*: add more metrics and some temporary adjustment (#276)

Author: IANTHEREAL

cmd/drainer/main.go

@@ -24,11 +24,10 @@ func main() {
 		log.Fatalf("verifying flags error, See 'drainer --help'. %s", errors.ErrorStack(err))
 	}
 
+	drainer.InitLogger(cfg)
 	version.PrintVersionInfo()
 	log.Infof("use config: %+v", cfg)
 
-	drainer.InitLogger(cfg)
-
 	bs, err := drainer.NewServer(cfg)
 	if err != nil {
 		log.Fatalf("create drainer server error, %s", errors.ErrorStack(err))

drainer/collector.go

@@ -167,6 +167,11 @@ func (c *Collector) updateCollectStatus(synced bool) {
 // updateStatus queries pumps' status , deletes the offline pump
 // and updates pumps' latest ts
 func (c *Collector) updateStatus(ctx context.Context) error {
+	begin := time.Now()
+	defer func() {
+		publishBinlogHistogram.WithLabelValues("drainer").Observe(time.Since(begin).Seconds())
+	}()
+
 	if err := c.updatePumpStatus(ctx); err != nil {
 		log.Errorf("DetectPumps error: %v", errors.ErrorStack(err))
 		c.updateCollectStatus(false)
@@ -303,6 +308,7 @@ func (c *Collector) LoadHistoryDDLJobs() ([]*model.Job, error) {
 
 // publishBinlogs collects binlogs whose commitTS are in (minTS, maxTS], then publish them in ascending commitTS order
 func (c *Collector) publishBinlogs(ctx context.Context, minTS, maxTS int64) {
+	begin := time.Now()
 	// multiple ways sort:
 	// 1. get multiple way sorted binlogs
 	// 2. use heap to merge sort
@@ -316,11 +322,13 @@ func (c *Collector) publishBinlogs(ctx context.Context, minTS, maxTS int64) {
 			bss[id] = bs
 			binlogOffsets[id] = 1
 			// first push the first item into heap every pump
-			c.bh.push(ctx, bs[0])
+			c.bh.push(ctx, bs[0], false)
 		}
 		total += bs.Len()
 	}
+	publishBinlogHistogram.WithLabelValues("drainer_collector").Observe(time.Since(begin).Seconds())
 
+	begin = time.Now()
 	item := c.bh.pop()
 	for item != nil {
 		c.syncer.Add(item)
@@ -329,11 +337,12 @@ func (c *Collector) publishBinlogs(ctx context.Context, minTS, maxTS int64) {
 			delete(bss, item.nodeID)
 		} else {
 			// push next item into heap and increase the offset
-			c.bh.push(ctx, bss[item.nodeID][binlogOffsets[item.nodeID]])
+			c.bh.push(ctx, bss[item.nodeID][binlogOffsets[item.nodeID]], false)
 			binlogOffsets[item.nodeID] = binlogOffsets[item.nodeID] + 1
 		}
 		item = c.bh.pop()
 	}
+	publishBinlogHistogram.WithLabelValues("drainer_merge_sort").Observe(time.Since(begin).Seconds())
 
 	publishBinlogCounter.WithLabelValues("drainer").Add(float64(total))
 }

drainer/heap.go

@@ -67,14 +67,14 @@ func newBinlogHeap(size int) *binlogHeap {
 	}
 }
 
-func (b *binlogHeap) push(ctx context.Context, item *binlogItem) {
+func (b *binlogHeap) push(ctx context.Context, item *binlogItem, check bool) {
 	for {
 		select {
 		case <-ctx.Done():
 			return
 		default:
 			b.Lock()
-			if b.bh.Len() == b.size {
+			if check && b.bh.Len() == b.size {
 				b.Unlock()
 				time.Sleep(pushRetryTime)
 				continue
@@ -97,18 +97,3 @@ func (b *binlogHeap) pop() *binlogItem {
 	b.Unlock()
 	return item.(*binlogItem)
 }
-
-func (b *binlogHeap) peek() *binlogItem {
-	b.Lock()
-	if b.bh.Len() == 0 {
-		b.Unlock()
-		return nil
-	}
-
-	item := heap.Pop(b.bh)
-	heap.Push(b.bh, item)
-	b.Unlock()
-
-	return item.(*binlogItem)
-
-}

drainer/heap_test.go

@@ -46,7 +46,7 @@ func (s *testDrainerSuite) TestHeap(c *C) {
 		wg.Add(1)
 		defer wg.Done()
 		for _, cs := range testCase {
-			bh.push(ctx, newBinlogItem(cs, pb.Pos{}, "testnode"))
+			bh.push(ctx, newBinlogItem(cs, pb.Pos{}, "testnode"), true)
 		}
 	}()
 
