podcastdistributiona/worker/index.ts

import "dotenv/config";
import { getBoss } from "@/lib/queue/pgboss";
import {
  QUEUES,
  generateEpisodePayloadSchema,
  echoPayloadSchema,
  type GenerateEpisodePayload,
  type EchoPayload,
} from "@/lib/queue/jobs";
import { runEpisodeGeneration, refundEpisodeUsage } from "@/lib/ai/pipeline/generate-episode";
import { failEpisode } from "@/lib/episodes/status";
import { recordHeartbeat } from "@/lib/queue/health";

const HEARTBEAT_NAME = "generation-worker";

const CONCURRENCY = Math.max(1, Number(process.env.WORKER_CONCURRENCY ?? "2"));

async function main() {
  const boss = await getBoss({ supervise: true });
  console.log(`[worker] started (concurrency=${CONCURRENCY})`);

  // Liveness heartbeat so the admin health page can tell the worker is running.
  await recordHeartbeat(HEARTBEAT_NAME).catch(() => {});
  setInterval(() => {
    recordHeartbeat(HEARTBEAT_NAME).catch((e) => console.error("[worker] heartbeat failed", e));
  }, 15_000);

  // Proof-of-loop queue used by health checks / verification.
  await boss.work<EchoPayload>(QUEUES.echo, { batchSize: 1 }, async (jobs) => {
    for (const job of jobs) {
      const parsed = echoPayloadSchema.safeParse(job.data);
      if (!parsed.success) {
        console.error("[echo] invalid payload — skipping job", parsed.error.issues);
        continue;
      }
      console.log("[echo]", parsed.data);
    }
  });

  // Episode generation. batchSize 1 = independent retries per job.
  await boss.work<GenerateEpisodePayload>(
    QUEUES.generateEpisode,
    { batchSize: 1, includeMetadata: true },
    async (jobs) => {
      for (const job of jobs) {
        await handleGenerate(job);
      }
    }
  );

  const shutdown = async (signal: string) => {
    console.log(`[worker] ${signal} — shutting down`);
    await boss.stop({ graceful: true });
    process.exit(0);
  };
  process.on("SIGINT", () => void shutdown("SIGINT"));
  process.on("SIGTERM", () => void shutdown("SIGTERM"));
}

async function handleGenerate(job: {
  data: GenerateEpisodePayload;
  retryCount?: number;
  retryLimit?: number;
}) {
  // Runtime-validate the payload at the consume boundary. An invalid payload is
  // not retryable, so skip it rather than throwing into the work loop.
  const parsed = generateEpisodePayloadSchema.safeParse(job.data);
  if (!parsed.success) {
    console.error("[generate] invalid payload — skipping job", parsed.error.issues);
    return;
  }
  const { episodeId, type } = parsed.data;
  try {
    await runEpisodeGeneration(episodeId, type ?? "full");
  } catch (err) {
    const message = err instanceof Error ? err.message : String(err);
    const retryCount = job.retryCount ?? 0;
    const retryLimit = job.retryLimit ?? 0;
    const exhausted = retryCount >= retryLimit;
    console.error(
      `[generate] ${episodeId} failed (attempt ${retryCount}/${retryLimit}): ${message}`
    );
    if (exhausted) {
      // Terminal failure: refund the usage the enqueuing caller reserved up
      // front, so a failed generation doesn't permanently consume quota.
      // (The worker never increments; see lib/usage/meter.ts invariant.)
      try {
        await refundEpisodeUsage(episodeId, type ?? "full");
      } catch (refundErr) {
        console.error(`[generate] ${episodeId} usage refund failed`, refundErr);
      }
      // Raw `message` goes to the job row (admin-only); the user sees generic copy.
      await failEpisode(episodeId, message);
    } else {
      throw err; // let pg-boss retry with backoff
    }
  }
}

main().catch((err) => {
  console.error("[worker] fatal", err);
  process.exit(1);
});