Merge pull request #6281 from wuji-labs/feat/http2-long-stream-keepalive

feat(upstream): add HTTP/2 PING keepalive for long LLM generation streams

backend/internal/service/http_upstream_profile.go

@@ -7,9 +7,10 @@ import "context"
 type HTTPUpstreamProfile string
 
 const (
-	HTTPUpstreamProfileDefault HTTPUpstreamProfile = ""
-	HTTPUpstreamProfileOpenAI  HTTPUpstreamProfile = "openai"
-	HTTPUpstreamProfileGrok    HTTPUpstreamProfile = "grok"
+	HTTPUpstreamProfileDefault    HTTPUpstreamProfile = ""
+	HTTPUpstreamProfileOpenAI     HTTPUpstreamProfile = "openai"
+	HTTPUpstreamProfileGrok       HTTPUpstreamProfile = "grok"
+	HTTPUpstreamProfileLongStream HTTPUpstreamProfile = "long_stream"
 )
 
 type httpUpstreamProfileContextKey struct{}
@@ -37,7 +38,7 @@ func HTTPUpstreamProfileFromContext(ctx context.Context) HTTPUpstreamProfile {
 		return HTTPUpstreamProfileDefault
 	}
 	switch profile {
-	case HTTPUpstreamProfileOpenAI, HTTPUpstreamProfileGrok:
+	case HTTPUpstreamProfileOpenAI, HTTPUpstreamProfileGrok, HTTPUpstreamProfileLongStream:
 		return profile
 	default:
 		return HTTPUpstreamProfileDefault

backend/internal/service/http_upstream_profile_test.go

@@ -20,6 +20,13 @@ func TestWithHTTPUpstreamProfile_OpenAI(t *testing.T) {
 	}
 }
 
+func TestWithHTTPUpstreamProfile_LongStream(t *testing.T) {
+	ctx := WithHTTPUpstreamProfile(context.TODO(), HTTPUpstreamProfileLongStream)
+	if profile := HTTPUpstreamProfileFromContext(ctx); profile != HTTPUpstreamProfileLongStream {
+		t.Fatalf("expected profile %q, got %q", HTTPUpstreamProfileLongStream, profile)
+	}
+}
+
 func TestWithHTTPUpstreamRedirectsDisabled(t *testing.T) {
 	//nolint:staticcheck // Exercises the defensive nil-context fallback.
 	ctx := WithHTTPUpstreamRedirectsDisabled(nil)