Merge branch 'master' into name8

Makefile

@@ -315,9 +315,7 @@ CLEAN_UT_BINARY := find . -name '*.test.bin'| xargs rm -f
 
 clean-test:
 	# Cleaning test tmp...
-	rm -rf /tmp/test_pd*
-	rm -rf /tmp/pd-tests*
-	rm -rf /tmp/test_etcd*
+	rm -rf /tmp/pd_tests*
 	rm -f $(REAL_CLUSTER_TEST_PATH)/playground.log
 	go clean -testcache
 	@$(CLEAN_UT_BINARY)

client/go.mod

@@ -10,7 +10,7 @@ require (
 	github.com/opentracing/opentracing-go v1.2.0
 	github.com/pingcap/errors v0.11.5-0.20211224045212-9687c2b0f87c
 	github.com/pingcap/failpoint v0.0.0-20210918120811-547c13e3eb00
-	github.com/pingcap/kvproto v0.0.0-20240716095229-5f7ffec83ea7
+	github.com/pingcap/kvproto v0.0.0-20240910154453-b242104f8d31
 	github.com/pingcap/log v1.1.1-0.20221110025148-ca232912c9f3
 	github.com/prometheus/client_golang v1.18.0
 	github.com/stretchr/testify v1.8.2

client/go.sum

@@ -46,8 +46,8 @@ github.com/pingcap/errors v0.11.5-0.20211224045212-9687c2b0f87c h1:xpW9bvK+HuuTm
 github.com/pingcap/errors v0.11.5-0.20211224045212-9687c2b0f87c/go.mod h1:X2r9ueLEUZgtx2cIogM0v4Zj5uvvzhuuiu7Pn8HzMPg=
 github.com/pingcap/failpoint v0.0.0-20210918120811-547c13e3eb00 h1:C3N3itkduZXDZFh4N3vQ5HEtld3S+Y+StULhWVvumU0=
 github.com/pingcap/failpoint v0.0.0-20210918120811-547c13e3eb00/go.mod h1:4qGtCB0QK0wBzKtFEGDhxXnSnbQApw1gc9siScUl8ew=
-github.com/pingcap/kvproto v0.0.0-20240716095229-5f7ffec83ea7 h1:V9XS3FQ/P6u+kFaoSyY5DBswIA774BMpIOLDBMrpxKc=
-github.com/pingcap/kvproto v0.0.0-20240716095229-5f7ffec83ea7/go.mod h1:rXxWk2UnwfUhLXha1jxRWPADw9eMZGWEWCg92Tgmb/8=
+github.com/pingcap/kvproto v0.0.0-20240910154453-b242104f8d31 h1:6BY+3T6Hqpw9UZ/D7Om/xB+Xik3NkkYxBV6qCzUdUvU=
+github.com/pingcap/kvproto v0.0.0-20240910154453-b242104f8d31/go.mod h1:rXxWk2UnwfUhLXha1jxRWPADw9eMZGWEWCg92Tgmb/8=
 github.com/pingcap/log v1.1.1-0.20221110025148-ca232912c9f3 h1:HR/ylkkLmGdSSDaD8IDP+SZrdhV1Kibl9KrHxJ9eciw=
 github.com/pingcap/log v1.1.1-0.20221110025148-ca232912c9f3/go.mod h1:DWQW5jICDR7UJh4HtxXSM20Churx4CQL0fwL/SoOSA4=
 github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=

client/metrics.go

@@ -39,13 +39,14 @@ func initAndRegisterMetrics(constLabels prometheus.Labels) {
 }
 
 var (
-	cmdDuration         *prometheus.HistogramVec
-	cmdFailedDuration   *prometheus.HistogramVec
-	requestDuration     *prometheus.HistogramVec
-	tsoBestBatchSize    prometheus.Histogram
-	tsoBatchSize        prometheus.Histogram
-	tsoBatchSendLatency prometheus.Histogram
-	requestForwarded    *prometheus.GaugeVec
+	cmdDuration              *prometheus.HistogramVec
+	cmdFailedDuration        *prometheus.HistogramVec
+	requestDuration          *prometheus.HistogramVec
+	tsoBestBatchSize         prometheus.Histogram
+	tsoBatchSize             prometheus.Histogram
+	tsoBatchSendLatency      prometheus.Histogram
+	requestForwarded         *prometheus.GaugeVec
+	ongoingRequestCountGauge *prometheus.GaugeVec
 )
 
