代理 503 抖动缓冲：把单次重试升级为多次退避重试策略
问题/场景：上游代理短暂 503 时，仅 1 次固定 2.5s 重试不足以穿越抖动。前置条件：你有代理链路且可观测 503。实施步骤：先确认 transient HTTP 失败模式→将容忍窗口扩到 2-3 次递增延迟→监控成功率与总延迟→再决定默认策略。关键配置：重试次数与 backoff。验证：短时代理抖动期间用户侧失败率下降。风险：重试过多会放大尾延迟。
GITHUBDiscovered 2026-02-15Author hou-rong
Prerequisites
- You operate through an upstream LLM proxy where transient 503s can occur.
- Request failure/success metrics are available for before-after comparison.
Steps
- Capture baseline: count transient HTTP failures and retry outcomes in current config.
- Increase retry strategy to multi-attempt with incremental delays (example pattern: 2.5s → 5s → 7.5s).
- Run controlled outage simulation (or replay real incidents) to compare recovered vs failed requests.
- Tune retry budget to balance availability gains and latency impact.
Commands
openclaw gateway statusopenclaw gateway restartVerify
During brief proxy outages, more requests recover successfully and user-visible raw 503 errors decrease.
Caveats
- Exact retry knobs depend on runtime implementation/version and may require code-level changes（需验证）.
- Retry storms can overload recovering proxies; monitor upstream saturation closely.
Source attribution
This tip is aggregated from community/public sources and preserved with attribution.
Open original source ↗