← All incidents

[Eval][v3][zh] Third-party payment gateway timeouts cascade into checkout outage

service: checkout-svccreated: 6/2/2026, 11:59:40 PM

Raw incident context

Time: 18:40 UTC. checkout-svc starts returning HTTP 504 to ~78% of checkout attempts.

Symptoms:
- p99 latency: 28s (hitting our 30s gateway timeout)
- Success rate: 22% (down from 99.5%)
- Thread pool utilization: 100% (all 200 worker threads blocked)
- Inbound queue depth: 4200 requests (queue limit 5000)

Downstream calls (from APM):
- Stripe Connect API (/v1/payment_intents): p99 jumped from 800ms to 27s
- Stripe status page: "Investigating elevated latency for Connect endpoints in us-east-1" (posted 18:35 UTC)
- All other downstream deps (auth-svc, fraud-svc) normal

Our config:
- Stripe call timeout: 30s (matches inbound)
- No circuit breaker on Stripe client
- No bulkhead — Stripe calls share the main worker thread pool

Customer impact:
- Failed checkouts: ~3000 in 5min
- Estimated lost GMV: $180k (so far)

On-call:
- 18:38 — pager
- 18:41 — confirmed Stripe is the culprit (status page + APM)
- 18:43 — debating: wait it out vs disable Stripe path entirely

Summary

SEV1

结账服务 checkout-svc 因下游 Stripe Connect API 延迟飙升导致线程池耗尽,成功率从 99.5% 骤降至 22%,p99 延迟达到 28 秒(接近 30 秒超时),约 3000 笔交易失败,预计损失 GMV 18 万美元。

Severity reasoning: 用户可见的结账失败率高达 78%,超过 1% 阈值且持续超过 5 分钟;收入路径(支付)完全中断,符合 SEV1 定义。

deepseek-chat·prompt v3·output: zh·11535ms·2413↑ / 1771↓ tok·$0.00260

Root cause hypotheses

  • highStripe Connect API 在 us-east-1 区域出现延迟飙升,导致 checkout-svc 的 HTTP 调用阻塞线程池。

    Evidence: APM 显示 Stripe /v1/payment_intents 的 p99 从 800ms 升至 27s;Stripe 状态页确认 'Investigating elevated latency for Connect endpoints in us-east-1'(18:35 UTC)。

  • highcheckout-svc 未对 Stripe 调用设置熔断器或隔离机制,导致单个下游故障耗尽所有工作线程。

    Evidence: 配置中无 circuit breaker,且 Stripe 调用与主线程池共享(无 bulkhead),线程池利用率 100%,入队队列深度 4200。

  • mediumStripe 调用超时设置(30s)与网关超时一致,导致请求在超时前长时间占用线程。

    Evidence: 配置显示 Stripe 调用超时 30s,与 inbound 网关超时相同,线程被阻塞直至超时。

  • lowStripe 客户端连接池耗尽,导致新请求排队等待连接。

    Evidence: 无直接证据,但线程池耗尽时连接池也可能成为瓶颈。

Investigation checklist

  1. 确认 Stripe 延迟是否仍在持续
    curl -w '%{time_total}' -o /dev/null -s https://api.stripe.com/v1/payment_intents -H 'Authorization: Bearer sk_test_xxx'

    Expected: 响应时间应 < 1s;若 > 5s 则 Stripe 仍异常

  2. 检查线程池和队列深度
    curl http://checkout-svc:8080/actuator/health | jq '.threadPool'

    Expected: activeThreads < 200,queueDepth 接近 0

  3. 查看 Stripe 调用超时日志
    kubectl logs -n prod -l app=checkout-svc --since=10m | grep -i 'stripe.*timeout' | tail -20

    Expected: 大量 'timeout' 日志,确认超时是主要错误

  4. 检查 Stripe 客户端连接池状态
    kubectl exec -n prod deploy/checkout-svc -- curl -s localhost:8080/actuator/metrics/http.client.requests | jq '.measurements[0].value'

    Expected: 连接数应接近配置的最大值,若持续增长则连接池耗尽

Mitigation plan

  • 启用熔断器:在 Stripe 客户端配置熔断器,失败率达到 50% 时快速失败,避免线程池耗尽。

    Risk: 熔断后所有支付请求立即失败,但可保护线程池;需确保熔断器半开恢复机制。

    Rollback: 移除熔断器配置或设置阈值为 100%

  • 降级 Stripe 支付路径:临时禁用 Stripe 支付,仅保留其他支付方式(如 PayPal)。

    Risk: 部分用户无法使用信用卡支付,但可保留结账功能;需确认其他支付方式正常。

    Rollback: 重新启用 Stripe 支付

  • 降低 Stripe 调用超时:从 30s 降至 5s,减少线程阻塞时间。

    Risk: 部分正常请求可能因超时过短而失败,但可缓解线程池压力。

    Rollback: 恢复超时为 30s

Customer impact

约 78% 的结账尝试失败,用户看到 HTTP 504 错误或超时。过去 5 分钟内约 3000 笔交易失败,预计损失 GMV 18 万美元。已确认 Stripe 支付服务异常,其他支付方式可能正常。

Postmortem draft

事后复盘

摘要

[FILL IN] 2023-xx-xx 18:40 UTC,checkout-svc 因下游 Stripe Connect API 延迟飙升导致线程池耗尽,成功率从 99.5% 降至 22%,影响持续约 [FILL IN] 分钟。

时间线(UTC)

  • 18:35 Stripe 状态页发布延迟问题
  • 18:38 收到告警
  • 18:41 确认 Stripe 为根因
  • 18:43 讨论降级方案
  • [FILL IN] 实施熔断器
  • [FILL IN] 恢复

影响

  • 失败结账:约 3000 笔
  • 损失 GMV:约 $180k
  • 影响用户:约 [FILL IN] 人

根因

Stripe Connect API 在 us-east-1 区域延迟飙升,checkout-svc 缺乏熔断器和线程隔离,导致所有工作线程被阻塞。

检测

APM 和 Stripe 状态页及时提供了信号,但告警延迟了 3 分钟。

响应

  • 快速识别根因
  • 降级决策耗时较长

做得好的

  • 及时查看 Stripe 状态页
  • APM 数据清晰

做得不好的

  • 无熔断器
  • 无线程隔离
  • 超时设置过长

行动项

  • [ ] 为 Stripe 客户端添加熔断器(P0)
  • [ ] 实现 bulkhead 隔离(P0)
  • [ ] 缩短 Stripe 超时(P1)
  • [ ] 增加降级自动化(P1)

Follow-ups

  • P0为 Stripe 客户端添加熔断器,配置失败率阈值和半开恢复支付平台团队
  • P0实现线程池隔离(bulkhead),将 Stripe 调用与主线程池分离平台工程团队
  • P1缩短 Stripe 调用超时至 5s,并评估对正常请求的影响支付平台团队
  • P1增加自动降级机制:检测到下游故障时自动切换支付方式SRE 团队
  • P2完善告警规则:增加线程池利用率 > 80% 的告警可观测性团队