │
│  https://api.nomadkaraoke.com  (origin lock: rejects direct-    │
│  to-origin requests lacking X-Edge-Auth, health/root exempt)    │
│                                                                 │
│  Workers:                                                       │
│  • Audio Worker    - Stem separation via Cloud Run GPU (L4)              │
│  • Lyrics Worker   - Transcription via AudioShake               │
│  • Screens Worker  - Title/end screen generation                │
│  • Render Worker   - Karaoke video with lyrics                  │
│  • Video Worker    - Final assembly and encoding                │
└───────────────────────────┬─────────────────────────────────────┘
                            │
              ┌─────────────┼─────────────┬───────────────┐
              ▼             ▼             ▼               ▼
        ┌──────────┐  ┌──────────┐  ┌──────────────┐ ┌──────────────┐
        │ Firestore│  │   GCS    │  │Secret Manager│ │GCE Encoding  │
        │ (jobs)   │  │ (files)  │  │  (API keys)  │ │  Worker (*)  │
        └──────────┘  └──────────┘  └──────────────┘ └──────────────┘

(*) GCE Encoding Worker: a ranked pool of 6 x86_64 32-vCPU highcpu machine
    families (c4d/c4/n4d/c2d/n2d/n2). c4d-highcpu-32 (AMD EPYC Turin) is the
    fastest and stays the primary/preferred pick; the other families are
    stockout-resilience fallbacks selected fastest-first with a capacity cooldown.
    Used for both final video encoding and preview video generation.
    Uses immutable deployment pattern - see infrastructure/encoding-worker/README.md
    and the "Encoding Worker" section below.

Error Monitor

Cloud Logging (all services) ──→ Error Monitor (Cloud Run Job, */15 min)
Frontend clients ───────────────→ POST /api/client-errors ──→ Firestore
                                    │
                                    ├─ Normalize & deduplicate
                                    ├─ LLM incident grouping (Gemini Flash)
                                    ├─ Discord alerts (new patterns, spikes, daily digest)
                                    ├─ Firestore: error_patterns, error_incidents
                                    └─ Auto-resolve silent patterns

Entry point: python -m backend.services.error_monitor.monitor
Design spec: docs/archive/2026-04-12-error-monitor-design.md
Frontend crash reporting: docs/archive/2026-04-19-frontend-crash-reporting-plan.md

Frontend crashes are captured via window.onerror, unhandledrejection, Next.js error.tsx, and a CrashReportBoundary around the review UIs. They POST to /api/client-errors, which normalises and upserts into the same error_patterns collection with service="frontend" so they flow through the existing Discord-alerting pipeline. ChunkLoadError after deploy triggers an automatic hard reload; the CrashReport card also checks /version.json and shows an "Update now" CTA when the running bundle is stale.

Noise is dropped before alerting: benign messages (media AbortError, ResizeObserver loop) and opaque cross-origin Script error. events (no error object — thrown by third-party/extension scripts) are skipped client-side, and /api/client-errors accepts-but-ignores reports from bot user agents and Script error. window errors with a blank filename/lineno (covers older cached bundles).

Edge Security (Cloudflare)

api.nomadkaraoke.com is proxied through Cloudflare (orange-cloud) — the same zone that already fronts the marketing site, gen, decide, and the tenant portals. Cut over 2026-07-20 after a vuln-scanner hammered the previously directly-exposed Cloud Run origin. Managed as IaC in infrastructure/modules/edge_security.py (pulumi_cloudflare).

Request path: client → Cloudflare edge → Cloud Run. At the edge:

  • WAF (http_request_firewall_custom) blocks exploit/secret-exposure paths (/.env, /.git/, /.ssh/, wp-config, /etc/passwd, …) — uses the contains operator (regex matches needs a Business plan; zone is Free).
  • Rate limiting (http_ratelimit) — per-IP flood control (Free-plan limits: period=10s, characteristics must include cf.colo.id). Excludes /api/internal/* (scheduler/Cloud-Tasks cron).
  • Origin lock — a Transform Rule injects a secret X-Edge-Auth header (http_request_late_transform); the backend's EdgeAuthMiddleware (backend/middleware/edge_auth.py, EDGE_AUTH_MODE=enforce) rejects public requests lacking it, so hitting the raw *.run.app origin directly is blocked (health/root exempt for Cloud Run probes). Internal callers reach the backend via api.nomadkaraoke.com (Cloud Tasks use CLOUD_RUN_SERVICE_URL), so they traverse the edge and carry the header.

