This is foolhardy at best, and dangerous if some startup tries it on their MVP because while the MVP will run, adding features that require returning multiple documents on non-keyed field will require downloading all documents that might have the correct data and filtering it. If you haven't run into this then you will be amazed by how abysmally slow this gets as your data grows. It will be at this moment that you wish you had done it the old fashioned way with a properly normalized RDBMS.