mirror of
https://github.com/coder/coder.git
synced 2026-09-21 20:51:01 +08:00
Marcin Tojek
10bbe3b140
Fixes: https://github.com/coder/internal/issues/1451 ## Problem Flaky test `TestStreamAgentReinitEvents/doesn't_transmit_events_if_the_transmitter_context_is_canceled` (coder/internal#1451): ``` agentsdk_test.go:84: Error: Received unexpected error: Get "http://127.0.0.1:XXXXX": net/http: HTTP/1.x transport connection broken: http: CloseIdleConnections called ``` ## Root cause The subtests used `client := &http.Client{}`. A client with a nil `Transport` uses the process-global `http.DefaultTransport`, which is shared by every parallel test in the test binary. `httptest.Server.Close()` calls `http.DefaultTransport.CloseIdleConnections()`. When any other parallel test closes its `httptest.Server` while this test's request is in flight, the shared transport tears the connection down and `client.Do(req)` fails with `http: CloseIdleConnections called`. This is the same class of flake already documented/fixed in `testutil/oauth2.go` and the `mcphttpclient` helpers, and related to coder/internal#1020. ## Fix Give each client a dedicated `*http.Transport` (`&http.Client{Transport: &http.Transport{}}`) so cross-test `CloseIdleConnections` calls cannot break its requests. The construction is extracted into a small `newReinitTestClient()` helper used by all three subtests, with a comment documenting the reason. ## Verification `go test ./codersdk/agentsdk -run TestStreamAgentReinitEvents -count=20` passes. The flake was reproduced against the exact failing subtest logic (real `NewSSEAgentReinitTransmitter` with a pre-canceled transmit context, same client pattern) under a `CloseIdleConnections` stress loop: - Fix reverted to `&http.Client{}`: reliably FAILs (e.g. 15 broken requests in 10s). - Fix present: 0 broken requests across repeated runs. <details> <summary>Optional stress harness to reproduce/verify locally (not committed)</summary> Drop this into `codersdk/agentsdk/` as a throwaway `*_test.go` file. It runs the verbatim body of the failing subtest in a loop while parallel goroutines call `CloseIdleConnections` (exactly what `httptest.Server.Close()` does). With the fix present it reports `closeIdleErrs=0`; revert `newReinitTestClient` to `&http.Client{}` to reproduce. ```go package agentsdk_test import ( "context" "net/http" "net/http/httptest" "strings" "sync" "sync/atomic" "testing" "time" "github.com/google/uuid" "cdr.dev/slog/v3/sloggers/slogtest" "github.com/coder/coder/v2/codersdk/agentsdk" ) func TestFlakeReproRealSubtest(t *testing.T) { t.Parallel() ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) defer cancel() var wg sync.WaitGroup var closeIdleErrs int64 var sample atomic.Value defaultTransport := http.DefaultTransport.(*http.Transport) for range 8 { wg.Add(1) go func() { defer wg.Done() for ctx.Err() == nil { defaultTransport.CloseIdleConnections() } }() } for range 32 { wg.Add(1) go func() { defer wg.Done() for ctx.Err() == nil { // Verbatim body of the failing subtest. eventToSend := agentsdk.ReinitializationEvent{ WorkspaceID: uuid.New(), Reason: agentsdk.ReinitializeReasonPrebuildClaimed, } events := make(chan agentsdk.ReinitializationEvent, 1) events <- eventToSend transmitCtx, cancelTransmit := context.WithCancel(context.Background()) cancelTransmit() transmitErrCh := make(chan error, 1) srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { transmitter := agentsdk.NewSSEAgentReinitTransmitter(slogtest.Make(t, nil), w, r) transmitErrCh <- transmitter.Transmit(transmitCtx, events) })) req, err := http.NewRequestWithContext(ctx, "GET", srv.URL, nil) if err != nil { srv.Close() continue } client := newReinitTestClient() // revert to &http.Client{} to reproduce resp, err := client.Do(req) if err != nil { if strings.Contains(err.Error(), "CloseIdleConnections called") { atomic.AddInt64(&closeIdleErrs, 1) sample.CompareAndSwap(nil, err.Error()) } srv.Close() continue } resp.Body.Close() srv.Close() } }() } wg.Wait() t.Logf("closeIdleErrs=%d", atomic.LoadInt64(&closeIdleErrs)) if n := atomic.LoadInt64(&closeIdleErrs); n > 0 { t.Fatalf("reproduced coder/internal#1451 on the real subtest: %d requests broken (e.g. %v)", n, sample.Load()) } } ``` Example output with the fix reverted to `&http.Client{}`: ``` flakerepro_test.go:88: closeIdleErrs=15 flakerepro_test.go:90: reproduced coder/internal#1451 on the real subtest: 15 requests broken (e.g. Get "http://127.0.0.1:43755": net/http: HTTP/1.x transport connection broken: http: CloseIdleConnections called) --- FAIL: TestFlakeReproRealSubtest (10.14s) ``` With the fix present: `closeIdleErrs=0` and PASS. </details> --- Generated by Coder Agents on behalf of @mtojek.