Seems to me like anywhere Python is too bulky to work, a layer serializing and deserializing JSON at every pipeline step is likely to exhaust your memory too.