@@ -64,12 +64,12 @@ func (s *testDrainerSuite) TestHeap(c *C) {
 
 	//test push block and cancel push operator
 	bh = newBinlogHeap(1)
-	bh.push(ctx, newBinlogItem(testCase[0], pb.Pos{}, "testnode"))
+	bh.push(ctx, newBinlogItem(testCase[0], pb.Pos{}, "testnode"), true)
 	go func() {
 		wg.Add(1)
 		defer wg.Done()
 		for _, cs := range testCase {
-			bh.push(ctx, newBinlogItem(cs, pb.Pos{}, "testnode"))
+			bh.push(ctx, newBinlogItem(cs, pb.Pos{}, "testnode"), true)
 		}
 	}()
 

drainer/metrics.go

@@ -26,6 +26,24 @@ var (
 			Help:      "offset for each pump.",
 		}, []string{"nodeID"})
 
+	findMatchedBinlogHistogram = prometheus.NewHistogramVec(
+		prometheus.HistogramOpts{
+			Namespace: "binlog",
+			Subsystem: "drainer",
+			Name:      "find_matched_binlog_duration_time",
+			Help:      "Bucketed histogram of find a matched binlog.",
+			Buckets:   prometheus.ExponentialBuckets(0.00005, 2, 18),
+		}, []string{"nodeID"})
+
+	publishBinlogHistogram = prometheus.NewHistogramVec(
+		prometheus.HistogramOpts{
+			Namespace: "binlog",
+			Subsystem: "drainer",
+			Name:      "publish_binlog_duration_time",
+			Help:      "Bucketed histogram of publish a binlog.",
+			Buckets:   prometheus.ExponentialBuckets(0.00005, 2, 18),
+		}, []string{"nodeID"})
+
 	publishBinlogCounter = prometheus.NewCounterVec(
 		prometheus.CounterOpts{
 			Namespace: "binlog",
@@ -114,6 +132,8 @@ func init() {
 	prometheus.MustRegister(txnHistogram)
 	prometheus.MustRegister(readBinlogHistogram)
 	prometheus.MustRegister(readBinlogSizeHistogram)
+	prometheus.MustRegister(publishBinlogHistogram)
+	prometheus.MustRegister(findMatchedBinlogHistogram)
 }
 
 type metricClient struct {

drainer/pump.go

@@ -1,6 +1,7 @@
 package drainer
 
 import (
+	"fmt"
 	"io"
 	"strconv"
 	"sync"
@@ -21,6 +22,9 @@ import (
 	pb "github.com/pingcap/tipb/go-binlog"
 )
 
+// sleep 10 millisecond to wait matched binlog
+var waitMatchedTime = 10 * time.Millisecond
+
 type binlogEntity struct {
 	tp       pb.BinlogType
 	startTS  int64
@@ -157,10 +161,12 @@ func (p *Pump) publish(t *tikv.LockResolver) {
 		case entity = <-p.binlogChan:
 		}
 
+		begin := time.Now()
 		switch entity.tp {
 		case pb.BinlogType_Prewrite:
 			// while we meet the prebinlog we must find it's mathced commit binlog
 			p.mustFindCommitBinlog(t, entity.startTS)
+			findMatchedBinlogHistogram.WithLabelValues(p.nodeID).Observe(time.Since(begin).Seconds())
 		case pb.BinlogType_Commit, pb.BinlogType_Rollback:
 			// if the commitTs is larger than maxCommitTs,
 			// we would publish all binlogs:
@@ -175,6 +181,7 @@ func (p *Pump) publish(t *tikv.LockResolver) {
 				} else {
 					binlogs = make(map[int64]*binlogItem)
 				}
+				publishBinlogHistogram.WithLabelValues(p.nodeID).Observe(time.Since(begin).Seconds())
 			}
 		}
 
