client/controller: wait for tokens to reduce the debet

Signed-off-by: nolouch <[email protected]>

client/resource_group/controller/controller.go

@@ -46,6 +46,8 @@ const (
 	slowNotifyFilterDuration = 10 * time.Millisecond
 
 	watchRetryInterval = 30 * time.Second
+
+	bigRequestThreshold = 4 * 1024 * 1024 // 4MB -> 16 RRU
 )
 
 type selectType int
@@ -69,6 +71,9 @@ type ResourceGroupKVInterceptor interface {
 	OnRequestWait(ctx context.Context, resourceGroupName string, info RequestInfo) (*rmpb.Consumption, *rmpb.Consumption, time.Duration, uint32, error)
 	// OnResponse is used to consume tokens after receiving response.
 	OnResponse(resourceGroupName string, req RequestInfo, resp ResponseInfo) (*rmpb.Consumption, error)
+	// OnResponseWait is used to consume tokens after receiving a response. If the response requires many tokens, we need to wait for the tokens.
+	// This is an optimized version of OnResponse for cases where the response requires many tokens, making the debt smaller and smoother.
+	OnResponseWait(ctx context.Context, resourceGroupName string, req RequestInfo, resp ResponseInfo) (*rmpb.Consumption, time.Duration, error)
 	// IsBackgroundRequest If the resource group has background jobs, we should not record consumption and wait for it.
 	IsBackgroundRequest(ctx context.Context, resourceGroupName, requestResource string) bool
 }
@@ -642,7 +647,7 @@ func (c *ResourceGroupsController) OnRequestWait(
 	if err != nil {
 		return nil, nil, time.Duration(0), 0, err
 	}
-	return gc.onRequestWait(ctx, info)
+	return gc.onRequestWaitImpl(ctx, info)
 }
 
 // OnResponse is used to consume tokens after receiving response
@@ -654,7 +659,19 @@ func (c *ResourceGroupsController) OnResponse(
 		log.Warn("[resource group controller] resource group name does not exist", zap.String("name", resourceGroupName))
 		return &rmpb.Consumption{}, nil
 	}
-	return gc.onResponse(req, resp)
+	return gc.onResponseImpl(req, resp)
+}
+
+// OnResponseWait is used to consume tokens after receiving response
+func (c *ResourceGroupsController) OnResponseWait(
+	ctx context.Context, resourceGroupName string, req RequestInfo, resp ResponseInfo,
+) (*rmpb.Consumption, time.Duration, error) {
+	gc, ok := c.loadGroupController(resourceGroupName)
+	if !ok {
+		log.Warn("[resource group controller] resource group name does not exist", zap.String("name", resourceGroupName))
+		return &rmpb.Consumption{}, time.Duration(0), nil
+	}
+	return gc.onResponseWaitImpl(ctx, req, resp)
 }
 
 // IsBackgroundRequest If the resource group has background jobs, we should not record consumption and wait for it.
@@ -722,6 +739,8 @@ type groupCostController struct {
 
 	// fast path to make once token limit with un-limit burst.
 	burstable *atomic.Bool
+	// is throttled
+	isThrottled *atomic.Bool
 
 	lowRUNotifyChan       chan<- notifyMsg
 	tokenBucketUpdateChan chan<- *groupCostController
@@ -770,6 +789,8 @@ type groupMetricsCollection struct {
 	failedRequestCounterWithOthers    prometheus.Counter
 	failedRequestCounterWithThrottled prometheus.Counter
 	tokenRequestCounter               prometheus.Counter
+	runningKVRequestCounter           prometheus.Gauge
+	consumeTokenHistogram             prometheus.Observer
 }
 
 func initMetrics(oldName, name string) *groupMetricsCollection {
@@ -784,6 +805,8 @@ func initMetrics(oldName, name string) *groupMetricsCollection {
 		failedRequestCounterWithThrottled: failedRequestCounter.WithLabelValues(oldName, name, throttledType),
 		requestRetryCounter:               requestRetryCounter.WithLabelValues(oldName, name),
 		tokenRequestCounter:               resourceGroupTokenRequestCounter.WithLabelValues(oldName, name),
+		runningKVRequestCounter:           groupRunningKVRequestCounter.WithLabelValues(name),
+		consumeTokenHistogram:             tokenConsumedByReadHistogram.WithLabelValues(name),
 	}
 }
 
