The memory footprint of the current approach to generating number district labels is a bit heavy. I think it's worth trying an approach that relies on the centers of mass and checking on the main thread map versus the more expensive but more correct approach that uses a secondary raster canvas on the worker thread.