SSL: zone TLS mode is "full" (not strict). The Cloud Run domain-mapping managed cert cannot provision behind Cloudflare (ACME challenge not visible), but "full" mode doesn't validate the origin cert, so the edge works and is renewal-safe — no run.app Origin Rule or load balancer needed. Never switch the zone to "full (strict)". Full rationale + cutover runbook: docs/archive/2026-07-20-edge-security-hardening-plan.md and docs/archive/2026-07-20-edge-security-cutover-runbook.md.

Processing Pipeline

1. UPLOAD
   POST /api/jobs/upload
   • Save audio to GCS
   • Create job in Firestore
   • Trigger parallel workers

2. PARALLEL PROCESSING (with Worker Registry coordination)
   ┌─────────────────┐    ┌─────────────────┐
   │  Audio Worker   │    │  Lyrics Worker  │
   │  Cloud Run GPU      │    │  AudioShake     │
   │  2-stage sep    │    │  Auto-correct   │
   │  ↓ register()   │    │  ↓ register()   │
   │  ↓ unregister() │    │  ↓ unregister() │
   └────────┬────────┘    └────────┬────────┘
            └──────────┬───────────┘
                       ▼
3. SCREENS GENERATION + BACKING VOCALS ANALYSIS + AUDIO TRANSCODING
   • Title screen (5s video)
   • End screen (5s video)
   • Analyze backing vocals track (for instrumental selection)
   • Transcode audio to OGG Opus for review UI (~3 MB vs 35 MB FLAC)
                       ▼
4. COMBINED HUMAN REVIEW (AWAITING_REVIEW)
   • React UI for lyrics correction
   • Preview video generation (optionally via GCE for speed)
   • Instrumental track selection (clean vs with_backing)
   • Single submit saves both lyrics + instrumental choice
                       ▼
5. VIDEO RENDERING (optionally via GCE for speed + memory)
   • Merge corrections with original data
   • Generate karaoke video with lyrics (OutputGenerator + ffmpeg/libass)
   • Style assets cached on GCE worker disk between jobs
                       ▼
6. FINAL VIDEO (INSTRUMENTAL_SELECTED, via GCE)
   • Remux with selected instrumental
   • Concatenate: title + karaoke + end
   • Encode to 4 formats
                       ▼
7. COMPLETE
   • User downloads via signed URLs

Note: As of 2026-01, instrumental selection is combined with lyrics review. Users complete both in a single session rather than in separate steps.

Job States

State Description Human Action
pending Job created -
downloading Processing audio -
awaiting_duration_confirm Waiting for human — duration-based credit cost confirmation before separation/transcription (see Duration-Based Pricing) Confirm credit charge (or buy more credits)
separating_stage1 Audio separation stage 1 -
separating_stage2 Audio separation stage 2 -
transcribing Lyrics transcription -
generating_screens Creating title/end screens + backing vocals analysis -
awaiting_review Waiting for human Review lyrics + select instrumental
in_review Human actively reviewing Edit lyrics + select instrumental
review_complete Review submitted with instrumental selection -
rendering_video Generating karaoke video -
instrumental_selected Instrumental confirmed, ready for final video -
generating_video Final encoding -
complete All done Download files
failed Error occurred -

Note: awaiting_instrumental_selection exists for backwards compatibility with historical jobs but is no longer used. Instrumental selection is now part of the combined review (awaiting_review → in_review → review_complete).

awaiting_duration_confirm details: This state is a blocking human checkpoint inserted before separating_stage1. It is entered when the actual measured duration costs more than was already charged (stored in state_data.duration_confirm_reason as "reconcile"). This applies to all job types (URL, upload, search): 1 credit is charged at creation; ffprobe measures the actual audio via a signed URL just before workers start; if the true cost exceeds the hold, state_data.pending_additional_credits records the delta owed and the job pauses here.

POST /api/jobs/{id}/confirm-duration deducts the owed credits and triggers workers. If the actual duration exceeds 60 min all charged credits are refunded and the job is cancelled. If it costs less than charged an auto-refund is issued silently. Admin/made-for-you jobs bypass this state entirely. See docs/archive/2026-06-04-long-duration-input-handling-design.md.

