velox VS onnx-models

4. The workers (if distributed) or the engine (if single-box) executing the execution plan, by performing the actual computations on actual bytes.

(Note that this is a cartoon version and any given engine will have differences, eg Presto/Trino does not have a clear distinction between LP and Physical Plan.)

1) and 4) are broadly similar across engines, while 2) and 3) vary widely partially because of different requirements (reliability, latency, etc). Projects such as Apache Arrow and Velox (https://github.com/facebookincubator/velox) are making common tools for 4), and as mentioned ANSI SQL, ZetaSQL, Calcite, and Substrait are making common tools for 1).

No, the inference accuracy of the image classifier is dependent on the model used and this is a demo of the code executing the model in a browser with GPU acceleration not the model itself. You can plug and play any model in the onnx format e.g. https://github.com/onnx/models. As a comparison, complaining about the "abysmal quality" of the dummy model on display here is like saying blender is bad 3d modeling software after opening it for the first time because all it models is a blank cube.