statistics: ease the impact of stats feedback on cluster

config/config.go

@@ -617,7 +617,7 @@ var defaultConf = Config{
 		RunAutoAnalyze:       true,
 		StmtCountLimit:       5000,
 		FeedbackProbability:  0.05,
-		QueryFeedbackLimit:   1024,
+		QueryFeedbackLimit:   512,
 		PseudoEstimateRatio:  0.8,
 		ForcePriority:        "NO_PRIORITY",
 		BindInfoLease:        "3s",

config/config.toml.example

@@ -227,7 +227,7 @@ run-auto-analyze = true
 feedback-probability = 0.05
 
 # The max number of query feedback that cache in memory.
-query-feedback-limit = 1024
+query-feedback-limit = 512
 
 # Pseudo stats will be used if the ratio between the modify count and
 # row count in statistics of a table is greater than it.

executor/analyze_test.go

@@ -418,10 +418,13 @@ func (s *testSuite1) testAnalyzeIncremental(tk *testkit.TestKit, c *C) {
 	// Test analyze incremental with feedback.
 	tk.MustExec("insert into t values (3,3)")
 	oriProbability := statistics.FeedbackProbability.Load()
+	oriMinLogCount := handle.MinLogScanCount
 	defer func() {
 		statistics.FeedbackProbability.Store(oriProbability)
+		handle.MinLogScanCount = oriMinLogCount
 	}()
 	statistics.FeedbackProbability.Store(1)
+	handle.MinLogScanCount = 0
 	is := s.dom.InfoSchema()
 	table, err := is.TableByName(model.NewCIStr("test"), model.NewCIStr("t"))
 	c.Assert(err, IsNil)

statistics/feedback.go

@@ -78,6 +78,64 @@ func NewQueryFeedback(physicalID int64, hist *Histogram, expected int64, desc bo
 	}
 }
 
+// QueryFeedbackKey is the key for a group of feedbacks on the same index/column.
+type QueryFeedbackKey struct {
+	PhysicalID int64
+	HistID     int64
+	Tp         int
+}
+
+// QueryFeedbackMap is the collection of feedbacks.
+type QueryFeedbackMap struct {
+	Size      int
+	Feedbacks map[QueryFeedbackKey][]*QueryFeedback
+}
+
+// NewQueryFeedbackMap builds a feedback collection.
+func NewQueryFeedbackMap() *QueryFeedbackMap {
+	return &QueryFeedbackMap{Feedbacks: make(map[QueryFeedbackKey][]*QueryFeedback)}
+}
+
+// Append adds a feedback into map.
+func (m *QueryFeedbackMap) Append(q *QueryFeedback) {
+	k := QueryFeedbackKey{
+		PhysicalID: q.PhysicalID,
+		HistID:     q.Hist.ID,
+		Tp:         q.Tp,
+	}
+	m.append(k, []*QueryFeedback{q})
+	return
+}
+
+// MaxQueryFeedbackCount is the max number of feedbacks that are cached in memory.
+var MaxQueryFeedbackCount = atomic.NewInt64(1 << 9)
+
+func (m *QueryFeedbackMap) append(k QueryFeedbackKey, qs []*QueryFeedback) bool {
+	remained := MaxQueryFeedbackCount.Load() - int64(m.Size)
+	if remained <= 0 {
+		return false
+	}
+	s, ok := m.Feedbacks[k]
+	if !ok || s == nil {
+		s = make([]*QueryFeedback, 0, 8)
+	}
+	l := mathutil.MinInt64(int64(len(qs)), remained)
+	s = append(s, qs[:l]...)
+	m.Feedbacks[k] = s
+	m.Size = m.Size + int(l)
+	return true
+}
+
+// Merge combines 2 collections of feedbacks.
+func (m *QueryFeedbackMap) Merge(r *QueryFeedbackMap) {
+	for k, qs := range r.Feedbacks {
+		if !m.append(k, qs) {
+			break
+		}
+	}
+	return
+}
+
 var (
 	// MaxNumberOfRanges is the max number of ranges before split to collect feedback.
 	MaxNumberOfRanges = 20
@@ -201,7 +259,7 @@ func (q *QueryFeedback) Actual() int64 {
 // Update updates the query feedback. `startKey` is the start scan key of the partial result, used to find
 // the range for update. `counts` is the scan counts of each range, used to update the feedback count info.
 func (q *QueryFeedback) Update(startKey kv.Key, counts []int64) {
-	// Older version do not have the counts info.
+	// Older versions do not have the counts info.
 	if len(counts) == 0 {
 		q.Invalidate()
 		return
@@ -246,6 +304,43 @@ func (q *QueryFeedback) Update(startKey kv.Key, counts []int64) {
 	}
 }
 