State Machine Robustness (Feb 2026): State transitions are enforced by STATE_TRANSITIONS in models/job.py. Invalid transitions raise InvalidStateTransitionError by default. Use scripts/generate_state_diagram.py to generate a Mermaid diagram of valid transitions. The /health/job-consistency endpoint detects jobs stuck in invalid states. See docs/archive/2026-02-02-state-machine-robustness-plan.md for implementation details.

GCS File Structure

jobs/{job_id}/
├── input/
│   └── original.flac              # Uploaded audio
├── stems/
│   ├── vocals.flac
│   ├── instrumental_clean.flac
│   ├── instrumental_with_backing.flac
│   └── ...
├── lyrics/
│   ├── corrections.json           # Auto-corrected
│   ├── corrections_updated.json   # After human review
│   ├── edit_log_{session}.json    # User edit log with feedback
│   ├── annotations.json           # Correction annotations
│   └── karaoke.ass
├── screens/
│   ├── title.mov
│   └── end.mov
├── videos/
│   └── with_vocals.mkv            # Raw karaoke video (FLAC audio, large)
└── finals/
    ├── with_vocals_mp4.mp4        # Encoded with-vocals (H264/AAC, ~3x smaller)
    ├── lossless_4k.mp4
    ├── lossless_4k.mkv
    ├── lossy_4k.mp4
    └── lossy_720p.mp4

LyricsTranscriber Integration

Key Design Decision: We use LyricsTranscriber as a library, not a server.

CLI Mode (Original):              Cloud Mode (Our Approach):
────────────────────              ──────────────────────────
LyricsTranscriber                 LyricsTranscriber
├── transcribe()                  ├── transcribe()    ✅
├── correct()                     ├── correct()       ✅
├── ReviewServer.start()          │
│   └── BLOCKS (waits)            │  (Skip server, save JSON)
│                                 │  ... ASYNC GAP ...
│                                 │  (Human reviews via React)
└── OutputGenerator               │
                                  └── OutputGenerator  ✅
                                      (called after review)

Use: LyricsTranscriber, CorrectionResult, OutputGenerator, OutputConfig Don't use: ReviewServer, blocking waits

External Services

Service Purpose Required
Cloud Run GPU Job (L4, us-east4) Audio stem separation — Separator runs directly in the audio worker Job (no HTTP hop) Yes
AudioShake Lyrics transcription Yes
Vertex AI Agentic AI correction (Gemini 3.8 Flash) Default off (SKIP_CORRECTION=false to enable)
Genius Reference lyrics Yes
Flacfetch Audio downloads (YouTube, torrents) Recommended*
YouTube API Video upload Optional
Postmark Email notifications Optional
Cloud Tasks Delayed task scheduling (idle reminders, YouTube queue) Optional
Cloud Scheduler Hourly YouTube upload queue processing Optional
Cloud Monitoring Cross-reference YouTube API quota consumption from GCP metrics Optional
karaoke-decide Song catalog (MusicBrainz + Spotify) for autocomplete Optional
KaraokeNerds Community karaoke version detection Optional

*Flacfetch runs on a dedicated GCE VM with YouTube cookies and tracker access. Without it, YouTube downloads will fail due to bot detection on Cloud Run IPs.

Data Pipeline Cloud Functions

Three Cloud Functions manage the community karaoke data pipeline, running daily on automated schedules:

Function Schedule Purpose
divebar-mirror Daily 2:00 AM ET Indexes 48K+ files from diveBar Karaoke Google Drive into BigQuery divebar_catalog
kn-data-sync Daily 3:05 AM + 4:30 AM ET Fetches KaraokeNerds catalog (281K songs) and community tracks (58K with YouTube URLs) to BigQuery + GCS
divebar-lookup Daily 6:00 AM ET + on-demand API Public search/lookup API for KJ Controller; rebuilds KN↔Divebar cross-reference index

Infrastructure (Pulumi-managed in infrastructure/):

  • Service accounts: divebar-mirror@, kn-data-sync@, divebar-lookup@ (least-privilege IAM)
  • GCS: nomadkaraoke-kn-data (raw exports), 3 source buckets for function code
  • BigQuery tables: divebar_catalog, karaokenerds_community, kn_divebar_xref (in karaoke_decide dataset)
  • Secret: karaokenerds-api-key (KN API access)