@@ -841,6 +864,7 @@ func newGroupCostController(
 		tokenBucketUpdateChan: tokenBucketUpdateChan,
 		lowRUNotifyChan:       lowRUNotifyChan,
 		burstable:             &atomic.Bool{},
+		isThrottled:           &atomic.Bool{},
 	}
 
 	switch gc.mode {
@@ -1179,6 +1203,7 @@ func (gc *groupCostController) modifyTokenCounter(counter *tokenCounter, bucket
 		if cfg.NewBurst < 0 {
 			cfg.NewTokens = float64(counter.getTokenBucketFunc().Settings.FillRate)
 		}
+		gc.isThrottled.Store(false)
 	} else {
 		// Otherwise the granted token is delivered to the client by fill rate.
 		cfg.NewTokens = 0
@@ -1199,6 +1224,7 @@ func (gc *groupCostController) modifyTokenCounter(counter *tokenCounter, bucket
 		counter.notify.setupNotificationThreshold = 1
 		counter.notify.mu.Unlock()
 		counter.lastDeadline = deadline
+		gc.isThrottled.Store(true)
 		select {
 		case gc.tokenBucketUpdateChan <- gc:
 		default:
@@ -1317,7 +1343,7 @@ func (gc *groupCostController) calcRequest(counter *tokenCounter) float64 {
 	return value
 }
 
-func (gc *groupCostController) onRequestWait(
+func (gc *groupCostController) onRequestWaitImpl(
 	ctx context.Context, info RequestInfo,
 ) (*rmpb.Consumption, *rmpb.Consumption, time.Duration, uint32, error) {
 	delta := &rmpb.Consumption{}
@@ -1332,11 +1358,13 @@ func (gc *groupCostController) onRequestWait(
 
 	if !gc.burstable.Load() {
 		var err error
-		now := time.Now()
 		var i int
 		var d time.Duration
+		gc.metrics.runningKVRequestCounter.Inc()
+		defer gc.metrics.runningKVRequestCounter.Dec()
 	retryLoop:
 		for i = 0; i < gc.mainCfg.WaitRetryTimes; i++ {
+			now := time.Now()
 			switch gc.mode {
 			case rmpb.GroupMode_RawMode:
 				res := make([]*Reservation, 0, len(requestResourceLimitTypeList))
@@ -1399,7 +1427,7 @@ func (gc *groupCostController) onRequestWait(
 	return delta, penalty, waitDuration, gc.getMeta().GetPriority(), nil
 }
 
-func (gc *groupCostController) onResponse(
+func (gc *groupCostController) onResponseImpl(
 	req RequestInfo, resp ResponseInfo,
 ) (*rmpb.Consumption, error) {
 	delta := &rmpb.Consumption{}
@@ -1440,6 +1468,93 @@ func (gc *groupCostController) onResponse(
 	return delta, nil
 }
 
