The other thing you can try here is running the nodes/pipeline outside kedro just as pure Python functions. This should help to see whether this is a problem involving the kedro runner at all