mcp: fix goroutine leaks in unit tests #496

fgrosse · 2025-09-18T18:28:11Z

This PR fixes goroutine leaks in all unit tests.

To find the leaks I integrated https://github.com/uber-go/goleak and I suggest to keep using it to catch any future regressions.

I used the folowing script to more easily find individual tests which were leaking goroutines:

for t in $(go test -run=Nothing -list=. ./mcp | grep -v ok); do go test ./mcp -run="$t" > /dev/null && echo -n . || echo -e "\nDetected leak: $t"; done

Fixes modelcontextprotocol#489

…flag

fgrosse · 2025-09-19T09:21:00Z

mcp/cmd_test.go

-	// get a signal when the server process exits
-	onExit := make(chan struct{})
-	go func() {
-		cmd.Process.Wait()
-		close(onExit)
-	}()


This is just simplified to cmd.Process.Wait() further down below

But the previous logic was serving a purpose: the test would wait at most 5s.
Generally speaking, it's bad form to depend on timing in tests, but pragmatically it can be useful, and in any case we probably should separate this simplification into a separate CL, if we really care about it.

I'm fine with removing this if you prefer the more complex setup.

Just sharing my motivation: In my experience it's typically acceptable to let the runtime / CI enforce a timeout in these situations (even if the default timeout of 10m can feel long). For the happy path (no test failures) it makes no difference but you get slightly less complexity in your test. When the test starts to hang, I would notice during local development, where I would expect developers to run it in isolation and make use of the -timeout flag of go test

Ok, this is ifne.

mcp/mcp_test.go

mcp/server.go

fgrosse · 2025-09-19T09:24:58Z

mcp/server.go

+	if err := ss.mcpConn.Close(); err != nil {
+		connErr = fmt.Errorf("failed to close mcp connection: %w", err)
+	}


I checked the code, and it seemed like an oversight to me that the mcpConn was never closed, but please double-check if I missed something here.

I think this should just be ss.Close().

The mcpConn is closed automatically when the jsonrpc2 conn is closed (this happens in the overly complex connect function).

Fixed via a43fb6f

fgrosse · 2025-09-19T09:25:43Z

mcp/streamable_test.go

 	realServer := httptest.NewServer(NewStreamableHTTPHandler(func(*http.Request) *Server { return server }, nil))
-	defer realServer.Close()
+	t.Cleanup(func() {
+		t.Log("Closing real HTTP server")


Given this function currently does not pass the leak check because of #499 , I decided to leave these logs in here which I added during debugging.

Thanks, makes sense. How about also adding a comment referencing #499. Something like:

// Until we have a way to clean up abandoned sessions, this test will leak goroutines (see #499)

Done via f41256c

fgrosse · 2025-09-24T21:17:37Z

@findleyr let me know what you think when you have time :)

findleyr

Thanks for your patience, and for these improvements. I got busy and had to put down this review temporarily.

findleyr · 2025-09-29T17:29:05Z

mcp/server.go

+	if err := ss.mcpConn.Close(); err != nil {
+		connErr = fmt.Errorf("failed to close mcp connection: %w", err)
+	}


I think this should just be ss.Close().

The mcpConn is closed automatically when the jsonrpc2 conn is closed (this happens in the overly complex connect function).

findleyr · 2025-09-29T19:12:59Z

mcp/transport.go

+	rcErr := r.rc.Close()
+
+	var wcErr error
+	if r.wc != nil {


Let's add a comment that we only allow a nil writer for tests.

Done via 415ff56

findleyr · 2025-09-29T19:13:41Z

mcp/streamable_test.go

 	realServer := httptest.NewServer(NewStreamableHTTPHandler(func(*http.Request) *Server { return server }, nil))
-	defer realServer.Close()
+	t.Cleanup(func() {
+		t.Log("Closing real HTTP server")


Thanks, makes sense. How about also adding a comment referencing #499. Something like:

// Until we have a way to clean up abandoned sessions, this test will leak goroutines (see #499)

mcp/server.go

findleyr · 2025-09-29T19:18:25Z

mcp/cmd_test.go

-	// get a signal when the server process exits
-	onExit := make(chan struct{})
-	go func() {
-		cmd.Process.Wait()
-		close(onExit)
-	}()


But the previous logic was serving a purpose: the test would wait at most 5s.
Generally speaking, it's bad form to depend on timing in tests, but pragmatically it can be useful, and in any case we probably should separate this simplification into a separate CL, if we really care about it.

fgrosse · 2025-10-06T14:52:06Z

I have addressed all comments and merged the most recent main into this branch. Good to go?

findleyr

Thanks, just superficial comments at this point. Really appreciate your time and diligence in tracking this down.