+func (gc *groupCostController) onResponseWaitImpl(
+	ctx context.Context, req RequestInfo, resp ResponseInfo,
+) (*rmpb.Consumption, time.Duration, error) {
+	delta := &rmpb.Consumption{}
+	for _, calc := range gc.calculators {
+		calc.AfterKVRequest(delta, req, resp)
+	}
+	var (
+		waitDuration time.Duration
+		d            time.Duration
+		err          error
+	)
+	if !gc.burstable.Load() {
+		gc.metrics.runningKVRequestCounter.Inc()
+		defer gc.metrics.runningKVRequestCounter.Dec()
+	retryLoop:
+		for i := 0; i < gc.mainCfg.WaitRetryTimes; i++ {
+			now := time.Now()
+			switch gc.mode {
+			case rmpb.GroupMode_RawMode:
+				res := make([]*Reservation, 0, len(requestResourceLimitTypeList))
+				for typ, counter := range gc.run.resourceTokens {
+					if v := getRawResourceValueFromConsumption(delta, typ); v > 0 {
+						// allow debt for small request or not in throttled.
+						if delta.ReadBytes+delta.WriteBytes < bigRequestThreshold || !gc.isThrottled.Load() {
+							counter.limiter.RemoveTokens(time.Now(), v)
+							break retryLoop
+						}
+						res = append(res, counter.limiter.Reserve(ctx, gc.mainCfg.LTBMaxWaitDuration, now, v))
+					}
+				}
+				if d, err = WaitReservations(ctx, now, res); err == nil || errs.ErrClientResourceGroupThrottled.NotEqual(err) {
+					break retryLoop
+				}
+			case rmpb.GroupMode_RUMode:
+				res := make([]*Reservation, 0, len(requestUnitLimitTypeList))
+				for typ, counter := range gc.run.requestUnitTokens {
+					if v := getRUValueFromConsumption(delta, typ); v > 0 {
+						if gc.name != defaultResourceGroupName && enableControllerTraceLog.Load() {
+							gc.metrics.consumeTokenHistogram.Observe(v)
+						}
+						// allow debt for small request or not in throttled.
+						if delta.ReadBytes+delta.WriteBytes < bigRequestThreshold || !gc.isThrottled.Load() {
+							counter.limiter.RemoveTokens(time.Now(), v)
+							break retryLoop
+						}
+						res = append(res, counter.limiter.Reserve(ctx, gc.mainCfg.LTBMaxWaitDuration, now, v))
+					}
+				}
+				if d, err = WaitReservations(context.Background(), now, res); err == nil || errs.ErrClientResourceGroupThrottled.NotEqual(err) {
+					break retryLoop
+				}
+			}
+			gc.metrics.requestRetryCounter.Inc()
+			time.Sleep(gc.mainCfg.WaitRetryInterval)
+			waitDuration += gc.mainCfg.WaitRetryInterval
+		}
+		if err != nil {
+			if errs.ErrClientResourceGroupThrottled.Equal(err) {
+				gc.metrics.failedRequestCounterWithThrottled.Inc()
+				gc.metrics.failedLimitReserveDuration.Observe(d.Seconds())
+			} else {
+				gc.metrics.failedRequestCounterWithOthers.Inc()
+			}
+			return nil, waitDuration, err
+		}
+		gc.metrics.successfulRequestDuration.Observe(d.Seconds())
+		waitDuration += d
+	}
+
+	gc.mu.Lock()
+	// Record the consumption of the request
+	add(gc.mu.consumption, delta)
+	// Record the consumption of the request by store
+	count := &rmpb.Consumption{}
+	*count = *delta
+	// As the penalty is only counted when the request is completed, so here needs to calculate the write cost which is added in `BeforeKVRequest`
+	for _, calc := range gc.calculators {
+		calc.BeforeKVRequest(count, req)
+	}
+	add(gc.mu.storeCounter[req.StoreID()], count)
+	add(gc.mu.globalCounter, count)
+	gc.mu.Unlock()
+
+	return delta, waitDuration, nil
+}
+
 // GetActiveResourceGroup is used to get active resource group.
 // This is used for test only.
 func (c *ResourceGroupsController) GetActiveResourceGroup(resourceGroupName string) *rmpb.ResourceGroup {

client/resource_group/controller/controller_test.go

@@ -104,15 +104,15 @@ func TestRequestAndResponseConsumption(t *testing.T) {
 	kvCalculator := gc.getKVCalculator()
 	for idx, testCase := range testCases {
 		caseNum := fmt.Sprintf("case %d", idx)
-		consumption, _, _, priority, err := gc.onRequestWait(context.TODO(), testCase.req)
+		consumption, _, _, priority, err := gc.onRequestWaitImpl(context.TODO(), testCase.req)
 		re.NoError(err, caseNum)
 		re.Equal(priority, gc.meta.Priority)
 		expectedConsumption := &rmpb.Consumption{}
 		if testCase.req.IsWrite() {
 			kvCalculator.calculateWriteCost(expectedConsumption, testCase.req)
 			re.Equal(expectedConsumption.WRU, consumption.WRU)
 		}
-		consumption, err = gc.onResponse(testCase.req, testCase.resp)
+		consumption, err = gc.onResponseImpl(testCase.req, testCase.resp)
 		re.NoError(err, caseNum)
 		kvCalculator.calculateReadCost(expectedConsumption, testCase.resp)
 		kvCalculator.calculateCPUCost(expectedConsumption, testCase.resp)
@@ -121,6 +121,46 @@ func TestRequestAndResponseConsumption(t *testing.T) {
 	}
 }
 
