kv: Record a new AdmittedWorkDoneEvent

This commit creates a new kvpb.AdmittedWorkDoneEvent which the kvadmission
controller uses to record total CPU time consumed by KV.

This is done via span.RecordStructured.

The cpu time is added to the execstats.QueryLevelStats object and reported in
explain analyze. Note that CPUTime has been renamed to SQLCPUTime to be more
accurate and better distinguish the various timings recorded on this object.

Fixes: https://cockroachlabs.atlassian.net/browse/CRDB-55927
Release note (sql change): KV cpu time is now included in traces and
EXPLAIN ANALYZE.

Author: alyshanjahani-crl

pkg/kv/kvpb/api.proto

@@ -3872,6 +3872,12 @@ message ContentionEvent {
   bool is_latch = 4;
 }
 
+// AdmittedWorkDoneEvent is a message that is recorded on traces by the kvadmission
+// controller after admitted KV work is executed.
+message AdmittedWorkDoneEvent {
+ uint64 cpu_time = 1 [(gogoproto.customname) = "CPUTime"];
+}
+
 // ScanStats is a message that tracks miscellaneous statistics of all Gets,
 // Scans, and ReverseScans in a single BatchResponse.
 message ScanStats {

pkg/kv/kvserver/kvadmission/BUILD.bazel

@@ -21,6 +21,7 @@ go_library(
         "//pkg/util/log",
         "//pkg/util/stop",
         "//pkg/util/timeutil",
+        "//pkg/util/tracing",
         "@com_github_cockroachdb_errors//:errors",
         "@com_github_cockroachdb_pebble//:pebble",
     ],

pkg/kv/kvserver/kvadmission/kvadmission.go

@@ -27,6 +27,7 @@ import (
 	"github.com/cockroachdb/cockroach/pkg/util/log"
 	"github.com/cockroachdb/cockroach/pkg/util/stop"
 	"github.com/cockroachdb/cockroach/pkg/util/timeutil"
+	"github.com/cockroachdb/cockroach/pkg/util/tracing"
 	"github.com/cockroachdb/errors"
 	"github.com/cockroachdb/pebble"
 )
@@ -170,7 +171,7 @@ type Controller interface {
 	) (Handle, error)
 	// AdmittedKVWorkDone is called after the admitted KV work is done
 	// executing.
-	AdmittedKVWorkDone(Handle, *StoreWriteBytes)
+	AdmittedKVWorkDone(context.Context, Handle, *StoreWriteBytes)
 	// AdmitRangefeedRequest must be called before serving rangefeed requests.
 	// If enabled, it returns a non-nil Pacer that's to be used within rangefeed
 	// catchup scans (typically CPU-intensive and affecting scheduling
@@ -246,7 +247,9 @@ type Handle struct {
 	raftAdmissionMeta    *kvflowcontrolpb.RaftAdmissionMeta
 
 	callAdmittedWorkDoneOnKVAdmissionQ bool
-	cpuStart                           time.Duration
+	// Used for measuring post-admission CPU, even for cases that were not
+	// subject to admission control.
+	cpuStart time.Duration
 }
 
 // AnnotateCtx annotates the given context with request-scoped admission
@@ -293,7 +296,7 @@ func (n *controllerImpl) AdmitKVWork(
 	ba *kvpb.BatchRequest,
 ) (_ Handle, retErr error) {
 	if n.kvAdmissionQ == nil {
-		return Handle{}, nil
+		return Handle{cpuStart: grunning.Time()}, nil
 	}
 	admissionInfo := workInfoForBatch(n.settings, requestTenantID, rangeTenantID, ba)
 	ah := Handle{tenantID: admissionInfo.TenantID}
@@ -414,31 +417,34 @@ func (n *controllerImpl) AdmitKVWork(
 			if err != nil {
 				return Handle{}, err
 			}
-			if callAdmittedWorkDoneOnKVAdmissionQ {
-				// We include the time to do other activities like intent resolution,
-				// since it is acceptable to charge them to the tenant.
-				ah.cpuStart = grunning.Time()
-			}
 			ah.callAdmittedWorkDoneOnKVAdmissionQ = callAdmittedWorkDoneOnKVAdmissionQ
 		}
 	}
+	// We include the time to do other activities like intent resolution,
+	// since it is acceptable to charge them to the tenant.
+	ah.cpuStart = grunning.Time()
 	return ah, nil
 }
 