My largest comment was that I don't want to add a dependency on go-leak, and would prefer to do this analysis in an ad-hoc manner. (I actually thought I'd already left that feedback, but alas my review was still pending--sorry).

findleyr · 2025-10-07T14:16:44Z

mcp/client_example_test.go

 	// Connect the server and client...
 	t1, t2 := mcp.NewInMemoryTransports()
-	if _, err := s.Connect(ctx, t1, nil); err != nil {
+	sess1, err := s.Connect(ctx, t1, nil)


s/sess1/serverSession (or ss)
s/sess2/clientSession (or cs)

sess1 and sess2 obscures the fact that these variables have different types.

findleyr · 2025-10-07T14:16:55Z

mcp/cmd_test.go

-	// get a signal when the server process exits
-	onExit := make(chan struct{})
-	go func() {
-		cmd.Process.Wait()
-		close(onExit)
-	}()


Ok, this is ifne.

findleyr · 2025-10-07T14:18:20Z

mcp/streamable_example_test.go

 	handler := mcp.NewStreamableHTTPHandler(func(r *http.Request) *mcp.Server {
 		return server
-	}, &mcp.StreamableHTTPOptions{JSONResponse: true})
+	}, &mcp.StreamableHTTPOptions{JSONResponse: true, Stateless: true})


Why make this Stateless?

This example is demonstrating how to use the API, so using 'Stateless' here may be distracting. Would prefer to leave a comment that this test may leak goroutines, as you've done below.

findleyr · 2025-10-07T14:19:38Z

go.mod

 	github.com/google/go-cmp v0.7.0
 	github.com/google/jsonschema-go v0.3.0
 	github.com/yosida95/uritemplate/v3 v3.0.2
+	go.uber.org/goleak v1.3.0


Hmm, I don't think we should add an additional dependency just for this purpose. It seems like handling these as a one-off, every once in a while, is sufficient for now.

findleyr · 2025-10-11T14:26:27Z

Hi Friedrich, we really appreciate this contribution and would like to land it. Let me know if you'd like me to take it over (mea culpa for the review latency--it has been a busy few weeks!).

fgrosse added 3 commits September 18, 2025 17:50

mcp: use github.com/uber-go/goleak to test for leaks in unit tests

0af54b4

mcp: fix goroutine leaks in unit tests

c2458f9

Fixes modelcontextprotocol#489

mcp: fix more flaky tests

c9ef5b8

fgrosse mentioned this pull request Sep 18, 2025

Flaky Unit Tests #489

Closed

fgrosse added 3 commits September 18, 2025 21:40

Merge branch 'main' into fgrosse/goleaks

df334d6

mcp: fix more issues found in unit tests

99ab514

mcp: close the mcpConn in the ServerSession

eec8eeb

fgrosse changed the title ~~Fix goroutine leaks in unit tests~~ mcp: fix goroutine leaks in unit tests Sep 19, 2025

fgrosse added 2 commits September 19, 2025 09:56

mcp: debug TestClientReplay

2ff35bc

Make leak checking opt-in

37423ef

fgrosse mentioned this pull request Sep 19, 2025

Proposal: Add StreamableHTTPOptions.SessionTimeout (Ungraceful client disconnects leads to zombie sessions) #499

Open

fgrosse added 2 commits September 19, 2025 11:14

mcp: mention modelcontextprotocol#499 in TODO note of leakCheck test …

18d642b

…flag

mcp: minor polishing of changed test code before review

3b84ae3

fgrosse commented Sep 19, 2025

View reviewed changes

fgrosse marked this pull request as ready for review September 19, 2025 09:26

findleyr reviewed Sep 29, 2025

View reviewed changes

fgrosse added 4 commits October 6, 2025 16:32

mcp: simplify ServerSession.Close() based on review feedback

a43fb6f

Merge branch 'main' into fgrosse/goleaks

f5ab129

mcp: add TODO, referencing modelcontextprotocol#499

f41256c

mcp: add comment about nil write in unit tests

415ff56

findleyr reviewed Oct 7, 2025

View reviewed changes

mcp: fix goroutine leaks in unit tests #496

Are you sure you want to change the base?

mcp: fix goroutine leaks in unit tests #496

Uh oh!

Conversation

fgrosse commented Sep 18, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

fgrosse commented Sep 24, 2025

Uh oh!

findleyr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

fgrosse commented Oct 6, 2025

Uh oh!

findleyr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

findleyr commented Oct 11, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants