Decouple stream cancellation and clarify failover rules
This commit is contained in:
@@ -351,6 +351,48 @@ VALUES ($1, 5, '{"purpose":"core-flow"}'::jsonb)`, inviteCode); err != nil {
|
||||
t.Fatalf("unexpected compatible chat response: %+v", compatChat)
|
||||
}
|
||||
|
||||
cancelMarker := "cancel-stream-" + suffixText
|
||||
cancelCtx, cancelRequest := context.WithCancel(context.Background())
|
||||
cancelPayload := map[string]any{
|
||||
"model": defaultTextModel,
|
||||
"runMode": "simulation",
|
||||
"messages": []map[string]any{{"role": "user", "content": "cancelled stream"}},
|
||||
"stream": true,
|
||||
"simulation": true,
|
||||
"simulationDurationMs": 250,
|
||||
"cancelTestId": cancelMarker,
|
||||
}
|
||||
cancelRaw, err := json.Marshal(cancelPayload)
|
||||
if err != nil {
|
||||
t.Fatalf("marshal cancelled stream payload: %v", err)
|
||||
}
|
||||
cancelReq, err := http.NewRequestWithContext(cancelCtx, http.MethodPost, server.URL+"/v1/chat/completions", bytes.NewReader(cancelRaw))
|
||||
if err != nil {
|
||||
t.Fatalf("build cancelled stream request: %v", err)
|
||||
}
|
||||
cancelReq.Header.Set("Authorization", "Bearer "+apiKeyResponse.Secret)
|
||||
cancelReq.Header.Set("Content-Type", "application/json")
|
||||
cancelErrCh := make(chan error, 1)
|
||||
go func() {
|
||||
resp, err := http.DefaultClient.Do(cancelReq)
|
||||
if resp != nil {
|
||||
_, _ = io.ReadAll(resp.Body)
|
||||
_ = resp.Body.Close()
|
||||
}
|
||||
cancelErrCh <- err
|
||||
}()
|
||||
cancelTaskID := waitForTaskIDByRequestMarker(t, ctx, testPool, cancelMarker, 2*time.Second)
|
||||
cancelRequest()
|
||||
select {
|
||||
case <-cancelErrCh:
|
||||
case <-time.After(time.Second):
|
||||
t.Fatal("cancelled stream request did not return after client cancellation")
|
||||
}
|
||||
cancelledStreamTask := waitForTaskStatus(t, server.URL, apiKeyResponse.Secret, cancelTaskID, []string{"succeeded"}, 2*time.Second)
|
||||
if cancelledStreamTask.Status != "succeeded" {
|
||||
t.Fatalf("client-cancelled compatible stream should keep backend task running to success, got %+v", cancelledStreamTask)
|
||||
}
|
||||
|
||||
var imageResponse struct {
|
||||
Task struct {
|
||||
ID string `json:"id"`
|
||||
@@ -1398,6 +1440,26 @@ func waitForTaskStatus(t *testing.T, baseURL string, token string, taskID string
|
||||
return detail
|
||||
}
|
||||
|
||||
func waitForTaskIDByRequestMarker(t *testing.T, ctx context.Context, pool *pgxpool.Pool, marker string, timeout time.Duration) string {
|
||||
t.Helper()
|
||||
deadline := time.Now().Add(timeout)
|
||||
for time.Now().Before(deadline) {
|
||||
var taskID string
|
||||
err := pool.QueryRow(ctx, `
|
||||
SELECT id::text
|
||||
FROM gateway_tasks
|
||||
WHERE request->>'cancelTestId' = $1
|
||||
ORDER BY created_at DESC
|
||||
LIMIT 1`, marker).Scan(&taskID)
|
||||
if err == nil && taskID != "" {
|
||||
return taskID
|
||||
}
|
||||
time.Sleep(50 * time.Millisecond)
|
||||
}
|
||||
t.Fatalf("task with request marker %s was not created within %s", marker, timeout)
|
||||
return ""
|
||||
}
|
||||
|
||||
func assertLoadAvoidanceSimulatedRetryChain(t *testing.T, ctx context.Context, testPool *pgxpool.Pool, baseURL string, adminToken string, runtimeToken string, suffixText string) {
