dragonflight/services/worker/src/index.js

import 'dotenv/config';
import { Worker, Queue } from 'bullmq';
import { proxyWorker, thumbnailQueue as proxyThumbnailQueue } from './workers/proxy.js';
import { thumbnailWorker } from './workers/thumbnail.js';
import { filmstripWorker } from './workers/filmstrip.js';
import { conformWorker } from './workers/conform.js';
import { youtubeImportWorker, proxyQueue as youtubeProxyQueue } from './workers/youtube-import.js';
import { trimWorker } from './workers/trimWorker.js';
import { hlsWorker } from './workers/hls.js';
import { playoutStageWorker } from './workers/playout-stage.js';
import { promotionWorker } from './workers/promotion.js';
import { startPromotionScanner } from './workers/promotion-scanner.js';

const parseRedisUrl = (url) => {
  const parsed = new URL(url);
  return {
    host: parsed.hostname,
    port: parseInt(parsed.port, 10),
    password: parsed.password || undefined,
  };
};

const redisOptions = parseRedisUrl(process.env.REDIS_URL || 'redis://localhost:6379');

// Human-readable node/GPU label stamped onto each job so the Jobs UI can show
// which worker ran it (UI reads metadata.node). Set per container via env.
const WORKER_LABEL = process.env.WORKER_LABEL || process.env.HOSTNAME || 'worker';

const createWorker = (queueName, handler, overrides = {}) => {
  // Stamp node attribution into the job data on pickup (persists to Redis).
  const wrapped = async (job) => {
    try { await job.updateData({ ...job.data, node: WORKER_LABEL }); } catch (_) {}
    return handler(job);
  };
  const worker = new Worker(queueName, wrapped, {
    connection: redisOptions,
    // Stall detection: if a worker dies mid-job, BullMQ moves it back to wait
    // after stalledInterval. Without this a crashed run sits in active forever.
    stalledInterval: 30000,
    maxStalledCount: 1,
    lockDuration: 60000,
    lockRenewTime: 15000,
    ...overrides,
  });

  worker.on('completed', (job) => {
    console.log(`[${queueName}] Job ${job.id} completed`);
  });

  worker.on('failed', (job, err) => {
    console.error(`[${queueName}] Job ${job.id} failed:`, err.message);
  });

  worker.on('stalled', (jobId) => {
    console.warn(`[${queueName}] Job ${jobId} stalled — reclaimed`);
  });

  worker.on('progress', (job, progress) => {
    console.log(`[${queueName}] Job ${job.id} progress:`, progress);
  });

  return worker;
};

// Per-queue concurrency. Defaults to 1, which serialises every job in a
// queue — meaning a single stalled job blocks every other one. We want
// thumbnails (cheap, parallel-safe) to run several at a time so a slow
// outlier doesn't back the rest of the catalog up. Proxy + conform are
// heavier (ffmpeg transcode) so we keep them lower to avoid trashing
// the box; tune via env if a node has more headroom.
const PROXY_CONCURRENCY      = parseInt(process.env.PROXY_CONCURRENCY      || '2', 10);
const THUMBNAIL_CONCURRENCY  = parseInt(process.env.THUMBNAIL_CONCURRENCY  || '4', 10);
const FILMSTRIP_CONCURRENCY  = parseInt(process.env.FILMSTRIP_CONCURRENCY  || '2', 10);
const CONFORM_CONCURRENCY    = parseInt(process.env.CONFORM_CONCURRENCY    || '1', 10);
const TRIM_CONCURRENCY       = parseInt(process.env.TRIM_CONCURRENCY       || '4', 10);

