ttl: only gc in leader to save performance (#59358) (#59471)

close #59357

Author: ti-chi-bot

pkg/metrics/grafana/tidb.json

@@ -20236,7 +20236,7 @@
           "targets": [
             {
               "exemplar": true,
-              "expr": "avg(tidb_server_ttl_watermark_delay{k8s_cluster=\"$k8s_cluster\",tidb_cluster=\"$tidb_cluster\", type=\"schedule\"}) by (type, name)",
+              "expr": "max(tidb_server_ttl_watermark_delay{k8s_cluster=\"$k8s_cluster\",tidb_cluster=\"$tidb_cluster\", type=\"schedule\"}) by (type, name)",
               "interval": "",
               "legendFormat": "{{ name }}",
               "queryType": "randomWalk",

pkg/ttl/metrics/metrics.go

@@ -255,3 +255,8 @@ func UpdateDelayMetrics(records map[int64]*DelayMetricsRecord) {
 		metrics.TTLWatermarkDelay.With(prometheus.Labels{metrics.LblType: "schedule", metrics.LblName: delay}).Set(v)
 	}
 }
+
+// ClearDelayMetrics clears the metrics of TTL delay
+func ClearDelayMetrics() {
+	metrics.TTLWatermarkDelay.Reset()
+}

pkg/ttl/ttlworker/BUILD.bazel

@@ -101,6 +101,7 @@ go_test(
         "@com_github_ngaut_pools//:pools",
         "@com_github_pingcap_errors//:errors",
         "@com_github_pingcap_failpoint//:failpoint",
+        "@com_github_prometheus_client_golang//prometheus",
         "@com_github_prometheus_client_model//go",
         "@com_github_stretchr_testify//assert",
         "@com_github_stretchr_testify//mock",

pkg/ttl/ttlworker/job_manager.go

@@ -490,8 +490,15 @@ func (m *JobManager) reportMetrics(se session.Session) {
 	metrics.RunningJobsCnt.Set(runningJobs)
 	metrics.CancellingJobsCnt.Set(cancellingJobs)
 
+	if !m.isLeader() {
+		// only the leader can do collect delay metrics to reduce the performance overhead
+		metrics.ClearDelayMetrics()
+		return
+	}
+
 	if time.Since(m.lastReportDelayMetricsTime) > 10*time.Minute {
 		m.lastReportDelayMetricsTime = time.Now()
+		logutil.Logger(m.ctx).Info("TTL leader to collect delay metrics")
 		records, err := GetDelayMetricRecords(m.ctx, se, time.Now())
 		if err != nil {
 			logutil.Logger(m.ctx).Info("failed to get TTL delay metrics", zap.Error(err))
@@ -998,6 +1005,12 @@ func summarizeTaskResult(tasks []*cache.TTLTask) (*TTLSummary, error) {
 
 // DoGC deletes some old TTL job histories and redundant scan tasks
 func (m *JobManager) DoGC(ctx context.Context, se session.Session) {
+	if !m.isLeader() {
+		// only the leader can do the GC to reduce the performance impact
+		return
+	}
+
+	logutil.Logger(m.ctx).Info("TTL leader to DoGC")
 	// Remove the table not exist in info schema cache.
 	// Delete the table status before deleting the tasks. Therefore the related tasks
 	if err := m.updateInfoSchemaCache(se); err == nil {

pkg/ttl/ttlworker/job_manager_integration_test.go

@@ -30,6 +30,7 @@ import (
 	"github.com/pingcap/failpoint"
 	"github.com/pingcap/tidb/pkg/domain"
 	"github.com/pingcap/tidb/pkg/kv"
+	metrics2 "github.com/pingcap/tidb/pkg/metrics"
 	"github.com/pingcap/tidb/pkg/parser/ast"
 	"github.com/pingcap/tidb/pkg/parser/model"
 	dbsession "github.com/pingcap/tidb/pkg/session"
@@ -44,6 +45,7 @@ import (
 	"github.com/pingcap/tidb/pkg/ttl/session"
 	"github.com/pingcap/tidb/pkg/ttl/ttlworker"
 	"github.com/pingcap/tidb/pkg/util/logutil"
+	"github.com/prometheus/client_golang/prometheus"
 	dto "github.com/prometheus/client_model/go"
 	"github.com/stretchr/testify/require"
 	"go.uber.org/atomic"
@@ -807,10 +809,15 @@ func TestGCScanTasks(t *testing.T) {
 	addScanTaskRecord(3, 2, 1)
 	addScanTaskRecord(3, 2, 2)
 
+	isLeader := false
 	m := ttlworker.NewJobManager("manager-1", nil, store, nil, func() bool {
-		return true
+		return isLeader
 	})
 	se := session.NewSession(tk.Session(), tk.Session(), func(_ session.Session) {})
+	// only leader can do GC
+	m.DoGC(context.TODO(), se)
+	tk.MustQuery("select count(1) from mysql.tidb_ttl_task").Check(testkit.Rows("6"))
+	isLeader = true
 	m.DoGC(context.TODO(), se)
 	tk.MustQuery("select job_id, scan_id from mysql.tidb_ttl_task order by job_id, scan_id asc").Check(testkit.Rows("1 1", "1 2"))
 }
@@ -826,10 +833,15 @@ func TestGCTableStatus(t *testing.T) {
 	// insert table status without corresponding table
 	tk.MustExec("INSERT INTO mysql.tidb_ttl_table_status (table_id,parent_table_id) VALUES (?, ?)", 2024, 2024)
 
