[pramsey]

There's some work to be done to make all functions parallelizeable. After that it gets harder. For large scale raster analysis, moving the processing closer to the data becomes paramount. I feel like purpose-built raster processing will always be better. Maybe something with FDW or other client/service architecture, where the database sends off the request and the bulk processing happens close to the data. Also, the row-based model of SQL sort of breaks down against raster where the model expects much much larger objects and presenting the tiled-up model to users just adds complexity without getting back performance or flexibility in return.