Improve metrics

Author: dingxiaoshuai123

codis/pkg/proxy/proxy.go

@@ -558,8 +558,8 @@ type Stats struct {
 		PrimaryOnly bool `json:"primary_only"`
 	} `json:"backend"`
 
-	Runtime          *RuntimeStats `json:"runtime,omitempty"`
-	TimeoutCmdNumber int64         `json:"timeout_cmd_number"`
+	Runtime      *RuntimeStats `json:"runtime,omitempty"`
+	SlowCmdCount int64         `json:"slow_cmd_count"` // Cumulative count of slow log
 }
 
 type RuntimeStats struct {
@@ -668,6 +668,6 @@ func (p *Proxy) Stats(flags StatsFlags) *Stats {
 		stats.Runtime.NumCgoCall = runtime.NumCgoCall()
 		stats.Runtime.MemOffheap = unsafe2.OffheapBytes()
 	}
-	stats.TimeoutCmdNumber = TimeoutCmdNumberInSecond.Int64()
+	stats.SlowCmdCount = SlowCmdCount.Int64()
 	return stats
 }

codis/pkg/proxy/session.go

@@ -236,13 +236,14 @@ func (s *Session) loopWriter(tasks *RequestChan) (err error) {
 		} else {
 			s.incrOpStats(r, resp.Type)
 		}
+		nowTime := time.Now().UnixNano()
+		duration := int64((nowTime - r.ReceiveTime) / 1e3)
+		s.updateMaxDelay(duration, r)
 		if fflush {
 			s.flushOpStats(false)
 		}
-		nowTime := time.Now().UnixNano()
-		duration := int64((nowTime - r.ReceiveTime) / 1e3)
 		if duration >= s.config.SlowlogLogSlowerThan {
-			TimeoutCmdNumber.Incr()
+			SlowCmdCount.Incr()
 			//client -> proxy -> server -> porxy -> client
 			//Record the waiting time from receiving the request from the client to sending it to the backend server
 			//the waiting time from sending the request to the backend server to receiving the response from the server
@@ -759,3 +760,10 @@ func (s *Session) handlePConfig(r *Request) error {
 	}
 	return nil
 }
+
+func (s *Session) updateMaxDelay(duration int64, r *Request) {
+	e := s.getOpStats(r.OpStr)
+	if duration > e.maxDelay.Int64() {
+		e.maxDelay.Set(duration)
+	}
+}

codis/pkg/proxy/stats.go

@@ -14,10 +14,7 @@ import (
 	"pika/codis/v2/pkg/utils/sync2/atomic2"
 )
 
-var (
-	TimeoutCmdNumber         atomic2.Int64
-	TimeoutCmdNumberInSecond atomic2.Int64
-)
+var SlowCmdCount atomic2.Int64 // Cumulative count of slow log
 
 type opStats struct {
 	opstr string
@@ -27,14 +24,16 @@ type opStats struct {
 	redis struct {
 		errors atomic2.Int64
 	}
+	maxDelay atomic2.Int64
 }
 
 func (s *opStats) OpStats() *OpStats {
 	o := &OpStats{
-		OpStr: s.opstr,
-		Calls: s.calls.Int64(),
-		Usecs: s.nsecs.Int64() / 1e3,
-		Fails: s.fails.Int64(),
+		OpStr:    s.opstr,
+		Calls:    s.calls.Int64(),
+		Usecs:    s.nsecs.Int64() / 1e3,
+		Fails:    s.fails.Int64(),
+		MaxDelay: s.maxDelay.Int64(),
 	}
 	if o.Calls != 0 {
 		o.UsecsPercall = o.Usecs / o.Calls
@@ -50,6 +49,7 @@ type OpStats struct {
 	UsecsPercall int64  `json:"usecs_percall"`
 	Fails        int64  `json:"fails"`
 	RedisErrType int64  `json:"redis_errtype"`
+	MaxDelay     int64  `json:"max_delay"`
 }
 
 var cmdstats struct {
@@ -67,8 +67,7 @@ var cmdstats struct {
 
 func init() {
 	cmdstats.opmap = make(map[string]*opStats, 128)
-	TimeoutCmdNumber.Set(0)
-	TimeoutCmdNumberInSecond.Set(0)
+	SlowCmdCount.Set(0)
 	go func() {
 		for {
 			start := time.Now()
@@ -77,9 +76,17 @@ func init() {
 			delta := cmdstats.total.Int64() - total
 			normalized := math.Max(0, float64(delta)) * float64(time.Second) / float64(time.Since(start))
 			cmdstats.qps.Set(int64(normalized + 0.5))
+		}
+	}()
 
