This message was deleted.

I think it should work, from the perspective of BentoML it doesn't know(or care about) what you are sending in as input, since frames are essentially images/tensors/ndarrays

Great thanks @Jian Shen Yap, So our current workflows are to download video from URL, and then read in frame by frame. It looks like the only suitable IO descriptor might be a

multipart

file. Because a video file is so large, is there a different method that might work? Perhaps specifying an input file URL? Then could we specify a runner to stream results out to some data store? Seeing that most requests are RESTful, are you suggesting that we send frame data into the REST API over HTTP? For retrieving the results, because there is so much data and so many frames, can data be piped into a file? Last question, if we need to stream the data (ie. frame by frame to Bento ML -> Which would not be ideal) -> Would this only be through REST?

What you mentioned above could be a way to go, i.e. streaming data frame by frame. It could be via http or grpc. If latency is not something that matters, you could send a url as text, and configure the runner/service to download the file, you have full flexibility here. I agree streaming frame by frame is probably too heavy but it really depends on your engineering SLAs!