+	isLeader := false
 	m := ttlworker.NewJobManager("manager-1", nil, store, nil, func() bool {
-		return true
+		return isLeader
 	})
 	se := session.NewSession(tk.Session(), tk.Session(), func(_ session.Session) {})
+	// only leader can do GC
+	m.DoGC(context.TODO(), se)
+	tk.MustQuery("select count(1) from mysql.tidb_ttl_table_status").Check(testkit.Rows("1"))
+	isLeader = true
 	m.DoGC(context.TODO(), se)
 	tk.MustQuery("select * from mysql.tidb_ttl_table_status").Check(nil)
 
@@ -887,11 +899,16 @@ func TestGCTTLHistory(t *testing.T) {
 	addHistory(6, 91)
 	addHistory(7, 100)
 
+	isLeader := false
 	m := ttlworker.NewJobManager("manager-1", nil, store, nil, func() bool {
-		return true
+		return isLeader
 	})
 	se := session.NewSession(tk.Session(), tk.Session(), func(_ session.Session) {})
 	m.DoGC(context.TODO(), se)
+	// only leader can go GC
+	tk.MustQuery("select count(1) from mysql.tidb_ttl_job_history").Check(testkit.Rows("7"))
+	isLeader = true
+	m.DoGC(context.TODO(), se)
 	tk.MustQuery("select job_id from mysql.tidb_ttl_job_history order by job_id asc").Check(testkit.Rows("1", "2", "3", "4", "5"))
 }
 
@@ -1057,6 +1074,53 @@ func TestDelayMetrics(t *testing.T) {
 	checkRecord(records, "t3", now.Add(-3*time.Hour))
 	checkRecord(records, "t4", now.Add(-3*time.Hour))
 	checkRecord(records, "t5", emptyTime)
+
+	metrics.ClearDelayMetrics()
+	getMetricCnt := func() int {
+		ch := make(chan prometheus.Metric)
+		go func() {
+			metrics2.TTLWatermarkDelay.Collect(ch)
+			close(ch)
+		}()
+
+		cnt := 0
+		for range ch {
+			cnt++
+		}
+		return cnt
+	}
+
+	isLeader := false
+	m := ttlworker.NewJobManager("test-ttl-job-manager", nil, store, nil, func() bool {
+		return isLeader
+	})
+	// If the manager is not leader, the metrics will be empty.
+	m.ReportMetrics(se)
+	require.Zero(t, getMetricCnt())
+	// leader will collect metrics
+	isLeader = true
+	m.SetLastReportDelayMetricsTime(time.Now().Add(-11 * time.Minute))
+	m.ReportMetrics(se)
+	require.Equal(t, len(metrics.WaterMarkScheduleDelayNames), getMetricCnt())
+	require.InDelta(t, time.Now().Unix(), m.GetLastReportDelayMetricsTime().Unix(), 5)
+	// will not collect metrics in 10 minutes
+	lastReportTime := time.Now().Add(-9 * time.Minute)
+	m.SetLastReportDelayMetricsTime(lastReportTime)
+	m.ReportMetrics(se)
+	require.Equal(t, len(metrics.WaterMarkScheduleDelayNames), getMetricCnt())
+	require.Equal(t, lastReportTime.Unix(), m.GetLastReportDelayMetricsTime().Unix(), 5)
+	// when back to non-leader, the metrics will be empty and last report time will not be updated.
+	isLeader = false
+	lastReportTime = time.Now().Add(-11 * time.Minute)
+	m.SetLastReportDelayMetricsTime(lastReportTime)
+	m.ReportMetrics(se)
+	require.Zero(t, getMetricCnt())
+	require.Equal(t, lastReportTime.Unix(), m.GetLastReportDelayMetricsTime().Unix())
+	// when back to leader again, the metrics will be collected.
+	isLeader = true
+	m.ReportMetrics(se)
+	require.Equal(t, len(metrics.WaterMarkScheduleDelayNames), getMetricCnt())
+	require.InDelta(t, time.Now().Unix(), m.GetLastReportDelayMetricsTime().Unix(), 5)
 }
 
 func TestManagerJobAdapterCanSubmitJob(t *testing.T) {

pkg/ttl/ttlworker/job_manager_test.go

@@ -190,6 +190,16 @@ func (m *JobManager) UpdateHeartBeat(ctx context.Context, se session.Session, no
 	return m.updateHeartBeat(ctx, se, now)
 }
 
+// SetLastReportDelayMetricsTime sets the lastReportDelayMetricsTime for test
+func (m *JobManager) SetLastReportDelayMetricsTime(t time.Time) {
+	m.lastReportDelayMetricsTime = t
+}
+
+// GetLastReportDelayMetricsTime returns the lastReportDelayMetricsTime for test
+func (m *JobManager) GetLastReportDelayMetricsTime() time.Time {
+	return m.lastReportDelayMetricsTime
+}
+
 // ReportMetrics is an exported version of reportMetrics
 func (m *JobManager) ReportMetrics(se session.Session) {
 	m.reportMetrics(se)