Node.js 多模型轮询与故障切换：重试、熔断和监控实现

const models = [
  {
    id: 'glm-flash',
    model: 'glm-4.7-flash',
    url: 'https://open.bigmodel.cn/api/paas/v4/chat/completions',
    apiKey: process.env.ZHIPU_API_KEY,
    enabled: true,
    failures: 0,
    openUntil: 0,
  },
  {
    id: 'backup',
    model: process.env.BACKUP_MODEL,
    url: process.env.BACKUP_URL,
    apiKey: process.env.BACKUP_API_KEY,
    enabled: true,
    failures: 0,
    openUntil: 0,
  },
];

let cursor = 0;

function healthyModels() {
  const now = Date.now();
  return models.filter(m => m.enabled && m.apiKey && m.url && m.openUntil <= now);
}

function orderedCandidates() {
  const healthy = healthyModels();
  if (!healthy.length) throw new Error('没有可用模型');
  const start = cursor++ % healthy.length;
  return [...healthy.slice(start), ...healthy.slice(0, start)];
}

function isRetryable(status) {
  return status === 408 || status === 429 || status >= 500;
}

async function callOne(config, messages) {
  const response = await fetch(config.url, {
    method: 'POST',
    headers: {
      'Content-Type': 'application/json',
      Authorization: `Bearer ${config.apiKey}`,
    },
    body: JSON.stringify({
      model: config.model,
      messages,
      temperature: 0.2,
      max_tokens: 800,
    }),
    signal: AbortSignal.timeout(30_000),
  });

  const raw = await response.text();
  if (!response.ok) {
    const error = new Error(`${config.id} returned ${response.status}`);
    error.status = response.status;
    error.retryable = isRetryable(response.status);
    error.detail = raw.slice(0, 300);
    throw error;
  }

  const payload = JSON.parse(raw);
  const content = payload.choices?.[0]?.message?.content?.trim();
  if (!content) throw new Error(`${config.id} returned empty content`);
  return { content, usage: payload.usage, responseModel: payload.model };
}

export async function generate(messages) {
  const errors = [];
  for (const model of orderedCandidates()) {
    try {
      const result = await callOne(model, messages);
      model.failures = 0;
      return { ...result, configuredModel: model.id };
    } catch (error) {
      errors.push({ model: model.id, message: error.message });
      model.failures += 1;
      if (model.failures >= 3) {
        model.openUntil = Date.now() + 60_000;
      }
      if (!error.retryable && error.status) break;
    }
  }
  throw new AggregateError(errors, '所有候选模型均调用失败');
}

const wait = ms => new Promise(resolve => setTimeout(resolve, ms));

async function backoff(attempt) {
  const base = Math.min(500 * 2 ** attempt, 8_000);
  await wait(base + Math.random() * 300);
}