接入运维真实 metrics：负载、请求错误率与服务探测

- 新增 GET /admin/ops/metrics 与进程内请求统计中间件 - 运维页展示 CPU/内存/磁盘、24h 曲线与 WS/DB/存储状态 - 去掉未使用的消息队列占位项
2026-07-15 15:01:18 +08:00
parent a1336c630b
commit 9251815695
14 changed files with 797 additions and 248 deletions
@@ -0,0 +1,140 @@
+package metrics
+
+import (
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/gin-gonic/gin"
+)
+
+// Collector 进程内请求统计：近 24 小时按整点小时桶累加。
+type Collector struct {
+	mu     sync.Mutex
+	byHour map[int64]*hourBucket
+
+	// 自进程启动以来的累计
+	totalReq   atomic.Int64
+	totalErr   atomic.Int64 // HTTP status >= 400
+	total5xx   atomic.Int64
+	inFlight   atomic.Int64
+	startedAt  time.Time
+}
+
+type hourBucket struct {
+	requests int64
+	errors   int64
+	status5  int64
+}
+
+// Default 全局采集器
+var Default = NewCollector()
+
+func NewCollector() *Collector {
+	return &Collector{
+		byHour:    make(map[int64]*hourBucket),
+		startedAt: time.Now(),
+	}
+}
+
+func hourKey(t time.Time) int64 {
+	return t.Unix() / 3600
+}
+
+func (c *Collector) Record(status int) {
+	c.totalReq.Add(1)
+	isErr := status >= 400
+	is5 := status >= 500
+	if isErr {
+		c.totalErr.Add(1)
+	}
+	if is5 {
+		c.total5xx.Add(1)
+	}
+
+	h := hourKey(time.Now())
+	c.mu.Lock()
+	b := c.byHour[h]
+	if b == nil {
+		b = &hourBucket{}
+		c.byHour[h] = b
+		c.pruneLocked(h)
+	}
+	b.requests++
+	if isErr {
+		b.errors++
+	}
+	if is5 {
+		b.status5++
+	}
+	c.mu.Unlock()
+}
+
+func (c *Collector) pruneLocked(nowHour int64) {
+	cutoff := nowHour - 25
+	for k := range c.byHour {
+		if k < cutoff {
+			delete(c.byHour, k)
+		}
+	}
+}
+
+// Snapshot24h 返回近 24 个整点小时（含当前小时）的请求量与错误率（0–100）。
+// 下标 0 为 23 小时前，23 为当前小时；hours 为当地小时 0–23。
+func (c *Collector) Snapshot24h() (requests []int64, errorRates []float64, hours []int, totalReq, totalErr int64) {
+	now := time.Now()
+	nowH := hourKey(now)
+	loc := now.Location()
+
+	requests = make([]int64, 24)
+	errorRates = make([]float64, 24)
+	hours = make([]int, 24)
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	for i := 0; i < 24; i++ {
+		hKey := nowH - int64(23-i)
+		t := time.Unix(hKey*3600, 0).In(loc)
+		hours[i] = t.Hour()
+		b := c.byHour[hKey]
+		var r, e int64
+		if b != nil {
+			r, e = b.requests, b.errors
+		}
+		requests[i] = r
+		if r > 0 {
+			errorRates[i] = float64(e) / float64(r) * 100
+		} else {
+			errorRates[i] = 0
+		}
+		totalReq += r
+		totalErr += e
+	}
+	return
+}
+
+func (c *Collector) Totals() (req, err4, err5 int64, startedAt time.Time) {
+	return c.totalReq.Load(), c.totalErr.Load(), c.total5xx.Load(), c.startedAt
+}
+
+func (c *Collector) InFlight() int64 {
+	return c.inFlight.Load()
+}
+
+// Middleware 记录每个 HTTP 请求的状态码（跳过 WebSocket 升级失败后的统计由 status 决定）。
+func Middleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		// 健康检查可不计入业务错误曲线，但仍记请求量更直观；这里全部计入
+		Default.inFlight.Add(1)
+		c.Next()
+		Default.inFlight.Add(-1)
+
+		// WebSocket 升级成功后 status 可能是 101
+		status := c.Writer.Status()
+		if status == 0 {
+			status = 200
+		}
+		Default.Record(status)
+	}
+}
@@ -0,0 +1,32 @@
+package metrics
+
+import (
+	"testing"
+)
+
+func TestRecordAndSnapshot(t *testing.T) {
+	c := NewCollector()
+	c.Record(200)
+	c.Record(200)
+	c.Record(500)
+	c.Record(404)
+
+	reqs, rates, hours, totalReq, totalErr := c.Snapshot24h()
+	if len(reqs) != 24 || len(rates) != 24 || len(hours) != 24 {
+		t.Fatalf("snapshot 长度应为 24: req=%d rates=%d hours=%d", len(reqs), len(rates), len(hours))
+	}
+	if totalReq != 4 || totalErr != 2 {
+		t.Fatalf("窗口统计错误: totalReq=%d totalErr=%d", totalReq, totalErr)
+	}
+	// 当前小时桶应有 4 请求
+	if reqs[23] != 4 {
+		t.Fatalf("当前小时请求量 = %d, 期望 4", reqs[23])
+	}
+	if rates[23] < 49 || rates[23] > 51 {
+		t.Fatalf("当前小时错误率 = %v, 期望约 50", rates[23])
+	}
+	tr, te, t5, _ := c.Totals()
+	if tr != 4 || te != 2 || t5 != 1 {
+		t.Fatalf("累计统计错误: req=%d err=%d 5xx=%d", tr, te, t5)
+	}
+}