 func initMetrics(constLabels prometheus.Labels) {
@@ -117,6 +118,15 @@ func initMetrics(constLabels prometheus.Labels) {
 			Help:        "The status to indicate if the request is forwarded",
 			ConstLabels: constLabels,
 		}, []string{"host", "delegate"})
+
+	ongoingRequestCountGauge = prometheus.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace:   "pd_client",
+			Subsystem:   "request",
+			Name:        "ongoing_requests_count",
+			Help:        "Current count of ongoing batch tso requests",
+			ConstLabels: constLabels,
+		}, []string{"stream"})
 }
 
 var (

client/tso_batch_controller.go

@@ -19,53 +19,85 @@ import (
 	"runtime/trace"
 	"time"
 
-	"github.com/pingcap/errors"
-	"github.com/pingcap/log"
 	"github.com/tikv/pd/client/tsoutil"
-	"go.uber.org/zap"
 )
 
 type tsoBatchController struct {
 	maxBatchSize int
 	// bestBatchSize is a dynamic size that changed based on the current batch effect.
 	bestBatchSize int
 
-	tsoRequestCh          chan *tsoRequest
 	collectedRequests     []*tsoRequest
 	collectedRequestCount int
 
-	batchStartTime time.Time
+	// The time after getting the first request and the token, and before performing extra batching.
+	extraBatchingStartTime time.Time
 }
 
-func newTSOBatchController(tsoRequestCh chan *tsoRequest, maxBatchSize int) *tsoBatchController {
+func newTSOBatchController(maxBatchSize int) *tsoBatchController {
 	return &tsoBatchController{
 		maxBatchSize:          maxBatchSize,
 		bestBatchSize:         8, /* Starting from a low value is necessary because we need to make sure it will be converged to (current_batch_size - 4) */
-		tsoRequestCh:          tsoRequestCh,
 		collectedRequests:     make([]*tsoRequest, maxBatchSize+1),
 		collectedRequestCount: 0,
 	}
 }
 
 // fetchPendingRequests will start a new round of the batch collecting from the channel.
-// It returns true if everything goes well, otherwise false which means we should stop the service.
-func (tbc *tsoBatchController) fetchPendingRequests(ctx context.Context, maxBatchWaitInterval time.Duration) error {
-	var firstRequest *tsoRequest
-	select {
-	case <-ctx.Done():
-		return ctx.Err()
-	case firstRequest = <-tbc.tsoRequestCh:
-	}
-	// Start to batch when the first TSO request arrives.
-	tbc.batchStartTime = time.Now()
+// It returns nil error if everything goes well, otherwise a non-nil error which means we should stop the service.
+// It's guaranteed that if this function failed after collecting some requests, then these requests will be cancelled
+// when the function returns, so the caller don't need to clear them manually.
+func (tbc *tsoBatchController) fetchPendingRequests(ctx context.Context, tsoRequestCh <-chan *tsoRequest, tokenCh chan struct{}, maxBatchWaitInterval time.Duration) (errRet error) {
+	var tokenAcquired bool
+	defer func() {
+		if errRet != nil {
+			// Something went wrong when collecting a batch of requests. Release the token and cancel collected requests
+			// if any.
+			if tokenAcquired {
+				tokenCh <- struct{}{}
+			}
+			tbc.finishCollectedRequests(0, 0, 0, invalidStreamID, errRet)
+		}
+	}()
+
+	// Wait until BOTH the first request and the token have arrived.
+	// TODO: `tbc.collectedRequestCount` should never be non-empty here. Consider do assertion here.
 	tbc.collectedRequestCount = 0
-	tbc.pushRequest(firstRequest)
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case req := <-tsoRequestCh:
+			// Start to batch when the first TSO request arrives.
+			tbc.pushRequest(req)
+			// A request arrives but the token is not ready yet. Continue waiting, and also allowing collecting the next
+			// request if it arrives.
+			continue
+		case <-tokenCh:
+			tokenAcquired = true
+		}
+
+		// The token is ready. If the first request didn't arrive, wait for it.
+		if tbc.collectedRequestCount == 0 {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case firstRequest := <-tsoRequestCh:
+				tbc.pushRequest(firstRequest)
+			}
+		}
+
+		// Both token and the first request have arrived.
+		break
+	}
+
+	tbc.extraBatchingStartTime = time.Now()
 
