app/vtstorage/netselect: return 502 instead of empty 200 on non-200 vtstorage response

app/vtstorage/netselect/netselect.go

@@ -807,10 +807,14 @@ func getFirstError(errs []error, allowPartialResponse bool) error {
 
 	// allowPartialResponse == true. Return the error only if all the backends are unavailable
 	// or if some of the backends are improperly configured.
+	//
+	// Check all the errors, since nil may come from a backend with a canceled query,
+	// because another backend returned an error. See https://github.com/VictoriaMetrics/VictoriaLogs/issues/1935
+	hasSuccess := false
 	for _, err := range errs {
 		if err == nil {
-			// At least a single vtstorage returned full response.
-			return nil
+			hasSuccess = true
+			continue
 		}
 		if !isUnavailableBackendError(err) {
 			// Return the first error, which isn't related to the backend unavailability, to the client,
@@ -821,6 +825,11 @@ func getFirstError(errs []error, allowPartialResponse bool) error {
 		}
 	}
 
+	if hasSuccess {
+		// At least a single vtstorage returned full response.
+		return nil
+	}
+
 	err := fmt.Errorf("all the vtstorage nodes are unavailable for querying; a sample error: %w", errs[0])
 	return newStatusBadGatewayError(err)
 }

app/vtstorage/netselect/netselect_test.go

@@ -0,0 +1,51 @@
+package netselect
+
+import (
+	"errors"
+	"net/http"
+	"testing"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+)
+
+func TestGetFirstError(t *testing.T) {
+	errUnavailable := &errUnavailableBackend{
+		err: errors.New("cannot connect to storage node"),
+	}
+	errResponse := errors.New("unexpected response status code")
+
+	f := func(errs []error, allowPartialResponse bool, errExpected error) {
+		t.Helper()
+
+		err := getFirstError(errs, allowPartialResponse)
+		if errExpected == nil {
+			if err != nil {
+				t.Fatalf("unexpected error: %s", err)
+			}
+			return
+		}
+
+		if !errors.Is(err, errExpected) {
+			t.Fatalf("unexpected error\ngot\n%v\nwant the error wrapping\n%v", err, errExpected)
+		}
+		esc, ok := errors.AsType[*httpserver.ErrorWithStatusCode](err)
+		if !ok || esc.StatusCode != http.StatusBadGateway {
+			t.Fatalf("expecting %d status code for the error %v", http.StatusBadGateway, err)
+		}
+	}
+
+	// partial responses are disallowed
+	f([]error{nil, nil}, false, nil)
+	f([]error{nil, errUnavailable}, false, errUnavailable)
+
+	// partial responses are allowed
+	f([]error{nil, nil}, true, nil)
+	f([]error{nil, errUnavailable}, true, nil)
+	f([]error{errUnavailable, errUnavailable}, true, errUnavailable)
+	f([]error{errUnavailable, errResponse}, true, errResponse)
+	f([]error{errResponse, nil}, true, errResponse)
+
+	// handleError() returns nil for the query canceled because of an error at another backend.
+	// See https://github.com/VictoriaMetrics/VictoriaLogs/issues/1935
+	f([]error{nil, errResponse}, true, errResponse)
+}

docs/victoriatraces/changelog/CHANGELOG.md

@@ -12,6 +12,8 @@ The following `tip` changes can be tested by building VictoriaTraces components
 
 ## tip
 
+* BUGFIX: [VictoriaTraces cluster](https://docs.victoriametrics.com/victoriatraces/cluster/): properly return `502 Bad Gateway` error when one of `vtstorage` nodes returns non-200 status code and partial responses are allowed via `-search.allowPartialResponse`. Previously `vtselect` could return an empty or truncated response with `200 OK` status code depending on the order of nodes in `-storageNode`. See [VictoriaLogs issue #1935](https://github.com/VictoriaMetrics/VictoriaLogs/issues/1935).
+
 ## [v0.13.0](https://github.com/VictoriaMetrics/VictoriaTraces/releases/tag/v0.13.0)
 
 Released at 2026-11-05