-			TimeoutCmdNumberInSecond.Swap(TimeoutCmdNumber.Int64())
-			TimeoutCmdNumber.Set(0)
+	go func() {
+		for {
+			time.Sleep(15 * time.Second)
+			cmdstats.Lock()
+			for _, s := range cmdstats.opmap {
+				s.maxDelay = 0
+			}
+			cmdstats.Unlock()
 		}
 	}()
 }
@@ -175,6 +182,18 @@ func incrOpStats(e *opStats) {
 		s.redis.errors.Add(n)
 		cmdstats.redis.errors.Add(n)
 	}
+
+	for {
+		oldValue := s.maxDelay
+		if e.maxDelay > oldValue {
+			if s.maxDelay.CompareAndSwap(oldValue.Int64(), e.maxDelay.Int64()) {
+				e.maxDelay.Set(0)
+				break
+			}
+		} else {
+			break
+		}
+	}
 }
 
 var sessions struct {

include/pika_server.h

@@ -327,6 +327,7 @@ class PikaServer : public pstd::noncopyable {
   uint32_t SlowlogLen();
   void SlowlogObtain(int64_t number, std::vector<SlowlogEntry>* slowlogs);
   void SlowlogPushEntry(const PikaCmdArgsType& argv, int64_t time, int64_t duration);
+  uint64_t SlowlogCount();
 
   /*
    * Statistic used
@@ -680,6 +681,7 @@ class PikaServer : public pstd::noncopyable {
    * Slowlog used
    */
   uint64_t slowlog_entry_id_ = 0;
+  uint64_t slowlog_counter_ = 0;
   std::shared_mutex slowlog_protector_;
   std::list<SlowlogEntry> slowlog_list_;
 

src/pika_admin.cc

@@ -1056,7 +1056,7 @@ void InfoCmd::InfoStats(std::string& info) {
   tmp_stream << "is_slots_migrating:" << (is_migrating ? "Yes, " : "No, ") << start_migration_time_str << ", "
              << (is_migrating ? (current_time_s - start_migration_time) : (end_migration_time - start_migration_time))
              << "\r\n";
-
+  tmp_stream << "slow_logs_count:" << g_pika_server->SlowlogCount() << "\r\n";
   info.append(tmp_stream.str());
 }
 

src/pika_server.cc

@@ -1198,11 +1198,17 @@ void PikaServer::SlowlogPushEntry(const PikaCmdArgsType& argv, int64_t time, int
     entry.start_time = time;
     entry.duration = duration;
     slowlog_list_.push_front(entry);
+    slowlog_counter_++;
   }
 
   SlowlogTrim();
 }
 
+uint64_t PikaServer::SlowlogCount() {
+  std::shared_lock l(slowlog_protector_);
+  return slowlog_counter_;
+}
+
 void PikaServer::ResetStat() {
   statistic_.server_stat.accumulative_connections.store(0);
   statistic_.server_stat.qps.querynum.store(0);

tools/pika_exporter/discovery/codis_dashboard.go

@@ -49,6 +49,7 @@ type CmdInfo struct {
 	Calls         int64  `json:"calls"`
 	Usecs_percall int64  `json:"usecs_percall"`
 	Fails         int64  `json:"fails"`
+	MaxDelay      int64  `json:"max_delay"`
 }
 
 type ProxyOpsInfo struct {
@@ -97,9 +98,9 @@ type RunTimeInfo struct {
 }
 
 type ProxyStats struct {
-	Online           bool         `json:"online"`
-	Ops              ProxyOpsInfo `json:"ops"`
-	Rusage           RusageInfo   `json:"rusage"`
-	RunTime          RunTimeInfo  `json:"runtime"`
-	TimeoutCmdNumber int64        `json:"timeout_cmd_number"`
+	Online       bool         `json:"online"`
+	Ops          ProxyOpsInfo `json:"ops"`
+	Rusage       RusageInfo   `json:"rusage"`
+	RunTime      RunTimeInfo  `json:"runtime"`
+	SlowCmdCount int64        `json:"slow_cmd_count"`
 }