@@ -195,7 +202,7 @@ func (p *Pump) mustFindCommitBinlog(t *tikv.LockResolver, startTS int64) {
 
 		b, ok := p.getPrewriteBinlogEntity(startTS)
 		if ok {
-			time.Sleep(waitTime)
+			time.Sleep(waitMatchedTime)
 			// check again after sleep a moment
 			b, ok = p.getPrewriteBinlogEntity(startTS)
 			if ok {
@@ -296,7 +303,7 @@ func (p *Pump) putIntoHeap(items map[int64]*binlogItem) {
 			// if we meet a smaller binlog, we should ignore it. because we have published binlogs that before window low boundary
 			continue
 		}
-		p.bh.push(p.ctx, item)
+		p.bh.push(p.ctx, item, true)
 	}
 
 	errorBinlogCount.Add(float64(errorBinlogs))
@@ -352,8 +359,9 @@ func (p *Pump) getDDLJob(id int64) (*model.Job, error) {
 }
 
 func (p *Pump) collectBinlogs(windowLower, windowUpper int64) binlogItems {
+	begin := time.Now()
 	var bs binlogItems
-	item := p.bh.peek()
+	item := p.bh.pop()
 	for item != nil && item.binlog.CommitTs <= windowUpper {
 		// make sure to discard old binlogs whose commitTS is earlier or equal minTS
 		if item.binlog.CommitTs > windowLower {
@@ -363,10 +371,14 @@ func (p *Pump) collectBinlogs(windowLower, windowUpper int64) binlogItems {
 		if ComparePos(p.currentPos, item.pos) == -1 {
 			p.currentPos = item.pos
 		}
-		_ = p.bh.pop()
-		item = p.bh.peek()
+		item = p.bh.pop()
+	}
+	if item != nil {
+		p.bh.push(p.ctx, item, false)
 	}
 
+	publishBinlogHistogram.WithLabelValues(fmt.Sprintf("%s_collect_binlogs", p.nodeID)).Observe(time.Since(begin).Seconds())
+
 	return bs
 }
 

drainer/util.go

@@ -116,7 +116,10 @@ func genDrainerID(listenAddr string) (string, error) {
 }
 
 func execute(executor executor.Executor, sqls []string, args [][]interface{}, commitTSs []int64, isDDL bool) error {
-	// compute txn duration
+	if len(sqls) == 0 {
+		return nil
+	}
+
 	beginTime := time.Now()
 	defer func() {
 		txnHistogram.Observe(time.Since(beginTime).Seconds())

pkg/offsets/offset.go

@@ -163,6 +163,8 @@ func (ks *KafkaSeeker) getAndCompare(topic string, partition int32, offset int64
 	for msg := range pc.Messages() {
 		bp, err := ks.operator.Decode(msg)
 		if err != nil {
+			//log.Errorf("decode message(offset %d) error %v", message.Offset, err)
+			//return 1, -2, nil
 			return 0, bp, errors.Annotatef(err, "decode %s", msg)
 		}
 

pkg/sql/sql.go

@@ -41,7 +41,7 @@ func ExecuteSQLs(db *sql.DB, sqls []string, args [][]interface{}, isDDL bool) er
 			time.Sleep(RetryWaitTime)
 		}
 
-		err = appleTxn(db, sqls, args)
+		err = ExecuteTxn(db, sqls, args)
 		if err == nil {
 			return nil
 		}
@@ -50,7 +50,8 @@ func ExecuteSQLs(db *sql.DB, sqls []string, args [][]interface{}, isDDL bool) er
 	return errors.Trace(err)
 }
 
-func appleTxn(db *sql.DB, sqls []string, args [][]interface{}) error {
+// ExecuteTxn executes transaction
+func ExecuteTxn(db *sql.DB, sqls []string, args [][]interface{}) error {
 	txn, err := db.Begin()
 	if err != nil {
 		log.Errorf("exec sqls[%v] begin failed %v", sqls, errors.ErrorStack(err))