Source code: infrastructure/functions/{divebar_mirror,kn_data_sync,divebar_lookup}/

Firestore Collections

karaoke-gen shares a GCP project (nomadkaraoke) with karaoke-decide, but uses separate Firestore collections:

Collection Purpose Key Fields
gen_users karaoke-gen user accounts email, credits, role, is_active
jobs Karaoke generation jobs job_id, user_email, status, state_data, processing_metadata
jobs/{job_id}/logs Worker log entries (subcollection) timestamp, level, worker, message, ttl_expiry
sessions Magic link auth sessions user_email, token, expires_at
magic_links Passwordless auth tokens email, token, expires_at, used
user_feedback User feedback for credits (all users) user_email, ratings, comments, created_at
beta_feedback Beta program feedback (deprecated) user_email, ratings, comments
youtube_quota_pending Pending YouTube API quota buffer (~7min GCP delay) pending_units, pending_uploads[], updated_at
youtube_upload_queue Deferred YouTube uploads job_id, status, user_email, queued_at, youtube_url

Note: The users collection in the same Firestore instance belongs to karaoke-decide (different schema: user_id, is_guest, quiz_* fields). Don't use it for karaoke-gen.

Worker Logs: Stored in subcollection (jobs/{job_id}/logs) instead of embedded array to avoid Firestore's 1MB document size limit. TTL policy auto-deletes logs after 30 days. Feature flag: USE_LOG_SUBCOLLECTION (default: true). See LESSONS-LEARNED.md.

Encoding Worker: blue-green primary pair + multi-family fallback pool

The GCE encoding worker is a blue-green primary pair (encoding-worker-a/-b, both c4d-highcpu-32 in us-central1-c) plus a ranked pool of stockout-resilience fallback VMs spanning 6 x86_64 machine families (see "Multi-instance-type pool" below). 10 VMs total, all TERMINATED by default and started on demand.

Architecture

Frontend → Backend (Cloud Run) → Firestore config → active worker (primary, or a
                                                   ↑  capacity-fallback "override")
Cloud Scheduler (5 min) → Cloud Function → checks idle + stops VMs
                                                   ↑
CI (GitHub Actions) → selects a fresh "green" (ranked pool) → health + encode test
                       → promote (swap primary OR set active_override) → stop old

Multi-instance-type pool (stockout resilience, v0.195.0)

A single machine family can hit a region-wide ZONE_RESOURCE_POOL_EXHAUSTED stockout across every zone at once. To survive that, the fallback fleet spans 6 families (c4d, c4, n4d, c2d, n2d, n2 — all -highcpu-32, ≥32 GB) across zones a/b/c/f: encoding-worker-fallback-a/-b (c4d), -n2c/-n2f (n2), -c4a (c4), -n4db (n4d), -c2df (c2d), -n2da (n2d).

Candidate ordering is one shared pure module, backend/services/encoding_worker_preference.py::ordered_candidates(pool, capacity_state), used by BOTH runtime selection and deploy green selection so they can't drift:

  • Fastest-first by SPEED_RANK (c4d preferred whenever it can start).
  • Availability-aware: a family that recently hit a capacity/stockout error is recorded in the Firestore capacity_state map and demoted for a 15-min cooldown, then re-probed — so we stop wasting ~2 min probing a dead family but snap back to c4d the moment capacity returns.

The fallback fleet is defined in infrastructure/config.py::EncodingWorkerConfig.FALLBACKS and published to the backend/idle-fn via the encoding-worker-fallback-vms secret (entries: {vm, zone, ip, machine_type}).

Key Components

Primary pair + fallback fleet: All VMs TERMINATED by default. The primary pair (a/b) does blue-green deploys; the 8 fallbacks absorb capacity stockouts. Runtime warmup (encoding_worker_manager.ensure_any_running) tries the ranked candidates in order, falling through on stockout and recording an active_override when it lands on a fallback.

Firestore config doc (config/encoding-worker): source of truth for routing. Fields: primary_vm/primary_ip/primary_version, secondary_vm/…, the capacity-fallback pointer active_override_vm/_ip/_zone/_version (set when a fallback is serving), deploy_in_progress, last_activity_at, and the capacity_state map ({"<machine_type>@<zone>": last_stockout_iso}). active_url = override IP if set, else primary IP.

