Merge 7c68d27694 into f1ffb55c9a

* Only check latencies once every 10 seconds with `routeByLatency`

`routeByLatency` currently checks latencies any time a server returns
a MOVED or READONLY reply. When a shard is down, the ClusterClient
chooses to issue the request to a random server, which returns a MOVED
reply. This causes a state refresh and a latency update on all servers.
This can lead to significant ping load to clusters with a large number
of clients.

This introduces logic to ping only once every 10 seconds, only
performing a latency update on a node during the `GC` function if the
latency was set later than 10 seconds ago.

Fixes https://github.com/redis/go-redis/issues/2782

* use UnixNano instead of Unix for better precision

---------

Co-authored-by: ofekshenawa <104765379+ofekshenawa@users.noreply.github.com>

README.md

@@ -70,7 +70,7 @@ go-redis supports 2 last Go versions and requires a Go version with
 module:
 
 ```shell
-go mod init github.com/my/repo
+go mod init example.com/greetings
 ```
 
 Then install go-redis/**v9**:

osscluster.go

@@ -21,6 +21,10 @@ import (
 	"github.com/redis/go-redis/v9/internal/rand"
 )
 
+const (
+	minLatencyMeasurementInterval = 10 * time.Second
+)
+
 var errClusterNoNodes = fmt.Errorf("redis: cluster has no nodes")
 
 // ClusterOptions are used to configure a cluster client and should be
@@ -316,6 +320,10 @@ type clusterNode struct {
 	latency    uint32 // atomic
 	generation uint32 // atomic
 	failing    uint32 // atomic
+
+	// last time the latency measurement was performed for the node, stored in nanoseconds
+	// from epoch
+	lastLatencyMeasurement int64 // atomic
 }
 
 func newClusterNode(clOpt *ClusterOptions, addr string) *clusterNode {
@@ -368,6 +376,7 @@ func (n *clusterNode) updateLatency() {
 		latency = float64(dur) / float64(successes)
 	}
 	atomic.StoreUint32(&n.latency, uint32(latency+0.5))
+	n.SetLastLatencyMeasurement(time.Now())
 }
 
 func (n *clusterNode) Latency() time.Duration {
@@ -397,6 +406,10 @@ func (n *clusterNode) Generation() uint32 {
 	return atomic.LoadUint32(&n.generation)
 }
 
+func (n *clusterNode) LastLatencyMeasurement() int64 {
+	return atomic.LoadInt64(&n.lastLatencyMeasurement)
+}
+
 func (n *clusterNode) SetGeneration(gen uint32) {
 	for {
 		v := atomic.LoadUint32(&n.generation)
@@ -406,6 +419,15 @@ func (n *clusterNode) SetGeneration(gen uint32) {
 	}
 }
 
+func (n *clusterNode) SetLastLatencyMeasurement(t time.Time) {
+	for {
+		v := atomic.LoadInt64(&n.lastLatencyMeasurement)
+		if t.UnixNano() < v || atomic.CompareAndSwapInt64(&n.lastLatencyMeasurement, v, t.UnixNano()) {
+			break
+		}
+	}
+}
+
 //------------------------------------------------------------------------------
 
 type clusterNodes struct {
@@ -493,10 +515,11 @@ func (c *clusterNodes) GC(generation uint32) {
 	c.mu.Lock()
 
 	c.activeAddrs = c.activeAddrs[:0]
+	now := time.Now()
 	for addr, node := range c.nodes {
 		if node.Generation() >= generation {
 			c.activeAddrs = append(c.activeAddrs, addr)
-			if c.opt.RouteByLatency {
+			if c.opt.RouteByLatency && node.LastLatencyMeasurement() < now.Add(-minLatencyMeasurementInterval).UnixNano() {
 				go node.updateLatency()
 			}
 			continue

redis.go

@@ -176,8 +176,6 @@ func (hs *hooksMixin) withProcessPipelineHook(
 }
 
 func (hs *hooksMixin) dialHook(ctx context.Context, network, addr string) (net.Conn, error) {
-	hs.hooksMu.Lock()
-	defer hs.hooksMu.Unlock()
 	return hs.current.dial(ctx, network, addr)
 }
 

redis_test.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"net"
+	"sync"
 	"testing"
 	"time"
 
@@ -633,3 +634,67 @@ var _ = Describe("Hook with MinIdleConns", func() {
 		}))
 	})
 })
+
+var _ = Describe("Dialer connection timeouts", func() {
+	var client *redis.Client
+
+	const dialSimulatedDelay = 1 * time.Second
+
+	BeforeEach(func() {
+		options := redisOptions()
+		options.Dialer = func(ctx context.Context, network, addr string) (net.Conn, error) {
+			// Simulated slow dialer.
+			// Note that the following sleep is deliberately not context-aware.
+			time.Sleep(dialSimulatedDelay)
+			return net.Dial("tcp", options.Addr)
+		}
+		options.MinIdleConns = 1
+		client = redis.NewClient(options)
+	})
+
+	AfterEach(func() {
+		err := client.Close()
+		Expect(err).NotTo(HaveOccurred())
+	})
+
+	It("does not contend on connection dial for concurrent commands", func() {
+		var wg sync.WaitGroup
+
+		const concurrency = 10
+
+		durations := make(chan time.Duration, concurrency)
+		errs := make(chan error, concurrency)
+
+		start := time.Now()
+		wg.Add(concurrency)
+
+		for i := 0; i < concurrency; i++ {
+			go func() {
+				defer wg.Done()
+
+				start := time.Now()
+				err := client.Ping(ctx).Err()
+				durations <- time.Since(start)
+				errs <- err
+			}()
+		}
+
+		wg.Wait()
+		close(durations)
+		close(errs)
+
+		// All commands should eventually succeed, after acquiring a connection.
+		for err := range errs {
+			Expect(err).NotTo(HaveOccurred())
+		}
+
+		// Each individual command should complete within the simulated dial duration bound.
+		for duration := range durations {
+			Expect(duration).To(BeNumerically("<", 2*dialSimulatedDelay))
+		}
+
+		// Due to concurrent execution, the entire test suite should also complete within
+		// the same dial duration bound applied for individual commands.
+		Expect(time.Since(start)).To(BeNumerically("<", 2*dialSimulatedDelay))
+	})
+})