the session wont be removed while client disconnect the mcp-server in WebMvcSseServerTransportProvider

[Forest-Fairy]

in WebMvcSseServerTransportProvider, sseBuilder.onComplete and sseBuilder.onTimeout wont be trigger while the client disconnect. I use cherry studio as the client. and it causes the size of sessions map to increase endlessly.
I think we can add an interval to send ping to check the connection state

private ServerResponse handleSseConnection(ServerRequest request) {
		if (this.isClosing) {
			return ServerResponse.status(HttpStatus.SERVICE_UNAVAILABLE).body("Server is shutting down");
		}

		String sessionId = UUID.randomUUID().toString();
		logger.debug("Creating new SSE connection for session: {}", sessionId);

		// Send initial endpoint event
		try {
			return ServerResponse.sse(sseBuilder -> {
				sseBuilder.onComplete(() -> {
					logger.debug("SSE connection completed for session: {}", sessionId);
					sessions.remove(sessionId);
				});
				sseBuilder.onTimeout(() -> {
					logger.debug("SSE connection timed out for session: {}", sessionId);
					sessions.remove(sessionId);
				});

				WebMvcMcpSessionTransport sessionTransport = new WebMvcMcpSessionTransport(sessionId, sseBuilder);
				McpServerSession session = sessionFactory.create(sessionTransport);
				this.sessions.put(sessionId, session);

				try {
					sseBuilder.id(sessionId)
						.event(ENDPOINT_EVENT_TYPE)
						.data(this.baseUrl + this.messageEndpoint + "?sessionId=" + sessionId);
				}
				catch (Exception e) {
					logger.error("Failed to send initial endpoint event: {}", e.getMessage());
					sseBuilder.error(e);
				}
			}, Duration.ZERO);
		}
		catch (Exception e) {
			logger.error("Failed to send initial endpoint event to session {}: {}", sessionId, e.getMessage());
			sessions.remove(sessionId);
			return ServerResponse.status(HttpStatus.INTERNAL_SERVER_ERROR).build();
		}
	}

[sunanzhi]

I'm having this problem at the moment, and I've found that the session stays in the ConcurrentHashMap for a long time, which may OOM

[sunanzhi]

If you are using the webmvc pattern, you can try to use a timer to periodically notify the client to send a ping to process, code example:

public CustomMcpSseServerTransportProvider(ObjectMapper objectMapper, String baseUrl, String messageEndpoint, String sseEndpoint) {

// some init code

Executors.newScheduledThreadPool(1).scheduleWithFixedDelay(() -> {
try {
notifyClients(McpSchema.METHOD_PING, null).block();
}catch (Exception ignored) {}
}, 0, 1, TimeUnit.MINUTES);

}

[denniskawurek]

Looks like the callbacks aren't working due to some changes which have been done between bb6514e808ddfd2a746151ac955d2375db42a3f1 and the latest commit.

[Forest-Fairy]

If you are using the webmvc pattern, you can try to use a timer to periodically notify the client to send a ping to process, code example:

public CustomMcpSseServerTransportProvider(ObjectMapper objectMapper, String baseUrl, String messageEndpoint, String sseEndpoint) {

// some init code

Executors.newScheduledThreadPool(1).scheduleWithFixedDelay(() -> { try { notifyClients(McpSchema.METHOD_PING, null).block(); }catch (Exception ignored) {} }, 0, 1, TimeUnit.MINUTES);

}

i use a similar way to fix this, i define a method to get the last request timestamp in the mcp-session interface, and check every 1 mins, then close the session if it has not received any request in last 30 mins.

[Forest-Fairy]

Looks like the callbacks aren't working due to some changes which have been done between bb6514e808ddfd2a746151ac955d2375db42a3f1 and the latest commit.

I check the git history of HttpServletSseServerTransportProvider, and find nothing may have effects on it after bb6514e , but now I think this is normal for the http connection which is stateless that the server can not receive the callbackw after client close the connection

[yu199195]

@sunanzhi. can fix this ?

notifyClients(McpSchema.METHOD_PING, null).block();

[krrr]

The laziest solution is to change the second parameter of ServerResponse.sse to a non-zero value, such as one day.

[mofee]

If you are using the webmvc pattern, you can try to use a timer to periodically notify the client to send a ping to process, code example:

public CustomMcpSseServerTransportProvider(ObjectMapper objectMapper, String baseUrl, String messageEndpoint, String sseEndpoint) {

// some init code

Executors.newScheduledThreadPool(1).scheduleWithFixedDelay(() -> { try { notifyClients(McpSchema.METHOD_PING, null).block(); }catch (Exception ignored) {} }, 0, 1, TimeUnit.MINUTES);

}

this can fix the bug, but trigger an IOException log, that means each disconnected session will log an exception.

[SmellyTrotter]

#128

[chemicL]

Since this issue was created, the implementation went over a significant change and refactor to incorporate Streamable HTTP. Are you still experiencing problems described here?