perf(worker): 解耦准入调度与远端执行

原因：线上 P24 同构验收确认两台香港 Worker 的全局容量为 48，但跨地域逐行准入事务每轮只能形成 8 个活跃租约，队列最老等待超过 15 分钟。\n\n影响：新增可配置的异步准入 dispatcher 角色；生产两地 API 负责准入和过期回收，Worker 仅执行 River job。当前主库同站点 API 可低延迟填满容量，主库切换后另一地 API 通过既有数据库锁安全接管；未配置环境变量时保持原 Worker 一体化行为。\n\n验证：Go 全量测试、go vet、gofmt、真实 PostgreSQL 1000 任务双进程回归、前端 lint/test/build、Kubernetes server-side dry-run、Compose、ShellCheck、cluster/manual release tests 全部通过。
2026-08-01 23:24:30 +08:00
parent 9c093974a1
commit a4046c2b81
10 changed files with 96 additions and 27 deletions
@@ -274,13 +274,24 @@ func (s *Service) startRiverQueue(ctx context.Context, workerEnabled bool) error
 		return err
 	}
 	go s.refreshAsyncWorkerCapacity(ctx)
-	go s.dispatchWaitingAsyncAdmissions(ctx)
-	go s.reapExpiredTaskAdmissions(ctx)
 	go s.recoverOrphanedAsyncRiverJobs(ctx)
 	go s.stopAsyncWorkersOnShutdown(ctx)
 	return nil
 }

+// StartAsyncAdmissionDispatcher starts the lightweight admission/reaping
+// coordinator independently from River execution workers. This allows the
+// dispatcher to run beside PostgreSQL while Worker pods scale on other nodes;
+// the existing database task/scope locks make multiple API-site dispatchers
+// safe during primary failover.
+func (s *Service) StartAsyncAdmissionDispatcher(ctx context.Context) {
+	s.asyncAdmissionRunner.Do(func() {
+		go s.dispatchWaitingAsyncAdmissions(ctx)
+		go s.reapExpiredTaskAdmissions(ctx)
+		s.logger.Info("asynchronous admission dispatcher started")
+	})
+}
+
 func (s *Service) newRiverAsyncExecutionClient(capacity int) (*river.Client[pgx.Tx], error) {
 	workers := river.NewWorkers()
 	if err := river.AddWorkerSafely(workers, &asyncTaskWorker{service: s}); err != nil {