 	// This loop is for trying best to collect more requests, so we use `tbc.maxBatchSize` here.
 fetchPendingRequestsLoop:
 	for tbc.collectedRequestCount < tbc.maxBatchSize {
 		select {
-		case tsoReq := <-tbc.tsoRequestCh:
+		case tsoReq := <-tsoRequestCh:
 			tbc.pushRequest(tsoReq)
 		case <-ctx.Done():
 			return ctx.Err()
@@ -88,7 +120,7 @@ fetchPendingRequestsLoop:
 		defer after.Stop()
 		for tbc.collectedRequestCount < tbc.bestBatchSize {
 			select {
-			case tsoReq := <-tbc.tsoRequestCh:
+			case tsoReq := <-tsoRequestCh:
 				tbc.pushRequest(tsoReq)
 			case <-ctx.Done():
 				return ctx.Err()
@@ -103,7 +135,7 @@ fetchPendingRequestsLoop:
 	// we can adjust the `tbc.bestBatchSize` dynamically later.
 	for tbc.collectedRequestCount < tbc.maxBatchSize {
 		select {
-		case tsoReq := <-tbc.tsoRequestCh:
+		case tsoReq := <-tsoRequestCh:
 			tbc.pushRequest(tsoReq)
 		case <-ctx.Done():
 			return ctx.Err()
@@ -136,31 +168,16 @@ func (tbc *tsoBatchController) adjustBestBatchSize() {
 	}
 }
 
-func (tbc *tsoBatchController) finishCollectedRequests(physical, firstLogical int64, suffixBits uint32, err error) {
+func (tbc *tsoBatchController) finishCollectedRequests(physical, firstLogical int64, suffixBits uint32, streamID string, err error) {
 	for i := 0; i < tbc.collectedRequestCount; i++ {
 		tsoReq := tbc.collectedRequests[i]
 		// Retrieve the request context before the request is done to trace without race.
 		requestCtx := tsoReq.requestCtx
 		tsoReq.physical, tsoReq.logical = physical, tsoutil.AddLogical(firstLogical, int64(i), suffixBits)
+		tsoReq.streamID = streamID
 		tsoReq.tryDone(err)
 		trace.StartRegion(requestCtx, "pdclient.tsoReqDequeue").End()
 	}
 	// Prevent the finished requests from being processed again.
 	tbc.collectedRequestCount = 0
 }
-
-func (tbc *tsoBatchController) revokePendingRequests(err error) {
-	for i := 0; i < len(tbc.tsoRequestCh); i++ {
-		req := <-tbc.tsoRequestCh
-		req.tryDone(err)
-	}
-}
-
-func (tbc *tsoBatchController) clear() {
-	log.Info("[pd] clear the tso batch controller",
-		zap.Int("max-batch-size", tbc.maxBatchSize), zap.Int("best-batch-size", tbc.bestBatchSize),
-		zap.Int("collected-request-count", tbc.collectedRequestCount), zap.Int("pending-request-count", len(tbc.tsoRequestCh)))
-	tsoErr := errors.WithStack(errClosing)
-	tbc.finishCollectedRequests(0, 0, 0, tsoErr)
-	tbc.revokePendingRequests(tsoErr)
-}

client/tso_client.go

@@ -203,6 +203,7 @@ func (c *tsoClient) getTSORequest(ctx context.Context, dcLocation string) *tsoRe
 	req.physical = 0
 	req.logical = 0
 	req.dcLocation = dcLocation
+	req.streamID = ""
 	return req
 }