Merge pull request #9977 from hashicorp/grpc-tuning

streaming: Grpc tuning

agent/cache-types/streaming_health_services.go

@@ -113,9 +113,13 @@ func newMaterializer(
 		Logger: deps.Logger,
 		Waiter: &retry.Waiter{
 			MinFailures: 1,
-			MinWait:     0,
+			// Start backing off with small increments (200-400ms) which will double
-			MaxWait:     60 * time.Second,
+			// each attempt. (200-400, 400-800, 800-1600, 1600-3200, 3200-6000, 6000
-			Jitter:      retry.NewJitter(100),
+			// after that). (retry.Wait applies Max limit after jitter right now).
+			Factor:  200 * time.Millisecond,
+			MinWait: 0,
+			MaxWait: 60 * time.Second,
+			Jitter:  retry.NewJitter(100),
 		},
 		Request: newRequestFn,
 	}), nil

agent/grpc/client.go

@@ -5,8 +5,10 @@ import (
 	"fmt"
 	"net"
 	"sync"
+	"time"
 
 	"google.golang.org/grpc"
+	"google.golang.org/grpc/keepalive"
 
 	"github.com/hashicorp/consul/agent/metadata"
 	"github.com/hashicorp/consul/agent/pool"
@@ -64,7 +66,22 @@ func (c *ClientConnPool) ClientConn(datacenter string) (*grpc.ClientConn, error)
 		grpc.WithDisableRetry(),
 		grpc.WithStatsHandler(newStatsHandler(defaultMetrics())),
 		// nolint:staticcheck // there is no other supported alternative to WithBalancerName
-		grpc.WithBalancerName("pick_first"))
+		grpc.WithBalancerName("pick_first"),
+		// Keep alive parameters are based on the same default ones we used for
+		// Yamux. These are somewhat arbitrary but we did observe in scale testing
+		// that the gRPC defaults (servers send keepalives only every 2 hours,
+		// clients never) seemed to result in TCP drops going undetected until
+		// actual updates needed to be sent which caused unnecessary delays for
+		// deliveries. These settings should be no more work for servers than
+		// existing yamux clients but hopefully allow TCP drops to be detected
+		// earlier and so have a smaller chance of going unnoticed until there are
+		// actual updates to send out from the servers. The servers have a policy to
+		// not accept pings any faster than once every 15 seconds to protect against
+		// abuse.
+		grpc.WithKeepaliveParams(keepalive.ClientParameters{
+			Time:    30 * time.Second,
+			Timeout: 10 * time.Second,
+		}))
 	if err != nil {
 		return nil, err
 	}

agent/grpc/handler.go

@@ -6,8 +6,10 @@ package grpc
 import (
 	"fmt"
 	"net"
+	"time"
 
 	"google.golang.org/grpc"
+	"google.golang.org/grpc/keepalive"
 )
 
 // NewHandler returns a gRPC server that accepts connections from Handle(conn).
@@ -20,6 +22,9 @@ func NewHandler(addr net.Addr, register func(server *grpc.Server)) *Handler {
 	srv := grpc.NewServer(
 		grpc.StatsHandler(newStatsHandler(metrics)),
 		grpc.StreamInterceptor((&activeStreamCounter{metrics: metrics}).Intercept),
+		grpc.KeepaliveEnforcementPolicy(keepalive.EnforcementPolicy{
+			MinTime: 15 * time.Second,
+		}),
 	)
 	register(srv)