语音 AI 限流：并发、队列与 429 错误

import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";

const elevenlabs = new ElevenLabsClient({ apiKey: process.env.ELEVENLABS_API_KEY });

async function pool<T, R>(
  items: T[],
  maxInFlight: number,
  worker: (item: T) => Promise<R>,
): Promise<R[]> {
  const results: R[] = new Array(items.length);
  let next = 0;

  async function run(): Promise<void> {
    while (next < items.length) {
      const i = next++;
      results[i] = await worker(items[i]); // never more than maxInFlight of these run at once
    }
  }

  await Promise.all(
    Array.from({ length: Math.min(maxInFlight, items.length) }, run),
  );
  return results;
}

async function synthesize(text: string): Promise<Buffer> {
  const stream = await elevenlabs.textToSpeech.stream("JBFqnCBsd6RMkjVDRZzb", {
    text,
    modelId: "eleven_flash_v2_5",
    outputFormat: "mp3_44100_128",
  });
  const chunks: Buffer[] = [];
  for await (const chunk of stream) chunks.push(Buffer.from(chunk));
  return Buffer.concat(chunks);
}

// Plan Flash limit is, say, 10. Stay under it.
const texts = Array.from({ length: 50 }, (_, i) => `Sentence number ${i}.`);
const audio = await pool(texts, 8, synthesize); // never more than 8 in flight

class TokenBucket {
  private tokens: number;
  private updated = performance.now();

  constructor(private capacity: number, private refillPerSec: number) {
    this.tokens = capacity;
  }

  private refill(): void {
    const now = performance.now();
    const elapsed = (now - this.updated) / 1000;
    this.tokens = Math.min(this.capacity, this.tokens + elapsed * this.refillPerSec);
    this.updated = now;
  }

  tryAcquire(cost = 1): boolean {
    this.refill();
    if (this.tokens >= cost) {
      this.tokens -= cost;
      return true;
    }
    return false;
  }

  timeUntil(cost = 1): number {
    this.refill();
    return this.tokens >= cost ? 0 : ((cost - this.tokens) / this.refillPerSec) * 1000;
  }
}

class LeakyBucket {
  private next = performance.now();
  constructor(private intervalMs: number) {} // admit at most one item per intervalMs

  async acquire(): Promise<void> {
    const now = performance.now();
    const wait = Math.max(0, this.next - now);
    this.next = Math.max(now, this.next) + this.intervalMs;
    if (wait > 0) await new Promise((r) => setTimeout(r, wait));
  }
}

async function withBackoff<T>(
  call: () => Promise<T>,
  opts: { maxAttempts?: number; baseMs?: number; capMs?: number } = {},
): Promise<T> {
  const { maxAttempts = 5, baseMs = 500, capMs = 20_000 } = opts;
  let attempt = 0;
  for (;;) {
    try {
      return await call();
    } catch (e) {
      if (!(e instanceof RetryableError) || ++attempt >= maxAttempts) throw e;
      // honor Retry-After if present; otherwise capped exponential growth with full jitter
      const delay =
        e.retryAfterMs ?? Math.random() * Math.min(capMs, baseMs * 2 ** attempt);
      await new Promise((r) => setTimeout(r, delay));
    }
  }
}

class RetryableError extends Error {
  constructor(public status: number, public retryAfterMs?: number) {
    super(`retryable ${status}`);
  }
}

function classify(resp: Response): void {
  if ([429, 500, 502, 503, 504].includes(resp.status)) {
    const ra = resp.headers.get("retry-after");
    throw new RetryableError(resp.status, ra ? Number(ra) * 1000 : undefined);
  }
  if (!resp.ok) throw new Error(`non-retryable ${resp.status}`);
}

class CircuitBreaker {
  private failures = 0;
  private openedAt: number | null = null;
  constructor(private threshold = 5, private cooldownMs = 10_000) {}

  allow(): boolean {
    if (this.openedAt === null) return true;
    if (performance.now() - this.openedAt >= this.cooldownMs) {
      this.openedAt = null; // half-open: allow a probe
      this.failures = 0;
      return true;
    }
    return false;
  }

  record(ok: boolean): void {
    if (ok) {
      this.failures = 0;
      this.openedAt = null;
    } else if (++this.failures >= this.threshold) {
      this.openedAt = performance.now();
    }
  }
}

class MultiTenantAdmission {
  private tenantBuckets = new Map<string, TokenBucket>();
  constructor(private globalMaxInFlight: number) {}

  private bucket(tenant: string): TokenBucket {
    let b = this.tenantBuckets.get(tenant);
    if (!b) {
      // Each tenant: burst of 5, sustained 2 starts/sec. Tune per tier.
      b = new TokenBucket(5, 2);
      this.tenantBuckets.set(tenant, b);
    }
    return b;
  }

  async run<R>(tenant: string, work: () => Promise<R>): Promise<R> {
    const b = this.bucket(tenant);
    if (!b.tryAcquire()) {
      throw new RetryableError(429, b.timeUntil());
    }
    // ... then admit through the global limiter (e.g. the bounded pool above)
    return work();
  }
}

function recordHeadroom(resp: Response, metrics: Metrics): void {
  const cur = Number(resp.headers.get("current-concurrent-requests"));
  const max = Number(resp.headers.get("maximum-concurrent-requests"));
  if (Number.isFinite(cur) && Number.isFinite(max)) {
    metrics.gauge("el.concurrency.current", cur);
    metrics.gauge("el.concurrency.max", max);
    if (max > 0) metrics.gauge("el.concurrency.utilization", cur / max);
  }
}