feat: alert on gateway p99 latency

deploy/compose_test.go

@@ -50,6 +50,7 @@ type alertRulesDocument struct {
 	Groups []struct {
 		Rules []struct {
 			Expression string `yaml:"expr"`
+			Alert      string `yaml:"alert"`
 		} `yaml:"rules"`
 	} `yaml:"groups"`
 }
@@ -297,9 +298,11 @@ func TestObservabilityAssetsUseRegisteredLowCardinalityMetrics(t *testing.T) {
 	if err := yaml.Unmarshal(rulesPayload, &rules); err != nil {
 		t.Fatalf("parse Prometheus rules: %v", err)
 	}
+	alerts := make([]string, 0)
 	for _, group := range rules.Groups {
 		for _, rule := range group.Rules {
 			expressions = append(expressions, rule.Expression)
+			alerts = append(alerts, rule.Alert)
 		}
 	}
 	metricPattern := regexp.MustCompile(`proxy_pool_[a-z0-9_]+`)
@@ -324,6 +327,9 @@ func TestObservabilityAssetsUseRegisteredLowCardinalityMetrics(t *testing.T) {
 			t.Errorf("observability assets do not cover %s", required)
 		}
 	}
+	if !slices.Contains(alerts, "ProxyPoolGatewayHighLatency") {
+		t.Error("Prometheus rules do not alert on Gateway p99 latency")
+	}
 }
 
 func loadComposeDocument(t *testing.T) composeDocument {

deploy/prometheus/rules/proxy-pool.yml

@@ -10,6 +10,15 @@ groups:
           severity: warning
         annotations:
           summary: Gateway 代理尝试失败比例持续高于 10%
+      - alert: ProxyPoolGatewayHighLatency
+        expr: |
+          histogram_quantile(0.99,
+            sum by (le, protocol) (rate(proxy_pool_gateway_request_duration_seconds_bucket[5m]))) > 5
+        for: 10m
+        labels:
+          severity: warning
+        annotations:
+          summary: Gateway p99 请求时延持续高于 5 秒
       - alert: ProxyPoolGatewayOutcomeQueueDrops
         expr: sum(increase(proxy_pool_gateway_outcome_queue_dropped_total[5m])) > 0
         for: 5m

progress.md

@@ -21,7 +21,8 @@
   Gateway、Controller 容量、Provider、Extraction、Checker 与 Drain。部署测试会解析仪表盘和
   规则，拒绝未注册指标以及 `upstream`/`worker` 聚合或 selector 标签，防止观测资产再次漂移。
 - Gateway 新增按固定 `HTTP`/`CONNECT` protocol 标签聚合的请求总耗时 Histogram，覆盖从准入到
-  HTTP 响应或 CONNECT 隧道结束的完整生命周期；Grafana Overview 已恢复该真实指标的 p99 面板。
+  HTTP 响应或 CONNECT 隧道结束的完整生命周期；Grafana Overview 已恢复该真实指标的 p99 面板，
+  Prometheus 以相同固定标签集监控持续高于 5 秒的 p99。
 - 全仓 `go test -count=1 -timeout 60s ./...`、`go vet ./...`、`go build ./...`、
   Protobuf descriptor、Kustomize Base 渲染及开发证书 SAN/SPIFFE 校验均通过。Compose
   容器端到端启动在拉取 Dockerfile 前端与监控镜像时受 Docker Desktop HTTPS 代理缺失阻断，