perf(queue): 限制大媒体任务内存并消除准入惊群

将同步与异步非文本任务的准入唤醒改为按任务和全局队首推进，批量续租等待者，避免大量请求同时争抢 advisory lock 和数据库连接。对 Base64 请求解析、素材解码、上游媒体执行和结果物化增加分层并发限制，复用请求素材并释放重复 Gemini wire 数据；生产 API 默认入口解析并发 16，媒体物化并发 8。新增 1000 个同步 Gemini 图像编辑请求的模拟上游压力验收，10 MiB 输入和输出下全部成功，Heap 峰值增长约 2.58 GiB，并验证共享上传哈希、单次 attempt 与本地零落盘。验证：go test ./... -count=1；go vet ./...；gofmt -l 无输出；kubectl kustomize deploy/kubernetes/production；10 MiB Gemini Base64 千任务压力测试通过。
2026-07-30 13:13:32 +08:00
parent fd3b6bf042
commit ea58d21d03
26 changed files with 1980 additions and 74 deletions
@@ -402,6 +402,59 @@ VALUES ($1::uuid, 1, $2::uuid, $3::uuid, $4, 'running', 'submitting')`,
 		t.Fatalf("post-submission cancellation changed=%v err=%v, want unchanged", changed, err)
 	}

+	queuedAtomicTask := createTask(true)
+	queuedHookFailure := errors.New("synthetic queued hook failure")
+	_, err = first.QueueTaskAdmissionWithHook(ctx, inputFor(queuedAtomicTask, 100, ""), func(pgx.Tx) error {
+		return queuedHookFailure
+	})
+	if !errors.Is(err, queuedHookFailure) {
+		t.Fatalf("queued hook failure = %v, want synthetic failure", err)
+	}
+	var queuedAdmissions, queuedLeases int
+	if err := first.pool.QueryRow(ctx, `
+SELECT
+  (SELECT COUNT(*) FROM gateway_task_admissions WHERE task_id = $1::uuid),
+  (SELECT COUNT(*) FROM gateway_concurrency_leases WHERE task_id = $1::uuid AND released_at IS NULL)`,
+		queuedAtomicTask.ID,
+	).Scan(&queuedAdmissions, &queuedLeases); err != nil {
+		t.Fatalf("read rolled back queued hook state: %v", err)
+	}
+	if queuedAdmissions != 0 || queuedLeases != 0 {
+		t.Fatalf("failed queued hook left admissions=%d leases=%d", queuedAdmissions, queuedLeases)
+	}
+	const queuedSyntheticRiverJobID int64 = 987654320
+	queuedAdmission, err := first.QueueTaskAdmissionWithHook(ctx, inputFor(queuedAtomicTask, 100, ""), func(tx pgx.Tx) error {
+		_, hookErr := tx.Exec(ctx, `
+UPDATE gateway_tasks
+SET river_job_id = $2
+WHERE id = $1::uuid`, queuedAtomicTask.ID, queuedSyntheticRiverJobID)
+		return hookErr
+	})
+	if err != nil || queuedAdmission.Status != "waiting" {
+		t.Fatalf("atomic queued admission=%+v err=%v, want waiting", queuedAdmission, err)
+	}
+	var riverJobID int64
+	if err := first.pool.QueryRow(ctx, `
+SELECT
+  river_job_id,
+  (SELECT COUNT(*)
+   FROM gateway_concurrency_leases
+   WHERE task_id = $1::uuid AND released_at IS NULL)
+FROM gateway_tasks
+WHERE id = $1::uuid`, queuedAtomicTask.ID).Scan(&riverJobID, &queuedLeases); err != nil {
+		t.Fatalf("read atomic queued River marker: %v", err)
+	}
+	if riverJobID != queuedSyntheticRiverJobID || queuedLeases != 0 {
+		t.Fatalf("queued River marker=%d leases=%d, want %d/0", riverJobID, queuedLeases, queuedSyntheticRiverJobID)
+	}
+	result, err = second.TryTaskAdmission(ctx, inputFor(queuedAtomicTask, 100, ""))
+	if err != nil || !result.Admitted || len(result.Leases) != 1 {
+		t.Fatalf("worker-time queued admission result=%+v err=%v", result, err)
+	}
+	if err := first.DeleteTaskAdmission(ctx, queuedAtomicTask.ID); err != nil {
+		t.Fatalf("release queued atomic task: %v", err)
+	}
+
 	atomicTask := createTask(true)
 	hookFailure := errors.New("synthetic admitted hook failure")
 	_, err = first.TryTaskAdmissionWithAdmittedHook(ctx, inputFor(atomicTask, 100, ""), func(pgx.Tx) error {
@@ -433,7 +486,6 @@ WHERE id = $1::uuid`, atomicTask.ID, syntheticRiverJobID)
 	if err != nil || !result.Admitted {
 		t.Fatalf("atomic admitted hook result=%+v err=%v", result, err)
 	}
-	var riverJobID int64
 	if err := first.pool.QueryRow(ctx, `
 SELECT river_job_id
 FROM gateway_tasks