Backend routing (encoding_interface.py): reads the active worker IP from the config doc with a 30-second TTL cache.

JIT startup: entering the lyrics-review page starts the worker on demand (~60 s natural warmup), avoiding 24/7 cost and cold-start latency.

Idle auto-shutdown: a Cloud Function (every 5 min) stops idle VMs (primary, secondary, and every fallback in the secret) based on Firestore activity — external to the VM so a sick worker still gets stopped.

Capacity-aware deploys: CI selects a FRESH "green" from the ranked pool via deploy_promote.select_green_candidates (the same preference module) — the c4d secondary if it can start, else a fallback family, never the currently-serving override. After a real encode test on the green, it promotes: a c4d-secondary green does the classic primary swap (and clears any stale override); a fallback green becomes the active_override. The retired VM is drained + stopped; a failed health check never promotes. Override routing keys off an explicit is_primary flag, not list position. See infrastructure/encoding-worker/deploy_promote.py.

Self-sufficient image: the Packer image (infrastructure/packer/scripts/provision.sh) bakes the full karaoke-gen dependency tree with CPU-only torch, so a fresh VM of any family boots ready to encode (no first-job dependency install).

Seeding the Config

Initial setup or reset:

python scripts/seed-encoding-worker-config.py <ip-a> <ip-b>

Deploy Flow (capacity-aware, v0.194.2+)

1. CI builds new karaoke-gen wheel + pushes to GCS
2. CI picks a FRESH "green" from the ranked pool (select_green_candidates →
   ordered_candidates): the c4d secondary if it can start, else the fastest
   available fallback family — never the currently-serving active_override.
3. CI starts the green, waits for /health, runs a real encode test against it.
4. If the test FAILS: CI stops the green; primary/override unchanged (no promote).
5. If the test PASSES: CI promotes the green:
   - green is the c4d secondary  → swap it to `primary_vm` (classic blue-green)
                                    AND clear any stale `active_override_*` so
                                    traffic returns to the fresh c4d primary.
   - green is a fallback family   → set it as `active_override_*` (the serving
                                    worker); leave primary/secondary so c4d
                                    resumes as primary when its capacity returns.
6. CI drains (waits active_jobs==0, bounded) + stops the retired VM — never the
   promoted green.
7. Last resort (no separate green can start, e.g. deep stockout): in-place restart
   of the current serving override — a brief blip, covered by client auto-resubmit.

Override routing keys off an explicit is_primary flag (not list position), so a fallback ranked ahead of a stocked-out primary is still routed correctly. Pure decision logic + unit tests: infrastructure/encoding-worker/deploy_promote.py.

Video Worker Orchestrator

The Video Worker uses an orchestrator pattern to ensure all features work regardless of encoding backend (local Cloud Run or GCE VM).

┌─────────────────────────────────────────────────────────────┐
│                 VideoWorkerOrchestrator                      │
│                                                             │
│  Stage 1: PACKAGING        Stage 2: ENCODING                │
│  ├─ CDG file generation    ├─ Local (FFmpeg)               │
│  └─ TXT lyrics export      └─ GCE (remote VM)              │
│                                                             │
│  Stage 3: ORGANIZATION     Stage 4: DISTRIBUTION           │
│  ├─ Final path setup       ├─ Dropbox upload               │
│  └─ File organization      ├─ Google Drive upload          │
│                            └─ YouTube upload               │
│                                                             │
│  Stage 5: NOTIFICATIONS                                     │
│  ├─ Discord webhooks                                        │
│  └─ Email notifications (completion, reminders)             │
└─────────────────────────────────────────────────────────────┘

Key Design Decision: The orchestrator coordinates stages rather than embedding logic in monolithic functions. This:

  • Ensures all features work with both local and GCE encoding
  • Makes each stage independently testable (139 new tests)
  • Allows feature flag rollback via USE_NEW_ORCHESTRATOR env var

Service Modules:

