fix(#2180 P0): VLM/LLM timeout 调大,适配方舟视觉API 24-38s响应 #2180
Reference in New Issue
Block a user
Delete Branch "fix/vlm-timeout-tune"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
背景
P0:staging VLM/文案全链路耗时超标,VLM预期15s实际129s,intent预期6s实际75s,script预期20s实际78s。
根因
运维从 staging 服务器直连 curl 方舟 VLM 视觉 API 实测:HTTP 200 但耗时 24.2s(
x-envoy-upstream-service-time: 23357ms),文本API 3~6s正常。代码中 timeout 设置:
vision_timeout=15s→ 每次15s被截断→3次重试全超时(46.6s)→ lite降级pro又3次超时(76.7s)→ 总计123s决定性证据:每次retry elapsed时间精确等于timeout值(15.1s/25.1s/40.1s),典型客户端read timeout先于服务端返回。
修改
验证
部署后新job应看到:
🚀 预览环境已部署
🗑️ 预览环境已清理
PR #2180 已关闭或合并,对应的预览环境已被清理。