diff --git a/pipe.go b/pipe.go
index 97ae6d82..de852a91 100644
--- a/pipe.go
+++ b/pipe.go
@@ -286,6 +286,7 @@ func (p *pipe) _backgroundWrite() (err error) {
 		ones  = make([]cmds.Completed, 1)
 		multi []cmds.Completed
 		ch    chan RedisResult
+		delay = p.maxFlushDelay
 	)
 
 	for atomic.LoadInt32(&p.state) < 3 {
@@ -293,16 +294,12 @@ func (p *pipe) _backgroundWrite() (err error) {
 			if p.w.Buffered() == 0 {
 				err = p.Error()
 			} else {
-				if p.maxFlushDelay == 0 {
+				if delay == 0 || atomic.LoadInt32(&p.waits) == 1 { // do not delay for sequential usage
 					err = p.w.Flush()
 				} else {
-					if atomic.LoadInt32(&p.waits) == 1 {
-						err = p.w.Flush()
-					} else {
-						ts := time.Now()
-						err = p.w.Flush()
-						time.Sleep(p.maxFlushDelay - time.Since(ts))
-					}
+					ts := time.Now()
+					err = p.w.Flush()
+					time.Sleep(delay - time.Since(ts)) // ref: https://github.com/rueian/rueidis/issues/156
 				}
 			}
 			if err == nil {
diff --git a/rueidis.go b/rueidis.go
index 51b62385..8ac168eb 100644
--- a/rueidis.go
+++ b/rueidis.go
@@ -110,6 +110,15 @@ type ClientOption struct {
 	// This default is ClientOption.Dialer.KeepAlive * (9+1), where 9 is the default of tcp_keepalive_probes on Linux.
 	ConnWriteTimeout time.Duration
 
+	// MaxFlushDelay when greater than zero pauses pipeline write loop for some time (not larger than MaxFlushDelay)
+	// after each flushing of data to the connection. This gives pipeline a chance to collect more commands to send
+	// to Redis. Adding this delay increases latency, reduces throughput – but in most cases may significantly reduce
+	// application and Redis CPU utilization due to less executed system calls. By default, Rueidis flushes data to the
+	// connection without extra delays. Depending on network latency and application-specific conditions the value
+	// of MaxFlushDelay may vary, sth like 20 microseconds should not affect latency/throughput a lot but still
+	// produce notable CPU usage reduction under load. Ref: https://github.com/rueian/rueidis/issues/156
+	MaxFlushDelay time.Duration
+
 	// ShuffleInit is a handy flag that shuffles the InitAddress after passing to the NewClient() if it is true
 	ShuffleInit bool
 	// DisableRetry disables retrying read-only commands under network errors
@@ -118,14 +127,6 @@ type ClientOption struct {
 	DisableCache bool
 	// AlwaysPipelining makes rueidis.Client always pipeline redis commands even if they are not issued concurrently.
 	AlwaysPipelining bool
-	// MaxFlushDelay when greater than zero pauses pipeline write loop for some time (not larger than MaxFlushDelay)
-	// after each flushing of data to the connection. This gives pipeline a chance to collect more commands to send
-	// to Redis. Adding this delay increases latency, reduces throughput – but in most cases may significantly reduce
-	// application and Redis CPU utilization due to less executed system calls. By default, Rueidis flushes data to the
-	// connection without extra delays. Depending on network latency and application-specific conditions the value
-	// of MaxFlushDelay may vary, sth like 20 microseconds should not affect latency/throughput a lot but still
-	// produce notable CPU usage reduction under load.
-	MaxFlushDelay time.Duration
 }
 
 // SentinelOption contains MasterSet,