tools/pika_exporter/exporter/metrics/parser.go

@@ -304,6 +304,8 @@ func (p *proxyParser) Parse(m MetricMeta, c Collector, opt ParseOption) {
 				metric.Value = convertToFloat64(strconv.FormatInt(v[1], 10))
 			case "fails":
 				metric.Value = convertToFloat64(strconv.FormatInt(v[2], 10))
+			case "max_delay":
+				metric.Value = convertToFloat64(strconv.FormatInt(v[3], 10))
 			}
 
 			if err := c.Collect(metric); err != nil {

tools/pika_exporter/exporter/metrics/proxy.go

@@ -6,20 +6,20 @@ func RegisterForProxy() {
 }
 
 var collectProxyMetrics map[string]MetricConfig = map[string]MetricConfig{
-	"ops_total": {
+	"total_ops": {
 		Parser: &normalParser{},
 		MetricMeta: &MetaData{
-			Name:      "ops_total",
+			Name:      "total_ops",
 			Help:      "proxy total ops",
 			Type:      metricTypeCounter,
 			Labels:    []string{LabelNameAddr, LabelID, LabelProductName},
 			ValueName: "ops_total",
 		},
 	},
-	"ops_fails": {
+	"total_ops_fails": {
 		Parser: &normalParser{},
 		MetricMeta: &MetaData{
-			Name:      "ops_fails",
+			Name:      "total_ops_fails",
 			Help:      "proxy fails counter",
 			Type:      metricTypeCounter,
 			Labels:    []string{LabelNameAddr, LabelID, LabelProductName},
@@ -66,23 +66,23 @@ var collectProxyMetrics map[string]MetricConfig = map[string]MetricConfig{
 			ValueName: "online",
 		},
 	},
-	"timeout_cmd_number": {
+	"total_slow_cmd": {
 		Parser: &normalParser{},
 		MetricMeta: &MetaData{
-			Name:      "timeout_cmd_number",
-			Help:      "The number of commands recorded in the slow log within the last second",
-			Type:      metricTypeGauge,
+			Name:      "total_slow_cmd",
+			Help:      "The number of commands recorded in the slow log",
+			Type:      metricTypeCounter,
 			Labels:    []string{LabelNameAddr, LabelID, LabelProductName},
-			ValueName: "timeout_cmd_number",
+			ValueName: "slow_cmd_count",
 		},
 	},
 }
 
 var collectPorxyCmdMetrics map[string]MetricConfig = map[string]MetricConfig{
-	"calls": {
+	"total_calls": {
 		Parser: &proxyParser{},
 		MetricMeta: &MetaData{
-			Name:      "calls",
+			Name:      "total_calls",
 			Help:      "the number of cmd calls",
 			Type:      metricTypeCounter,
 			Labels:    []string{LabelNameAddr, LabelID, LabelProductName, LabelOpstr},
@@ -99,14 +99,24 @@ var collectPorxyCmdMetrics map[string]MetricConfig = map[string]MetricConfig{
 			ValueName: "usecs_percall",
 		},
 	},
-	"fails": {
+	"total_fails": {
 		Parser: &proxyParser{},
 		MetricMeta: &MetaData{
-			Name:      "fails",
+			Name:      "total_fails",
 			Help:      "the number of cmd fail",
 			Type:      metricTypeCounter,
 			Labels:    []string{LabelNameAddr, LabelID, LabelProductName, LabelOpstr},
 			ValueName: "fails",
 		},
 	},
+	"max_delay": {
+		Parser: &proxyParser{},
+		MetricMeta: &MetaData{
+			Name:      "max_delay",
+			Help:      "The maximum time consumed by this command since the last collection.",
+			Type:      metricTypeGauge,
+			Labels:    []string{LabelNameAddr, LabelID, LabelProductName, LabelOpstr},
+			ValueName: "max_delay",
+		},
+	},
 }

tools/pika_exporter/exporter/metrics/stats.go

@@ -158,4 +158,14 @@ var collectStatsMetrics = map[string]MetricConfig{
 			Labels: []string{LabelNameAddr, LabelNameAlias, "is_compact", "compact_cron", "compact_interval"},
 		},
 	},
+	"total_slow_log": {
+		Parser: &normalParser{},
+		MetricMeta: &MetaData{
+			Name:      "total_slow_log",
+			Help:      "pika serve instance total count of slow log",
+			Type:      metricTypeCounter,
+			Labels:    []string{LabelNameAddr, LabelNameAlias},
+			ValueName: "slow_logs_count",
+		},
+	},
 }