Loosely inspired by Flask

Server:

Create app.py

import aflight

app = aflight.Aflight()

@app.command("predict")
def predict(batch):
    result = do_something_with(batch)
    return [result]

To Run:

aflight

Client:

from aflight.client import predict

data = some_pandas_data()
print(predict(data))

Current kserve, torch/serve etc. APIs are too reliant on HTTP request/response paradigm

Using Arrow as a data format and Arrow Flight as a service API produces a much better interface:

• standard
• has schema
• fast
• columnar
• batch-able