Module Purpose
youtube_download_service.py YouTube audio downloads (via remote flacfetch)
flacfetch_client.py Client for remote flacfetch HTTP API
audio_search_service.py Audio source search (YouTube, RED, OPS, Spotify)
youtube_upload_service.py YouTube OAuth flow and upload
discord_service.py Discord webhook notifications
packaging_service.py CDG/TXT package generation
local_encoding_service.py FFmpeg encoding (Cloud Run)
encoding_interface.py Abstract interface + GCE backend
video_worker_orchestrator.py Stage coordination
email_service.py Postmark email delivery with CC support
template_service.py GCS-backed email templates
job_notification_service.py Email orchestration (completion, reminders)
audio_transcoding_service.py Transcode FLAC → OGG Opus for review UI playback
youtube_quota_service.py Track YouTube API quota (GCP Cloud Monitoring + Firestore pending buffer)
email_validation_service.py Email validation, tiered disposable domain detection (static list + DeBounce + verifymail.io APIs), blocklist management
youtube_upload_queue_service.py Deferred YouTube upload queue management
catalog_proxy_service.py Proxy to karaoke-decide catalog API (artist/track search)
karaokenerds_service.py Scrape karaokenerds.com for community karaoke versions

Audio Source Download Flow

Audio can come from file upload or remote search (YouTube, torrents). All YouTube downloads route through YouTubeDownloadService to ensure consistent handling:

┌──────────────────────────────────────────────────────────────────────┐
│                     Audio Source Entry Points                         │
│                                                                      │
│  1. File Upload (/api/jobs/upload)                                   │
│     └─ Direct to GCS → triggers workers                              │
│                                                                      │
│  2. YouTube URL (/api/jobs/create-from-url)                          │
│     └─ YouTubeDownloadService → GCS → triggers workers               │
│                                                                      │
│  3. Audio Search + Select (/api/audio-search/...)                    │
│     ├─ YouTube  → YouTubeDownloadService → GCS → triggers workers    │
│     ├─ RED/OPS  → FlacfetchClient (torrent) → GCS → triggers workers │
│     └─ Spotify  → FlacfetchClient (spotify) → GCS → triggers workers │
└──────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│                    YouTubeDownloadService                             │
│                                                                      │
│  • Single entry point for ALL YouTube downloads in cloud             │
│  • Uses remote flacfetch when FLACFETCH_API_URL is set (recommended) │
│  • Falls back to local yt_dlp with warning (usually fails on CR)     │
│  • Handles all URL formats (watch, youtu.be, shorts, embed)          │
└──────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│                      Flacfetch VM (GCE)                               │
│                                                                      │
│  • Has YouTube cookies (avoids bot detection)                        │
│  • Has tracker credentials (RED, OPS private trackers)               │
│  • Runs BitTorrent client for seeding                                │
│  • Uploads directly to GCS (no data transfer through Cloud Run)      │
└──────────────────────────────────────────────────────────────────────┘

Key environment variables:

  • FLACFETCH_API_URL - URL of flacfetch VM (e.g., http://10.x.x.x:8080)
  • FLACFETCH_API_KEY - API key for authentication

Worker Execution

Workers run via two mechanisms depending on their processing duration:

Cloud Run Jobs (Long-Running Workers)

Audio and lyrics workers run as Cloud Run Jobs - standalone batch containers that run to completion without HTTP request lifecycle concerns. This eliminates the instance termination issue where Cloud Run would shut down instances mid-processing.

┌─────────────────────────────────────────────────────────────────┐
│  Cloud Run Jobs (via WorkerService)                             │
│                                                                 │
│  audio-download-job          - 30s-5 min (flacfetch/YouTube)     │
│  lyrics-transcription-job    - 5-15 min (AudioShake + correction)│
│  audio-separation-job        - 10-20 min (L4 GPU, us-east4, direct) │
│  video-encoding-job          - up to 60 min (optional)          │
│                                                                 │
│  Triggered via: google.cloud.run_v2.JobsClient.run_job()        │
│  Job ID passed as: --job-id argument                            │
└─────────────────────────────────────────────────────────────────┘

Why Cloud Run Jobs: When using FastAPI BackgroundTasks, Cloud Run would terminate instances when HTTP requests completed, even if background work was still running. Cloud Run Jobs solve this by running workers as standalone processes that complete naturally. This applies to audio downloads (jobs 51b8231d, 89e497b1 were killed mid-download), lyrics transcription (job c94cc9d6 killed mid-processing), and audio separation.