// Capability routing: a worker only subscribes to the queues named in
// WORKER_QUEUES (comma-separated). Unset = all queues (back-compat). This lets
// us pin heavy encodes (proxy/conform/trim) to strong GPUs and light jobs
// (thumbnail/filmstrip) to weak ones, each container bound to one GPU.
const _wq = (process.env.WORKER_QUEUES || '').trim();
const _enabled = _wq ? new Set(_wq.split(',').map(x => x.trim()).filter(Boolean)) : null;
const want = (q) => !_enabled || _enabled.has(q);

const workers = [
  want('proxy')     && createWorker('proxy',     proxyWorker,     { concurrency: PROXY_CONCURRENCY }),
  want('thumbnail') && createWorker('thumbnail', thumbnailWorker, { concurrency: THUMBNAIL_CONCURRENCY }),
  want('filmstrip') && createWorker('filmstrip', filmstripWorker, { concurrency: FILMSTRIP_CONCURRENCY }),
  want('conform')   && createWorker('conform',   conformWorker,   { concurrency: CONFORM_CONCURRENCY }),
  want('trim')      && createWorker('trim',      trimWorker,      { concurrency: TRIM_CONCURRENCY }),
  // HLS backfill remux is a light stream-copy. Run it wherever proxy runs so
  // existing proxy nodes pick up reprocess?type=hls jobs without an env change.
  (want('proxy') || want('hls')) && createWorker('hls', hlsWorker, { concurrency: 2 }),
  want('import')    && createWorker('import', youtubeImportWorker, {
    concurrency: 1,
    lockDuration: 10 * 60 * 1000,
    lockRenewTime: 60000,
  }),
  // playout-stage = S3 → /media volume + EBU R128 loudnorm. CPU/IO-bound;
  // colocate with workers that already have ffmpeg + the media mount.
  want('playout-stage') && createWorker('playout-stage', playoutStageWorker, { concurrency: 1 }),
  // promotion = growing-files promotion (S3 upload + DB update + queue proxy).
  // Triggered manually via POST /assets/:id/promote AND automatically by the
  // promotion scanner below once a pending_migration asset has been idle for
  // settings.growing_promote_after_seconds.
  want('promotion') && createWorker('promotion', promotionWorker, { concurrency: 1 }),
].filter(Boolean);
console.log(`WORKER_QUEUES=${_wq || '(all)'}`);

// Auto-promotion scanner — only on promotion-capable workers, and only ONE
// instance is needed cluster-wide, but the scan is idempotent (status guard +
// stable jobId) so running it on every promotion worker is safe.
let _promotionScanner = null;
if (want('promotion')) {
  _promotionScanner = startPromotionScanner(redisOptions);
}

// Filmstrip queue singleton — used by thumbnail worker to enqueue filmstrip jobs
export const filmstripQueue = new Queue('filmstrip', { connection: redisOptions });

console.log(`Concurrency: proxy=${PROXY_CONCURRENCY} thumbnail=${THUMBNAIL_CONCURRENCY} filmstrip=${FILMSTRIP_CONCURRENCY} conform=${CONFORM_CONCURRENCY} trim=${TRIM_CONCURRENCY} import=1`);

console.log('Wild Dragon Worker Service started');
console.log(`Redis: ${redisOptions.host}:${redisOptions.port}`);
console.log('Active queues: proxy, thumbnail, conform, trim, import, promotion');

process.on('SIGTERM', async () => {
  console.log('SIGTERM received, shutting down...');

  // BUG FIX #4 + #10: Close all BullMQ Workers AND all Queue client instances
  // on SIGTERM. Workers process jobs; Queues dispatch them. Both hold open
  // Redis connections that keep the event loop alive after workers are closed,
  // causing the process to hang indefinitely unless process.exit() is called.
  // Explicitly closing every Queue allows the event loop to drain naturally.
  await Promise.all([
    // Close all Worker instances (stops accepting new jobs, waits for active)
    ...workers.map(w => w.close()),
    // BUG FIX #7: Close the Queue singletons from worker modules.
    // proxyThumbnailQueue: thumbnailQueue in proxy.js (dispatches thumbnail jobs)
    // youtubeProxyQueue:   proxyQueue in youtube-import.js (dispatches proxy jobs)
    proxyThumbnailQueue.close().catch(() => {}),
    youtubeProxyQueue.close().catch(() => {}),
    filmstripQueue.close().catch(() => {}),
    _promotionScanner ? _promotionScanner.promotionQueue.close().catch(() => {}) : Promise.resolve(),
  ]);

  console.log('All workers and queues closed');
  process.exit(0);
});