mcp: add configurable keepalive failure threshold (#982)

mcp: add configurable keepalive failure threshold

Introduce `KeepAliveFailureThreshold` option in both `ClientOptions` and
`ServerOptions` to control how many consecutive keepalive ping failures
are tolerated before closing the session.

This aligns with the MCP spec's guidance that "multiple failed pings MAY
trigger a connection reset," allowing operators to tune resilience
against transient network hiccups without immediately tearing down
otherwise healthy sessions.

A threshold of 0 or 1 (the default) closes on the first failure,
preserving existing behavior. Higher values let isolated misses pass
while still closing the session once consecutive failures reach the
threshold. A successful ping resets the counter.

Tolerated failures are logged at WARN level; the final failure that
closes the session is logged at ERROR level.

This is rework of #979.

mcp/client.go

@@ -166,6 +166,13 @@ type ClientOptions struct {
 	// If the peer fails to respond to pings originating from the keepalive check,
 	// the session is automatically closed.
 	KeepAlive time.Duration
+	// KeepAliveFailureThreshold is the number of consecutive keepalive ping
+	// failures tolerated before the session is closed. A value of 0 or 1
+	// closes the session on the first failure (the default). Higher values
+	// align with the spec's "multiple failed pings MAY trigger a connection
+	// reset" guidance, letting a transient miss pass without tearing down an
+	// otherwise live session. Has no effect unless KeepAlive is non-zero.
+	KeepAliveFailureThreshold int
 }
 
 // toolContextKeyType is the context key type for passing tool definitions
@@ -441,7 +448,7 @@ func (cs *ClientSession) registerElicitationWaiter(elicitationID string) (await
 
 // startKeepalive starts the keepalive mechanism for this client session.
 func (cs *ClientSession) startKeepalive(interval time.Duration) {
-	startKeepalive(cs, interval, &cs.keepaliveCancel, cs.client.opts.Logger)
+	startKeepalive(cs, interval, cs.client.opts.KeepAliveFailureThreshold, &cs.keepaliveCancel, cs.client.opts.Logger)
 }
 
 // AddRoots adds the given roots to the client,

mcp/mcp_test.go

@@ -1920,6 +1920,82 @@ func TestKeepAliveFailure_Logged(t *testing.T) {
 	})
 }
 
+// scriptedKeepaliveSession is a keepaliveSession test double whose Ping
+// returns errors from a script (one entry consumed per call; the last entry
+// repeats once exhausted), and records how many times Close was called. Ping
+// returns immediately so the keepalive loop's pace is driven purely by the
+// ticker, making the test deterministic under synctest.
+type scriptedKeepaliveSession struct {
+	pingErrs   []error
+	pingCalls  atomic.Int64
+	closeCalls atomic.Int64
+}
+
+func (s *scriptedKeepaliveSession) Ping(context.Context, *PingParams) error {
+	n := int(s.pingCalls.Add(1)) - 1
+	if n >= len(s.pingErrs) {
+		n = len(s.pingErrs) - 1
+	}
+	return s.pingErrs[n]
+}
+
+func (s *scriptedKeepaliveSession) Close() error {
+	s.closeCalls.Add(1)
+	return nil
+}
+
+// TestStartKeepalive_FailureThreshold verifies that the session is kept alive
+// across consecutive ping failures below the threshold and only closed once the
+// threshold is reached.
+func TestStartKeepalive_FailureThreshold(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		const interval = 100 * time.Millisecond
+		sess := &scriptedKeepaliveSession{pingErrs: []error{errors.New("boom")}}
+		logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+		var cancel context.CancelFunc
+		startKeepalive(sess, interval, 3, &cancel, logger)
+		defer cancel()
+
+		// After two ticks → two failures, still below threshold 3: not closed.
+		time.Sleep(2*interval + interval/2)
+		synctest.Wait()
+		if got := sess.closeCalls.Load(); got != 0 {
+			t.Fatalf("session closed below threshold: closeCalls=%d (pingCalls=%d)", got, sess.pingCalls.Load())
+		}
+
+		// Third tick → third failure reaches threshold: session closed.
+		time.Sleep(interval)
+		synctest.Wait()
+		if got := sess.closeCalls.Load(); got != 1 {
+			t.Fatalf("expected one Close at threshold, got closeCalls=%d (pingCalls=%d)", got, sess.pingCalls.Load())
+		}
+	})
+}
+
+// TestStartKeepalive_SuccessResetsFailures verifies that a successful ping
+// resets the consecutive-failure counter, so an isolated failure between
+// successes never accumulates toward the threshold.
+func TestStartKeepalive_SuccessResetsFailures(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		const interval = 100 * time.Millisecond
+		// fail, success, fail, fail, then success (the tail repeats): the run
+		// never has 3 consecutive failures, so the session is never closed.
+		sess := &scriptedKeepaliveSession{pingErrs: []error{
+			errors.New("boom"), nil, errors.New("boom"), errors.New("boom"), nil,
+		}}
+		logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+		var cancel context.CancelFunc
+		startKeepalive(sess, interval, 3, &cancel, logger)
+		defer cancel()
+
+		time.Sleep(6 * interval)
+		synctest.Wait()
+		if got := sess.closeCalls.Load(); got != 0 {
+			t.Fatalf("session closed despite a success resetting the counter: closeCalls=%d (pingCalls=%d)", got, sess.pingCalls.Load())
+		}
+	})
+}
+
 func TestAddTool_DuplicateNoPanicAndNoDuplicate(t *testing.T) {
 	// Adding the same tool pointer twice should not panic and should not
 	// produce duplicates in the server's tool list.

