search

autotraderuk / fastapi-mlflow

Deploy mlflow models as JSON APIs with minimal new code
Apache License 2.0
19 stars 4 forks source link

Support binary input fields #2

Open Andrew-Crosby opened 2 years ago

Andrew-Crosby commented 2 years ago

JSON does not natively support byte data, so I don't think it's possible to deploy a model with a binary input field.

bloomonkey commented 2 years ago

Thanks for reporting @Andrew-Crosby. In an ideal world, how would you want this to be supported? The options that occur to me are:

POST the binary as a simple blob (i.e. with Content-Type: application/octet-stream instead of application/json)

For: simplest to implement in model and client(s) Against: API does not directly reflect model input schema Against: precludes support for models with multiple inputs

Assume binary data columns/fields will be sent as base64 encoded

For: retains JSON as the API and Content-Type - and so supports multiple inputs Against: API does not directly reflect model input schema

Explicitly state that binary is not supported, suggest that MLflow model be modified to decode a base64 encoded column

For: retains JSON as the API and Content-Type - supports multiple inputs Against: forces MLflow model to address a concern specific to real-time inference up-front