CLI Entry Points: Each worker has a main() function for Cloud Run Job execution:

  • python -m backend.workers.audio_download_worker --job-id abc123
  • python -m backend.workers.lyrics_worker --job-id abc123
  • python -m backend.workers.audio_worker --job-id abc123
  • python -m backend.workers.video_worker --job-id abc123

HTTP Workers (Fast Workers)

Screens and render workers run via internal HTTP endpoints with BackgroundTasks, as they complete quickly (~30 seconds). They still use the WorkerRegistry for coordination:

┌─────────────────────────────────────────────────────────────────┐
│                       WorkerRegistry                             │
│                                                                 │
│  register(job_id, "screens")   unregister(job_id, "screens")   │
│  register(job_id, "render")    unregister(job_id, "render")    │
│                                                                 │
│  has_active_workers() → bool                                    │
│  wait_for_completion(timeout) → bool                            │
└─────────────────────────────────────────────────────────────────┘

Shutdown Handler (in main.py lifespan):

  • On shutdown signal, checks worker_registry.has_active_workers()
  • If workers active, calls wait_for_completion(timeout=600) (10 min max)
  • Logs which workers are still running for debugging

Audio Separation Architecture

Audio stem separation runs directly inside the audio worker Cloud Run Job — the Separator class from audio-separator is imported and called in-process. There is no separate GPU microservice.

┌──────────────────────────────────────────────────────────────────┐
│  audio-separation-job  (Cloud Run Job, L4 GPU, us-east4)         │
│                                                                  │
│  AudioWorker                                                     │
│  ├── download audio from GCS                                     │
│  ├── create_audio_processor()  ← Separator(model_file_dir=...)   │
│  ├── Stage 1: instrumental_clean preset                          │
│  │   └── produces: vocals.flac, instrumental_clean.flac          │
│  ├── Stage 2: karaoke preset                                     │
│  │   └── produces: instrumental_with_backing.flac                │
│  └── upload stems to GCS                                         │
└──────────────────────────────────────────────────────────────────┘

Key decisions:

  • Direct GPU — No HTTP hop to a separate service. Jobs can run up to 24h; no timeout concerns.
  • Two-stage ensemble preset — instrumental_clean (stage 1) + karaoke (stage 2).
  • Model cache — Models are downloaded to model_file_dir (GCS-backed or local) and reused across executions.

Rollback: The old audio-separator Cloud Run Service (HTTP endpoint) is kept for rollback but no longer receives traffic. It will be removed in a future phase.

Multitenancy

The platform supports white-label B2B portals where business customers get their own branded karaoke generation experience.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  Tenant Detection Flow                                          │
│                                                                 │
│  Request arrives at vocalstar.nomadkaraoke.com                  │
│         │                                                       │
│         ▼                                                       │
│  Frontend: detectTenantFromUrl()                                │
│  - Extracts subdomain from hostname                             │
│  - Calls GET /api/tenant/config?tenant=vocalstar                │
│         │                                                       │
│         ▼                                                       │
│  Backend: TenantMiddleware                                      │
│  1. X-Tenant-ID header (from frontend)                          │
│  2. Query param (dev only, disabled in production)              │
│  3. Host header subdomain detection                             │
│         │                                                       │
│         ▼                                                       │
│  TenantService.get_tenant_config()                              │
│  - Loads from GCS: tenants/{tenant_id}/config.json              │
│  - Caches in memory (5 min TTL)                                 │
│         │                                                       │
│         ▼                                                       │
│  Request proceeds with tenant context                           │
│  - request.state.tenant_id                                      │
│  - request.state.tenant_config                                  │
└─────────────────────────────────────────────────────────────────┘

GCS Storage Layout

tenants/{tenant_id}/
├── config.json          # TenantConfig with branding, features, defaults
└── logo.jpg             # Tenant logo (optional)

themes/{theme_id}/
├── style_params.json    # Full theme configuration
└── assets/
    ├── intro_background.png
    ├── karaoke_background.jpg
    ├── end_background.png
    ├── font.ttf
    └── cdg_*.gif