 // AdmittedKVWorkDone implements the Controller interface.
-func (n *controllerImpl) AdmittedKVWorkDone(ah Handle, writeBytes *StoreWriteBytes) {
+func (n *controllerImpl) AdmittedKVWorkDone(
+	ctx context.Context, ah Handle, writeBytes *StoreWriteBytes,
+) {
 	n.elasticCPUGrantCoordinator.ElasticCPUWorkQueue.AdmittedWorkDone(ah.elasticCPUWorkHandle)
-	if ah.callAdmittedWorkDoneOnKVAdmissionQ {
-		cpuTime := grunning.Time() - ah.cpuStart
-		if cpuTime < 0 {
-			// We sometimes see cpuTime to be negative. We use 1ns here, arbitrarily.
-			// This issue is tracked by
-			// https://github.com/cockroachdb/cockroach/issues/126681.
-			if buildutil.CrdbTestBuild {
-				log.KvDistribution.Warningf(context.Background(), "grunning.Time() should be non-decreasing, cpuTime=%s", cpuTime)
-			}
-			cpuTime = 1
+	cpuTime := grunning.Time() - ah.cpuStart
+	if cpuTime < 0 {
+		// We sometimes see cpuTime to be negative. We use 1ns here, arbitrarily.
+		// This issue is tracked by
+		// https://github.com/cockroachdb/cockroach/issues/126681.
+		if buildutil.CrdbTestBuild {
+			log.KvDistribution.Warningf(context.Background(), "grunning.Time() should be non-decreasing, cpuTime=%s", cpuTime)
 		}
+		cpuTime = 1
+	}
+	if span := tracing.SpanFromContext(ctx); span != nil {
+		span.RecordStructured(&kvpb.AdmittedWorkDoneEvent{CPUTime: uint64(cpuTime.Nanoseconds())})
+	}
+	if ah.callAdmittedWorkDoneOnKVAdmissionQ {
 		n.kvAdmissionQ.AdmittedWorkDone(ah.tenantID, cpuTime)
 	}
 	if ah.storeAdmissionQ != nil {

pkg/server/node.go

@@ -1637,7 +1637,7 @@ func (n *Node) batchInternal(
 
 	var writeBytes *kvadmission.StoreWriteBytes
 	defer func() {
-		n.storeCfg.KVAdmissionController.AdmittedKVWorkDone(handle, writeBytes)
+		n.storeCfg.KVAdmissionController.AdmittedKVWorkDone(ctx, handle, writeBytes)
 		writeBytes.Release()
 	}()
 

pkg/sql/exec_log.go

@@ -404,7 +404,7 @@ func (p *planner) maybeLogStatementInternal(
 			KvTimeNanos:              queryLevelStats.KVTime.Nanoseconds(),
 			KvGrpcCalls:              queryLevelStats.KVBatchRequestsIssued,
 			NetworkMessages:          queryLevelStats.DistSQLNetworkMessages,
-			CpuTimeNanos:             queryLevelStats.CPUTime.Nanoseconds(),
+			CpuTimeNanos:             queryLevelStats.SQLCPUTime.Nanoseconds(),
 			IndexRecommendations:     indexRecs,
 			// TODO(mgartner): Use a slice of struct{uint64, uint64} instead of
 			// converting to strings.
@@ -506,7 +506,7 @@ func (p *planner) logTransaction(
 			ContentionTime:  int64(txnStats.ExecStats.ContentionTime.Seconds()),
 			NetworkMessages: txnStats.ExecStats.DistSQLNetworkMessages,
 			MaxDiskUsage:    txnStats.ExecStats.MaxDiskUsage,
-			CPUSQLNanos:     txnStats.ExecStats.CPUTime.Nanoseconds(),
+			CPUSQLNanos:     txnStats.ExecStats.SQLCPUTime.Nanoseconds(),
 			MVCCIteratorStats: eventpb.MVCCIteratorStats{
 				StepCount:                      txnStats.ExecStats.MvccSteps,
 				StepCountInternal:              txnStats.ExecStats.MvccStepsInternal,