|
||||
t.Helper()
|
||||
model := "load-avoidance-smoke-" + suffixText
|
||||
|
||||
@@ -1,6 +1,7 @@
|
||||
package httpapi
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"fmt"
|
||||
@@ -564,10 +565,15 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
|
||||
writeTaskAccepted(w, task)
|
||||
return
|
||||
}
|
||||
runCtx, cancelRun := s.requestExecutionContext(r)
|
||||
defer cancelRun()
|
||||
if compatible {
|
||||
if boolValue(body, "stream") {
|
||||
flusher := prepareCompatibleStream(w)
|
||||
result, runErr := s.runner.ExecuteStream(r.Context(), task, user, func(delta string) error {
|
||||
result, runErr := s.runner.ExecuteStream(runCtx, task, user, func(delta string) error {
|
||||
if !requestStillConnected(r) {
|
||||
return nil
|
||||
}
|
||||
writeCompatibleDelta(w, kind, model, delta)
|
||||
if flusher != nil {
|
||||
flusher.Flush()
|
||||
@@ -575,6 +581,9 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
|
||||
return nil
|
||||
})
|
||||
if runErr != nil {
|
||||
if !requestStillConnected(r) {
|
||||
return
|
||||
}
|
||||
status := statusFromRunError(runErr)
|
||||
errorPayload := map[string]any{
|
||||
"code": runErrorCode(runErr),
|
||||
@@ -593,29 +602,66 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
|
||||
}
|
||||
return
|
||||
}
|
||||
if !requestStillConnected(r) {
|
||||
return
|
||||
}
|
||||
writeCompatibleDone(w, kind, model, result.Output)
|
||||
if flusher != nil {
|
||||
flusher.Flush()
|
||||
}
|
||||
return
|
||||
}
|
||||
result, runErr := s.runner.Execute(r.Context(), task, user)
|
||||
result, runErr := s.runner.Execute(runCtx, task, user)
|
||||
if runErr != nil {
|
||||
if !requestStillConnected(r) {
|
||||
return
|
||||
}
|
||||
writeError(w, statusFromRunError(runErr), runErr.Error(), runErrorCode(runErr))
|
||||
return
|
||||
}
|
||||
if !requestStillConnected(r) {
|
||||
return
|
||||
}
|
||||
writeJSON(w, http.StatusOK, result.Output)
|
||||
return
|
||||
}
|
||||
result, runErr := s.runner.Execute(r.Context(), task, user)
|
||||
result, runErr := s.runner.Execute(runCtx, task, user)
|
||||
if runErr != nil {
|
||||
s.logger.Warn("task completed with failure", "kind", kind, "taskId", task.ID, "error", runErr)
|
||||
}
|
||||
|
||||
if !requestStillConnected(r) {
|
||||
return
|
||||
}
|
||||
writeTaskAccepted(w, result.Task)
|
||||
})
|
||||
}
|
||||
|
||||
func (s *Server) requestExecutionContext(r *http.Request) (context.Context, context.CancelFunc) {
|
||||
base := context.WithoutCancel(r.Context())
|
||||
if s.ctx == nil {
|
||||
return base, func() {}
|
||||
}
|
||||
ctx, cancel := context.WithCancel(base)
|
||||
go func() {
|
||||
select {
|
||||
case <-s.ctx.Done():
|
||||
cancel()
|
||||
case <-ctx.Done():
|
||||
}
|
||||
}()
|
||||
return ctx, cancel
|
||||
}
|
||||
|
||||
func requestStillConnected(r *http.Request) bool {
|
||||
select {
|
||||
case <-r.Context().Done():
|
||||
return false
|
||||
default:
|
||||
return true
|
||||
}
|
||||
}
|
||||
|
||||
func asyncRequest(r *http.Request) bool {
|
||||
value := strings.TrimSpace(strings.ToLower(r.Header.Get("x-async")))
|
||||
return value == "1" || value == "true" || value == "yes" || value == "on"
|
||||
|
||||
Reference in New Issue
Block a user