I am trying to connect a Fraud Detection ML model deployed on Azure ML to my Confluent Flink Streaming Workspace. I successfully set up the connection to AzureML, and created the model on Confluent.
When I try to run my model using:
SELECT * FROM ….my_table, LATERAL TABLE(ML_PREDICT(‘testmodel’, TransactionID, Product, card1, card2, card3,…))
,I get an “Internal Error” without any further specification.
My input consists of 300 features/columns, the output is one boolean.
I reproduced the same steps using Azure OpenAI (with one string-input and one string-output) and didn’t have any problems.
Did anyone experience the same problems or has an idea how to debug? Or at least has an idea how to find any information like proper logs?
We figured it out by trial and error: Confluent Flink remote model inference with Azure ML only accepts a maximum input of 84 features. The ominous “internal error” occurs when exceeding this limit. Our workaround: Parsing all inputs together as one JSON string. Then, in the model-run script accompanying your model on Azure ML, you can parse the string and convert it back to the accepted inputs. The input in COnfluent Flink AI then only exists of one string-feature