pkg/sql/execstats/traceanalyzer.go

@@ -95,14 +95,20 @@ func NewFlowsMetadata(flows map[base.SQLInstanceID]*execinfrapb.FlowSpec) *Flows
 // given traces and flow metadata.
 // NOTE: When adding fields to this struct, be sure to update Accumulate.
 type QueryLevelStats struct {
-	DistSQLNetworkBytesSent            int64
-	MaxMemUsage                        int64
-	MaxDiskUsage                       int64
-	KVBytesRead                        int64
-	KVPairsRead                        int64
-	KVRowsRead                         int64
-	KVBatchRequestsIssued              int64
-	KVTime                             time.Duration
+	DistSQLNetworkBytesSent int64
+	MaxMemUsage             int64
+	MaxDiskUsage            int64
+	KVBytesRead             int64
+	KVPairsRead             int64
+	KVRowsRead              int64
+	KVBatchRequestsIssued   int64
+	// KVTime is the cumulative time spent waiting for a KV request. This includes disk IO time
+	// and potentially network time (if any of the keys are not local).
+	KVTime time.Duration
+	// KVCPUTime is the total amount of CPU time spent by nodes doing KV work.
+	KVCPUTime time.Duration
+	// SQLCPUTime is the total amount of CPU time spent by nodes doing SQL work.
+	SQLCPUTime                         time.Duration
 	MvccSteps                          int64
 	MvccStepsInternal                  int64
 	MvccSeeks                          int64
@@ -122,7 +128,6 @@ type QueryLevelStats struct {
 	LatchWaitTime                      time.Duration
 	ContentionEvents                   []kvpb.ContentionEvent
 	RUEstimate                         float64
-	CPUTime                            time.Duration
 	// SQLInstanceIDs is an ordered list of SQL instances that were involved in
 	// query processing.
 	SQLInstanceIDs []int32
@@ -168,6 +173,7 @@ func (s *QueryLevelStats) Accumulate(other QueryLevelStats) {
 	s.KVRowsRead += other.KVRowsRead
 	s.KVBatchRequestsIssued += other.KVBatchRequestsIssued
 	s.KVTime += other.KVTime
+	s.KVCPUTime += other.KVCPUTime
 	s.MvccSteps += other.MvccSteps
 	s.MvccStepsInternal += other.MvccStepsInternal
 	s.MvccSeeks += other.MvccSeeks
@@ -187,7 +193,7 @@ func (s *QueryLevelStats) Accumulate(other QueryLevelStats) {
 	s.LatchWaitTime += other.LatchWaitTime
 	s.ContentionEvents = append(s.ContentionEvents, other.ContentionEvents...)
 	s.RUEstimate += other.RUEstimate
-	s.CPUTime += other.CPUTime
+	s.SQLCPUTime += other.SQLCPUTime
 	s.SQLInstanceIDs = util.CombineUnique(s.SQLInstanceIDs, other.SQLInstanceIDs)
 	s.KVNodeIDs = util.CombineUnique(s.KVNodeIDs, other.KVNodeIDs)
 	s.Regions = util.CombineUnique(s.Regions, other.Regions)
@@ -287,7 +293,7 @@ func (a *TraceAnalyzer) ProcessStats() {
 		s.LockWaitTime += stats.KV.LockWaitTime.Value()
 		s.LatchWaitTime += stats.KV.LatchWaitTime.Value()
 		s.RUEstimate += float64(stats.Exec.ConsumedRU.Value())
-		s.CPUTime += stats.Exec.CPUTime.Value()
+		s.SQLCPUTime += stats.Exec.CPUTime.Value()
 	}
 
 	// Process streamStats.
@@ -397,6 +403,25 @@ func getAllContentionEvents(trace []tracingpb.RecordedSpan) []kvpb.ContentionEve
 	return contentionEvents
 }
 
+// getKVCPUTime returns the sum of CPUTimes from all AdmittedWorkDoneEvents
+// found in the given trace.
+func getKVCPUTime(trace []tracingpb.RecordedSpan) time.Duration {
+	var totalCPUTimeNanos int64
+	var ev kvpb.AdmittedWorkDoneEvent
+	for i := range trace {
+		trace[i].Structured(func(any *pbtypes.Any, _ time.Time) {
+			if !pbtypes.Is(any, &ev) {
+				return
+			}
+			if err := pbtypes.UnmarshalAny(any, &ev); err != nil {
+				return
+			}
+			totalCPUTimeNanos += int64(ev.CPUTime)
+		})
+	}
+	return time.Duration(totalCPUTimeNanos)
+}
+
 // GetQueryLevelStats returns all the top-level stats in a QueryLevelStats
 // struct. GetQueryLevelStats tries to process as many stats as possible. If
 // errors occur while processing stats, GetQueryLevelStats returns the combined
@@ -417,5 +442,6 @@ func GetQueryLevelStats(
 		queryLevelStats.Accumulate(analyzer.GetQueryLevelStats())
 	}
 	queryLevelStats.ContentionEvents = getAllContentionEvents(trace)
+	queryLevelStats.KVCPUTime = getKVCPUTime(trace)
 	return queryLevelStats, errs
 }

