WebSocket over HTTP/2: server streams.active gauge leaks and connections.active goes negative after the connection closes

[LivingLikeKrillin]

Expected Behavior

After a WebSocket-over-HTTP/2 connection closes, the server-side gauges should return to the values they held before it was opened, the same way they do for WebSocket over HTTP/1.1:

• reactor.netty.http.server.streams.active back to 0
• reactor.netty.http.server.connections.active unchanged (0 in a single-connection test)

Actual Behavior

After the HTTP/2 stream carrying the WebSocket closes:

• reactor.netty.http.server.streams.active stays at 1. It is incremented when the upgrade request is read (recordOpenStream) and never decremented, so it leaks by one per WebSocket-over-HTTP/2 connection.
• reactor.netty.http.server.connections.active is decremented to -1.

The HTTP/1.1 WebSocket path is not affected; its gauges return to 0.

Root cause (verified against main): when the upgrade completes, WebsocketServerOperations#initHandshaker replaces the HTTP metrics handler in place with the pass-through WebsocketHttpServerMetricsHandler added in #3229 for HTTP/1.1. Its channelInactive records recorder.recordServerConnectionInactive(...) regardless of the channel type. On an Http2StreamChannel the active unit is a stream — counted earlier through recordOpenStream — so the close should decrement streams.active via recordClosedStream. Recording a connection-inactive instead leaves streams.active incremented and decrements connections.active for a unit that was never counted there (connections.active is only incremented for a SocketChannel, never for an Http2StreamChannel).

The base handler already dispatches correctly — AbstractHttpServerMetricsHandler#recordInactiveConnectionOrStream routes an Http2StreamChannel to recordClosedStream — but the pass-through predates the HTTP/2 WebSocket path (#3691) and overrides channelInactive without going through it.

Steps to Reproduce

The following test fails on main; the HTTP/1.1 sibling (testServerConnectionsWebsocketMicrometer) passes, which rules out a measurement artifact. It uses .handle(...) (Extended CONNECT uses the CONNECT method, so routing by GET would 404) and waits for the stream to close on the server before reading the gauges (StreamCloseHandler is a small doOnConnection helper that latches when the server-side stream channel goes inactive).

@Test
void testServerConnectionsWebsocketMicrometerHttp2() throws Exception {
	disposableServer =
			createServer()
					.channelGroup(group)
					.host("127.0.0.1")
					.protocol(HttpProtocol.H2C)
					.http2Settings(spec -> spec.connectProtocolEnabled(true))
					.metrics(true, Function.identity())
					.handle((req, res) -> res.sendWebsocket((in, out) -> out.sendString(Mono.just("Hello World!"))))
					.doOnConnection(cnx -> StreamCloseHandler.INSTANCE.register(cnx.channel()))
					.bindNow();

	String address = formatSocketAddress(disposableServer.address());

	httpClient.protocol(HttpProtocol.H2C)
	          .websocket()
	          .uri("/ws")
	          .handle((in, out) -> in.receive().aggregate().asString())
	          .as(StepVerifier::create)
	          .expectNext("Hello World!")
	          .expectComplete()
	          .verify(Duration.ofSeconds(30));

	// wait until the stream is closed on the server side (pass-through channelInactive ran)
	assertThat(StreamCloseHandler.INSTANCE.awaitClientClosedOnServer()).as("awaitClientClosedOnServer timeout").isTrue();

	Gauge streamsActiveGauge = registry.find(SERVER_STREAMS_ACTIVE).tags(URI, HTTP, LOCAL_ADDRESS, address).gauge();
	Gauge connectionsActiveGauge = registry.find(SERVER_CONNECTIONS_ACTIVE).tags(URI, HTTP, LOCAL_ADDRESS, address).gauge();
	double streamsActive = streamsActiveGauge != null ? streamsActiveGauge.value() : 0d;
	double connectionsActive = connectionsActiveGauge != null ? connectionsActiveGauge.value() : 0d;

	// currently observed: streams.active = 1.0, connections.active = -1.0
	assertThat(streamsActive).isEqualTo(0d);
	assertThat(connectionsActive).isEqualTo(0d);
}

Possible Solution

Have WebsocketHttpServerMetricsHandler#channelInactive dispatch on the channel type the way recordInactiveConnectionOrStream already does — for an Http2StreamChannel, record a stream close (recordClosedStream) instead of a connection-inactive. The HTTP/1.1 path keeps its current behavior. I'm happy to open a PR with the fix and this test if you'd like.

Your Environment

• Reactor version(s) used: reactor-netty-http 1.3.6 (also reproduces on main / 1.3.7-SNAPSHOT)
• Other relevant libraries versions: Netty 4.2.15.Final
• JVM version (java -version): OpenJDK 25.0.2 (build 25.0.2+10-LTS)
• OS and version: Windows 11

[violetagg]

@LivingLikeKrillin Yes, please open a PR. Thanks!