+func TestOnResponseWaitConsumption(t *testing.T) {
+	re := require.New(t)
+	gc := createTestGroupCostController(re)
+
+	req := &TestRequestInfo{
+		isWrite: false,
+	}
+	resp := &TestResponseInfo{
+		readBytes: 2000 * 64 * 1024, // 2000RU
+		succeed:   true,
+	}
+
+	consumption, waitTIme, err := gc.onResponseWaitImpl(context.TODO(), req, resp)
+	re.NoError(err)
+	re.Zero(waitTIme)
+	verify := func() {
+		expectedConsumption := &rmpb.Consumption{}
+		kvCalculator := gc.getKVCalculator()
+		kvCalculator.calculateReadCost(expectedConsumption, resp)
+		re.Equal(expectedConsumption.RRU, consumption.RRU)
+	}
+	verify()
+
+	// modify the counter, then on response should has wait time.
+	counter := gc.run.requestUnitTokens[rmpb.RequestUnitType_RU]
+	gc.modifyTokenCounter(counter, &rmpb.TokenBucket{
+		Settings: &rmpb.TokenLimitSettings{
+			FillRate:   1000,
+			BurstLimit: 1000,
+		},
+	},
+		int64(5*time.Second/time.Millisecond),
+	)
+
+	consumption, waitTIme, err = gc.onResponseWaitImpl(context.TODO(), req, resp)
+	re.NoError(err)
+	re.NotZero(waitTIme)
+	verify()
+}
+
 func TestResourceGroupThrottledError(t *testing.T) {
 	re := require.New(t)
 	gc := createTestGroupCostController(re)
@@ -129,7 +169,7 @@ func TestResourceGroupThrottledError(t *testing.T) {
 		writeBytes: 10000000,
 	}
 	// The group is throttled
-	_, _, _, _, err := gc.onRequestWait(context.TODO(), req)
+	_, _, _, _, err := gc.onRequestWaitImpl(context.TODO(), req)
 	re.Error(err)
 	re.True(errs.ErrClientResourceGroupThrottled.Equal(err))
 }

client/resource_group/controller/limiter.go

@@ -335,7 +335,7 @@ func (lim *Limiter) Reconfigure(now time.Time,
 ) {
 	lim.mu.Lock()
 	defer lim.mu.Unlock()
-	logControllerTrace("[resource group controller] before reconfigure", zap.String("name", lim.name), zap.Float64("old-tokens", lim.tokens), zap.Float64("old-rate", float64(lim.limit)), zap.Float64("old-notify-threshold", args.NotifyThreshold), zap.Int64("old-burst", lim.burst))
+	logControllerTrace("[resource group controller] before reconfigure", zap.Float64("old-tokens", lim.tokens), zap.Float64("old-rate", float64(lim.limit)), zap.Float64("old-notify-threshold", lim.notifyThreshold), zap.Int64("old-burst", lim.burst))
 	if args.NewBurst < 0 {
 		lim.last = now
 		lim.tokens = args.NewTokens

client/resource_group/controller/metrics.go

@@ -63,6 +63,14 @@ var (
 			Help:      "Counter of failed request.",
 		}, []string{resourceGroupNameLabel, newResourceGroupNameLabel, errType})
 
+	groupRunningKVRequestCounter = prometheus.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: namespace,
+			Subsystem: requestSubsystem,
+			Name:      "running_kv_request",
+			Help:      "Counter of running kv request.",
+		}, []string{newResourceGroupNameLabel})
+
 	requestRetryCounter = prometheus.NewCounterVec(
 		prometheus.CounterOpts{
 			Namespace: namespace,
@@ -95,6 +103,14 @@ var (
 			Name:      "low_token_notified",
 			Help:      "Counter of low token request.",
 		}, []string{newResourceGroupNameLabel})
+	tokenConsumedByReadHistogram = prometheus.NewHistogramVec(
+		prometheus.HistogramOpts{
+			Namespace: namespace,
+			Subsystem: tokenRequestSubsystem,
+			Name:      "consume",
+			Buckets:   []float64{0, .5, 1, 2, 5, 10, 15, 20, 40, 64, 128, 256, 512, 1024, 2048}, // 0 ~ 2048
+			Help:      "Bucketed histogram of token consume.",
+		}, []string{newResourceGroupNameLabel})
 )
 
 var (
@@ -108,8 +124,10 @@ func init() {
 	prometheus.MustRegister(successfulRequestDuration)
 	prometheus.MustRegister(failedRequestCounter)
 	prometheus.MustRegister(failedLimitReserveDuration)
+	prometheus.MustRegister(groupRunningKVRequestCounter)
 	prometheus.MustRegister(requestRetryCounter)
 	prometheus.MustRegister(tokenRequestDuration)
 	prometheus.MustRegister(resourceGroupTokenRequestCounter)
 	prometheus.MustRegister(lowTokenRequestNotifyCounter)
+	prometheus.MustRegister(tokenConsumedByReadHistogram)
 }