I already have a trained model (that does classification) that takes as input timeseries of shape [2500, 3]. I exported it on vertex as a custom model and created a public endpoint from it. I can request it with other timeseries and get them classified without any problem. But my issue is that, I actually need to make a lot of request: One instance of my object (on which I want to classify every timestamp) has a size of ~ [2 000 000, 3]. If I want to get a prediction at each timestamp, I'd need to take sliding windows (like here), and therefore I could have a lot of windows to predict (depending on the gap I choose between two consecutive windows, right now i'm using 500).
So I'm wondering, shouldn't I have some kind of pipeline on gcloud in which the sliding windows - and the classification is computed? Because right now I'm really limited by the 1.5MB (which is around ~10 windows of data for me) max payload size per request for vertex public endpoints. If so, which tool is designed to do this kind of stuff?