add retry limit for excluded backoff type to avoid infinite retry (#1002)

* add retry limit for excluded backoff type to avoid infinite retry

Signed-off-by: cfzjywxk <lsswxrxr@163.com>

* change log

Signed-off-by: cfzjywxk <lsswxrxr@163.com>

---------

Signed-off-by: cfzjywxk <lsswxrxr@163.com>

internal/retry/backoff.go

@@ -143,7 +143,13 @@ func (b *Backoffer) BackoffWithCfgAndMaxSleep(cfg *Config, maxSleepMs int, err e
 	if b.noop {
 		return err
 	}
-	if b.maxSleep > 0 && (b.totalSleep-b.excludedSleep) >= b.maxSleep {
+	maxBackoffTimeExceeded := (b.totalSleep - b.excludedSleep) >= b.maxSleep
+	maxExcludedTimeExceeded := false
+	if maxLimit, ok := isSleepExcluded[cfg.name]; ok {
+		maxExcludedTimeExceeded = b.excludedSleep >= maxLimit && b.excludedSleep >= b.maxSleep
+	}
+	maxTimeExceeded := maxBackoffTimeExceeded || maxExcludedTimeExceeded
+	if b.maxSleep > 0 && maxTimeExceeded {
 		longestSleepCfg, longestSleepTime := b.longestSleepCfg()
 		errMsg := fmt.Sprintf("%s backoffer.maxSleep %dms is exceeded, errors:", cfg.String(), b.maxSleep)
 		for i, err := range b.errors {
@@ -163,7 +169,8 @@ func (b *Backoffer) BackoffWithCfgAndMaxSleep(cfg *Config, maxSleepMs int, err e
 			backoffDetail.WriteString(":")
 			backoffDetail.WriteString(strconv.Itoa(times))
 		}
-		errMsg += fmt.Sprintf("\ntotal-backoff-times: %v, backoff-detail: %v", totalTimes, backoffDetail.String())
+		errMsg += fmt.Sprintf("\ntotal-backoff-times: %v, backoff-detail: %v, maxBackoffTimeExceeded: %v, maxExcludedTimeExceeded: %v",
+			totalTimes, backoffDetail.String(), maxBackoffTimeExceeded, maxExcludedTimeExceeded)
 		returnedErr := err
 		if longestSleepCfg != nil {
 			errMsg += fmt.Sprintf("\nlongest sleep type: %s, time: %dms", longestSleepCfg.String(), longestSleepTime)

internal/retry/backoff_test.go

@@ -98,3 +98,15 @@ func TestBackoffDeepCopy(t *testing.T) {
 		assert.ErrorIs(t, err, BoMaxDataNotReady.err)
 	}
 }
+
+func TestBackoffWithMaxExcludedExceed(t *testing.T) {
+	setBackoffExcluded(BoTiKVServerBusy.name, 1)
+	b := NewBackofferWithVars(context.TODO(), 1, nil)
+	err := b.Backoff(BoTiKVServerBusy, errors.New("server is busy"))
+	assert.Nil(t, err)
+
+	// As the total excluded sleep is greater than the max limited value, error should be returned.
+	err = b.Backoff(BoTiKVServerBusy, errors.New("server is busy"))
+	assert.NotNil(t, err)
+	assert.Greater(t, b.excludedSleep, b.maxSleep)
+}

internal/retry/config.go

@@ -130,11 +130,18 @@ var (
 	BoTxnLockFast = NewConfig(txnLockFastName, &metrics.BackoffHistogramLockFast, NewBackoffFnCfg(2, 3000, EqualJitter), tikverr.ErrResolveLockTimeout)
 )
 
-var isSleepExcluded = map[string]struct{}{
-	BoTiKVServerBusy.name: {},
+var isSleepExcluded = map[string]int{
+	BoTiKVServerBusy.name: 600000, // The max excluded limit is 10min.
 	// add BoTiFlashServerBusy if appropriate
 }
 
+// setBackoffExcluded is used for test only.
+func setBackoffExcluded(name string, maxVal int) {
+	if _, ok := isSleepExcluded[name]; ok {
+		isSleepExcluded[name] = maxVal
+	}
+}
+
 const (
 	// NoJitter makes the backoff sequence strict exponential.
 	NoJitter = 1 + iota