Tenant Config Schema

TenantConfig:
  id: str                    # e.g., "vocalstar"
  name: str                  # Display name: "Vocal Star"
  subdomain: str             # e.g., "vocalstar.nomadkaraoke.com"
  is_active: bool            # Enable/disable tenant
  branding:
    logo_url: str | None     # GCS path to logo
    logo_height: int         # Pixels
    primary_color: str       # Hex color
    secondary_color: str
    accent_color: str | None
    background_color: str | None
    favicon_url: str | None
    site_title: str
    tagline: str | None
  features:
    audio_search: bool       # Enable audio search (vs file upload only)
    file_upload: bool
    youtube_url: bool
    youtube_upload: bool     # YouTube distribution
    dropbox_upload: bool
    gdrive_upload: bool
    theme_selection: bool    # Allow user to pick theme
    color_overrides: bool
    enable_cdg: bool
    enable_4k: bool
    admin_access: bool
  defaults:
    theme_id: str | None     # Default theme if not locked
    locked_theme: str | None # Force this theme (user cannot change)
    distribution_mode: str   # "all", "download_only", etc.
  auth:
    allowed_email_domains: list[str]  # e.g., ["vocal-star.com"]
    require_email_domain: bool
    sender_email: str | None  # Override email sender

Feature Enforcement

Frontend: The useTenant() hook provides features object. UI components conditionally render based on enabled features:

const { features } = useTenant()
if (!features.audio_search) {
  // Hide audio search tab
}

Backend: Routes check tenant config from request state:

tenant_config = get_tenant_config_from_request(request)
if tenant_config and not tenant_config.features.audio_search:
    raise HTTPException(403, "Audio search not enabled for this tenant")

Active Tenants

Vocal Star (vocalstar.nomadkaraoke.com)

  • Features: File upload only (no audio search, no YouTube URL), streamlined 4-field form (Artist, Title, Mixed Audio, Instrumental Audio)
  • Distribution: Cloud only (Dropbox + Google Drive to tenant-specific folders, no YouTube)
  • Theme: Locked to "vocalstar" theme (yellow/blue)
  • Auth: Restricted to @vocal-star.com and @vocalstarmusic.com emails
  • All jobs automatically private with VSTAR brand prefix
  • Setup: python scripts/setup-vocalstar-tenant.py

Singa (singa.nomadkaraoke.com)

  • Features: File upload only (no audio search, no YouTube URL)
  • Distribution: Download only
  • Theme: Locked to "singa" theme (green/black)
  • Auth: Restricted to @singa.com emails
  • Setup: python scripts/setup-singa-tenant.py

Tenant Portal Frontend (B2B Deployment)

B2B tenants are served from a separate Cloudflare Pages project (karaoke-gen-tenant) built with NEXT_PUBLIC_TENANT_PORTAL=true. This build flag enables tree-shaking of consumer-only content (landing page, pricing) at build time, keeping the bundle lean for white-label deployments.

Runtime tenant injection: A Cloudflare Pages edge function intercepts HTML responses and injects tenant-specific configuration before the page reaches the browser:

// Injected by edge function into <head>
window.__TENANT_CONFIG__ = { id: "vocalstar", branding: { ... }, features: { ... } }
// CSS variables for dynamic theming
--tenant-primary-color: #ffd700;
--tenant-background-color: #1a1a2e;

The frontend's TenantProvider reads window.__TENANT_CONFIG__ on startup (no API round-trip needed for initial render). This approach works for both custom domains (vocalstar.nomadkaraoke.com) and the shared tenant portal domain.

Magic link emails: When a tenant user requests a magic link, the email uses the tenant-specific URL (e.g., https://vocalstar.nomadkaraoke.com/auth/verify?token=...) rather than the main app URL. The job_notification_service.py resolves the correct base URL from the tenant config's subdomain field.

Admin preview: Admins can preview any tenant's branding in the main app by appending ?preview_tenant=<tenant_id> to any URL. This bypasses subdomain detection and loads the specified tenant config, useful for verifying branding before DNS cutover.

Tech Stack

  • Backend: FastAPI, Python 3.12, Cloud Run
  • Frontend: Next.js, TypeScript, Cloudflare Pages
  • Database: Firestore
  • Storage: Google Cloud Storage
  • Secrets: Google Secret Manager
  • Task Queues: Cloud Tasks (idle reminders, YouTube upload queue)
  • Scheduling: Cloud Scheduler (hourly YouTube queue processing)
  • Email: Postmark
  • IaC: Pulumi
  • CI/CD: GitHub Actions