coordination: fix flaky TestCoordinationSemaphore without changing public API

CHANGELOG.md

@@ -1,3 +1,5 @@
+* Fixed a race in `coordination.Session` where `lastGoodResponseTime` could be zero when the keep-alive loop started, causing the keep-alive timer to fire immediately and triggering a spurious reconnect that would fail non-idempotent operations such as `CreateSemaphore`
+
 ## v3.135.10
 * Fixed the SDK's `database/sql` driver to consistently map session-invalidating YDB errors to `driver.ErrBadConn` where possible, so `database/sql` can detect and discard bad connections
 

internal/coordination/session.go

@@ -441,8 +441,12 @@ func (s *session) receiveLoop( //nolint:funlen
 
 			return
 		case *Ydb_Coordination.SessionResponse_SessionStarted_:
-			sessionStarted <- message.GetSessionStarted()
+			// Update the keep-alive timestamp BEFORE signaling mainLoop. If the update happened after
+			// the channel send, mainLoop could race into the keep-alive inner loop and see a zero
+			// lastGoodResponseTime, causing the keep-alive timer to fire immediately and triggering a
+			// spurious reconnect that would fail any non-idempotent operations in the queue.
 			s.updateLastGoodResponseTime()
+			sessionStarted <- message.GetSessionStarted()
 		case *Ydb_Coordination.SessionResponse_SessionStopped_:
 			sessionStopped <- message.GetSessionStopped()
 			s.cancel()

tests/integration/coordination_test.go

@@ -45,7 +45,7 @@ func TestCoordinationSemaphore(sourceTest *testing.T) {
 	err = db.Coordination().CreateNode(ctx, nodePath, coordination.NodeConfig{
 		Path:                     "",
 		SelfCheckPeriodMillis:    1000,
-		SessionGracePeriodMillis: 1000,
+		SessionGracePeriodMillis: 10000, //nolint:mnd
 		ReadConsistencyMode:      coordination.ConsistencyModeStrict,
 		AttachConsistencyMode:    coordination.ConsistencyModeStrict,
 		RatelimiterCountersMode:  coordination.RatelimiterCountersModeDetailed,
@@ -64,7 +64,9 @@ func TestCoordinationSemaphore(sourceTest *testing.T) {
 	}
 	fmt.Printf("node description: %+v\nnode config: %+v\n", e, c)
 
-	s, err := db.Coordination().Session(ctx, nodePath)
+	// Use a zero reconnect delay so the client re-establishes the gRPC stream immediately after
+	// Reconnect(), ensuring the session is restored well within the server's SessionGracePeriodMillis.
+	s, err := db.Coordination().Session(ctx, nodePath, options.WithSessionReconnectDelay(0))
 	if err != nil {
 		t.Fatalf("failed to create session: %v\n", err)
 	}