pkg/sql/execstats/traceanalyzer_test.go

@@ -247,7 +247,8 @@ func TestQueryLevelStatsAccumulate(t *testing.T) {
 		ContentionEvents:                   []kvpb.ContentionEvent{aEvent},
 		MaxDiskUsage:                       8,
 		RUEstimate:                         9,
-		CPUTime:                            10 * time.Second,
+		SQLCPUTime:                         10 * time.Second,
+		KVCPUTime:                          12 * time.Second,
 		MvccSteps:                          11,
 		MvccStepsInternal:                  12,
 		MvccSeeks:                          13,
@@ -283,7 +284,8 @@ func TestQueryLevelStatsAccumulate(t *testing.T) {
 		ContentionEvents:                   []kvpb.ContentionEvent{bEvent},
 		MaxDiskUsage:                       15,
 		RUEstimate:                         16,
-		CPUTime:                            17 * time.Second,
+		SQLCPUTime:                         17 * time.Second,
+		KVCPUTime:                          19 * time.Second,
 		MvccSteps:                          18,
 		MvccStepsInternal:                  19,
 		MvccSeeks:                          20,
@@ -318,7 +320,8 @@ func TestQueryLevelStatsAccumulate(t *testing.T) {
 		ContentionEvents:                   []kvpb.ContentionEvent{aEvent, bEvent},
 		MaxDiskUsage:                       15,
 		RUEstimate:                         25,
-		CPUTime:                            27 * time.Second,
+		SQLCPUTime:                         27 * time.Second,
+		KVCPUTime:                          31 * time.Second,
 		MvccSteps:                          29,
 		MvccStepsInternal:                  31,
 		MvccSeeks:                          33,

pkg/sql/instrumentation.go

@@ -905,7 +905,8 @@ func (ih *instrumentationHelper) emitExplainAnalyzePlanToOutputBuilder(
 			// is only collected for vectorized plans since it is gathered by the
 			// vectorizedStatsCollector operator.
 			// TODO(drewk): lift these restrictions.
-			ob.AddCPUTime(queryStats.CPUTime)
+			ob.AddCPUTime(queryStats.SQLCPUTime)
+			ob.AddKVCPUTime(queryStats.KVCPUTime)
 		}
 		if ih.isTenant && ih.vectorized {
 			// Only output RU estimate if this is a tenant. Additionally, RUs aren't

pkg/sql/opt/exec/explain/output.go

@@ -457,6 +457,12 @@ func (ob *OutputBuilder) AddCPUTime(cpuTime time.Duration) {
 	}
 }
 
+func (ob *OutputBuilder) AddKVCPUTime(cpuTime time.Duration) {
+	if !ob.flags.Deflake.HasAny(DeflakeVolatile) {
+		ob.AddTopLevelField("kv cpu time", string(humanizeutil.Duration(cpuTime)))
+	}
+}
+
 // AddRUEstimate adds a top-level field for the estimated number of RUs consumed
 // by the query.
 func (ob *OutputBuilder) AddRUEstimate(ru float64) {

pkg/sql/sqlstats/insights/util.go

@@ -22,8 +22,8 @@ func makeTxnInsight(value *sqlstats.RecordedTxnStats) *insightspb.Transaction {
 	}
 
 	var cpuSQLNanos int64
-	if value.ExecStats.CPUTime.Nanoseconds() >= 0 {
-		cpuSQLNanos = value.ExecStats.CPUTime.Nanoseconds()
+	if value.ExecStats.SQLCPUTime.Nanoseconds() >= 0 {
+		cpuSQLNanos = value.ExecStats.SQLCPUTime.Nanoseconds()
 	}
 
 	var errorCode string
@@ -71,7 +71,7 @@ func makeStmtInsight(value *sqlstats.RecordedStmtStats) *insightspb.Statement {
 	var cpuSQLNanos int64
 	if value.ExecStats != nil {
 		contention = &value.ExecStats.ContentionTime
-		cpuSQLNanos = value.ExecStats.CPUTime.Nanoseconds()
+		cpuSQLNanos = value.ExecStats.SQLCPUTime.Nanoseconds()
 	}
 
 	var errorCode string