+// NonOverlappedFeedbacks extracts a set of feedbacks which are not overlapped with each other.
+func NonOverlappedFeedbacks(sc *stmtctx.StatementContext, fbs []Feedback) ([]Feedback, bool) {
+	// Sort feedbacks by end point and start point incrementally, then pick every feedback that is not overlapped
+	// with the previous chosen feedbacks.
+	var existsErr bool
+	sort.Slice(fbs, func(i, j int) bool {
+		res, err := fbs[i].Upper.CompareDatum(sc, fbs[j].Upper)
+		if err != nil {
+			existsErr = true
+		}
+		if existsErr || res != 0 {
+			return res < 0
+		}
+		res, err = fbs[i].Lower.CompareDatum(sc, fbs[j].Lower)
+		if err != nil {
+			existsErr = true
+		}
+		return res < 0
+	})
+	if existsErr {
+		return fbs, false
+	}
+	resFBs := make([]Feedback, 0, len(fbs))
+	previousEnd := &types.Datum{}
+	for _, fb := range fbs {
+		res, err := previousEnd.CompareDatum(sc, fb.Lower)
+		if err != nil {
+			return fbs, false
+		}
+		if res <= 0 {
+			resFBs = append(resFBs, fb)
+			previousEnd = fb.Upper
+		}
+	}
+	return resFBs, true
+}
+
 // BucketFeedback stands for all the feedback for a bucket.
 type BucketFeedback struct {
 	feedback []Feedback   // All the feedback info in the same bucket.
@@ -478,39 +573,15 @@ func (b *BucketFeedback) mergeFullyContainedFeedback(sc *stmtctx.StatementContex
 	if len(feedbacks) == 0 {
 		return 0, 0, false
 	}
-	// Sort feedbacks by end point and start point incrementally, then pick every feedback that is not overlapped
-	// with the previous chosen feedbacks.
-	var existsErr bool
-	sort.Slice(feedbacks, func(i, j int) bool {
-		res, err := feedbacks[i].Upper.CompareDatum(sc, feedbacks[j].Upper)
-		if err != nil {
-			existsErr = true
-		}
-		if existsErr || res != 0 {
-			return res < 0
-		}
-		res, err = feedbacks[i].Lower.CompareDatum(sc, feedbacks[j].Lower)
-		if err != nil {
-			existsErr = true
-		}
-		return res < 0
-	})
-	if existsErr {
+	sortedFBs, ok := NonOverlappedFeedbacks(sc, feedbacks)
+	if !ok {
 		return 0, 0, false
 	}
-	previousEnd := &types.Datum{}
 	var sumFraction, sumCount float64
-	for _, fb := range feedbacks {
-		res, err := previousEnd.CompareDatum(sc, fb.Lower)
-		if err != nil {
-			return 0, 0, false
-		}
-		if res <= 0 {
-			fraction, _ := getOverlapFraction(fb, bkt)
-			sumFraction += fraction
-			sumCount += float64(fb.Count)
-			previousEnd = fb.Upper
-		}
+	for _, fb := range sortedFBs {
+		fraction, _ := getOverlapFraction(fb, bkt)
+		sumFraction += fraction
+		sumCount += float64(fb.Count)
 	}
 	return sumFraction, sumCount, true
 }

statistics/handle/handle.go

@@ -61,8 +61,8 @@ type Handle struct {
 		schemaVersion int64
 	}
 
-	// It can be read by multiply readers at the same time without acquire lock, but it can be
-	// written only after acquire the lock.
+	// It can be read by multiple readers at the same time without acquiring lock, but it can be
+	// written only after acquiring the lock.
 	statsCache struct {
 		sync.Mutex
 		atomic.Value
@@ -78,7 +78,7 @@ type Handle struct {
 	// globalMap contains all the delta map from collectors when we dump them to KV.
 	globalMap tableDeltaMap
 	// feedback is used to store query feedback info.
-	feedback []*statistics.QueryFeedback
+	feedback *statistics.QueryFeedbackMap
 
 	lease atomic2.Duration
 }
@@ -90,7 +90,7 @@ func (h *Handle) Clear() {
 	for len(h.ddlEventCh) > 0 {
 		<-h.ddlEventCh
 	}
-	h.feedback = h.feedback[:0]
+	h.feedback = statistics.NewQueryFeedbackMap()
 	h.mu.ctx.GetSessionVars().InitChunkSize = 1
 	h.mu.ctx.GetSessionVars().MaxChunkSize = 1
 	h.mu.ctx.GetSessionVars().EnableChunkRPC = false
@@ -101,16 +101,13 @@ func (h *Handle) Clear() {
 	h.mu.Unlock()
 }
 
-// MaxQueryFeedbackCount is the max number of feedback that cache in memory.
-var MaxQueryFeedbackCount = atomic2.NewInt64(1 << 10)
-
 // NewHandle creates a Handle for update stats.
 func NewHandle(ctx sessionctx.Context, lease time.Duration) *Handle {
 	handle := &Handle{
 		ddlEventCh: make(chan *util.Event, 100),
 		listHead:   &SessionStatsCollector{mapper: make(tableDeltaMap), rateMap: make(errorRateDeltaMap)},
 		globalMap:  make(tableDeltaMap),
-		feedback:   make([]*statistics.QueryFeedback, 0, MaxQueryFeedbackCount.Load()),
+		feedback:   statistics.NewQueryFeedbackMap(),
 	}
 	handle.lease.Store(lease)
 	// It is safe to use it concurrently because the exec won't touch the ctx.
@@ -133,10 +130,10 @@ func (h *Handle) SetLease(lease time.Duration) {
 	h.lease.Store(lease)
 }
 
-// GetQueryFeedback gets the query feedback. It is only use in test.
-func (h *Handle) GetQueryFeedback() []*statistics.QueryFeedback {
+// GetQueryFeedback gets the query feedback. It is only used in test.
+func (h *Handle) GetQueryFeedback() *statistics.QueryFeedbackMap {
 	defer func() {
-		h.feedback = h.feedback[:0]
+		h.feedback = statistics.NewQueryFeedbackMap()
 	}()
 	return h.feedback
 }