mcp/server.go

@@ -79,6 +79,13 @@ type ServerOptions struct {
 	// If the peer fails to respond to pings originating from the keepalive check,
 	// the session is automatically closed.
 	KeepAlive time.Duration
+	// KeepAliveFailureThreshold is the number of consecutive keepalive ping
+	// failures tolerated before the session is closed. A value of 0 or 1
+	// closes the session on the first failure (the default). Higher values
+	// align with the spec's "multiple failed pings MAY trigger a connection
+	// reset" guidance, letting a transient miss pass without tearing down an
+	// otherwise live session. Has no effect unless KeepAlive is non-zero.
+	KeepAliveFailureThreshold int
 	// Function called when a client session subscribes to a resource.
 	SubscribeHandler func(context.Context, *SubscribeRequest) error
 	// Function called when a client session unsubscribes from a resource.
@@ -1605,7 +1612,7 @@ func (ss *ServerSession) Wait() error {
 
 // startKeepalive starts the keepalive mechanism for this server session.
 func (ss *ServerSession) startKeepalive(interval time.Duration) {
-	startKeepalive(ss, interval, &ss.keepaliveCancel, ss.server.opts.Logger)
+	startKeepalive(ss, interval, ss.server.opts.KeepAliveFailureThreshold, &ss.keepaliveCancel, ss.server.opts.Logger)
 }
 
 // pageToken is the internal structure for the opaque pagination cursor.

mcp/shared.go

@@ -751,9 +751,20 @@ type keepaliveSession interface {
 // It assigns the cancel function to the provided cancelPtr and starts a goroutine
 // that sends ping messages at the specified interval.
 //
-// logger must be non-nil; ping failures (which terminate the keepalive loop and
-// close the session) are reported via logger so they are not silently dropped.
-func startKeepalive(session keepaliveSession, interval time.Duration, cancelPtr *context.CancelFunc, logger *slog.Logger) {
+// failureThreshold is the number of consecutive ping failures tolerated before
+// the session is closed; a value below 1 is treated as 1 (close on the first
+// failure). A successful ping resets the counter. This mirrors the spec's
+// "multiple failed pings MAY trigger a connection reset" language, letting a
+// transient miss pass without tearing down an otherwise live session.
+//
+// logger must be non-nil; ping failures (both the tolerated ones and the final
+// one that closes the session) are reported via logger so they are not silently
+// dropped.
+func startKeepalive(session keepaliveSession, interval time.Duration, failureThreshold int, cancelPtr *context.CancelFunc, logger *slog.Logger) {
+	if failureThreshold < 1 {
+		failureThreshold = 1
+	}
+
 	ctx, cancel := context.WithCancel(context.Background())
 	// Assign cancel function before starting goroutine to avoid race condition.
 	// We cannot return it because the caller may need to cancel during the
@@ -764,6 +775,7 @@ func startKeepalive(session keepaliveSession, interval time.Duration, cancelPtr
 		ticker := time.NewTicker(interval)
 		defer ticker.Stop()
 
+		consecutiveFailures := 0
 		for {
 			select {
 			case <-ctx.Done():
@@ -772,17 +784,32 @@ func startKeepalive(session keepaliveSession, interval time.Duration, cancelPtr
 				pingCtx, pingCancel := context.WithTimeout(context.Background(), interval/2)
 				err := session.Ping(pingCtx, nil)
 				pingCancel()
-				if err != nil {
-					if errors.Is(err, jsonrpc2.ErrMethodNotFound) {
-						// Peer doesn't support ping, stop the keepalive process.
-						return
-					}
-					// Ping failed; log it before closing the session so the
-					// failure is observable to operators. See #218.
-					logger.Error("keepalive ping failed; closing session", "error", err)
-					_ = session.Close()
+				if err == nil {
+					consecutiveFailures = 0
+					continue
+				}
+				if errors.Is(err, jsonrpc2.ErrMethodNotFound) {
+					// Peer doesn't support ping, stop the keepalive process.
 					return
 				}
+				consecutiveFailures++
+				if consecutiveFailures < failureThreshold {
+					// Tolerate transient failures below the threshold; log so
+					// the misses are still observable to operators. See #218.
+					logger.Warn("keepalive ping failed; tolerating below threshold",
+						"error", err,
+						"consecutiveFailures", consecutiveFailures,
+						"failureThreshold", failureThreshold)
+					continue
+				}
+				// Threshold reached; log before closing the session so the
+				// failure is observable to operators. See #218.
+				logger.Error("keepalive ping failed; closing session",
+					"error", err,
+					"consecutiveFailures", consecutiveFailures,
+					"failureThreshold", failureThreshold)
+				_ = session.Close()
+				return
 			}
 		}
 	}()