fix(cluster): 进一步收紧媒体 Worker 并发上限

全局 48 槽在持续观察中仍导致宁波 Worker OOM。按实测内存占用将生产全局异步执行上限降至 16，双节点各 8 槽，单节点接管最多 16 槽；高并发请求继续由持久队列吸收。\n\n未开启 queue_size 生产策略。验证：kubectl kustomize、服务端 dry-run。
2026-07-29 18:51:26 +08:00
parent 8ce1f857a4
commit 4371e57212
1 changed files with 1 additions and 1 deletions
@@ -35,5 +35,5 @@ data:
  AI_GATEWAY_LOCAL_RESULT_MAX_BYTES: "268435456"
  AI_GATEWAY_LOCAL_RESULT_MAX_TASK_BYTES: "536870912"
  AI_GATEWAY_ASYNC_QUEUE_WORKER_ENABLED: "true"
-  AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT: "48"
+  AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT: "16"
  AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS: "5"