Db2mlops

[pwharned]

I found a procedure for managing and storing models in DB2 that I thought would be useful.

1. We register a single python UDTF that also takes a b64 encoded string as an additional parameter
2. Store the model as a b64 encoded string in a table in the database
3. Model is passed into the UDTF at runtime

Provides a regular SQL interface for doing CRUD with new models and storing them in the database.

Improvement suggestions welcome - I'm not sure what is the best datatype for a variable length chararray/string in the UDTF...

[kdrodger]

Thanks Patrick. Just a couple of relatively minor comments.

[pwharned]

My understanding is that joblib is faster than pickle particularly with numpy arrays - not sure how significant that is.

[kdrodger]

My understanding is that joblib is faster than pickle particularly with numpy arrays - not sure how significant that is.

I think I've seen that in some cases as well, and agreed, not sure that it really matters in practice. The only thing I'd maybe consider is documenting in the comments if there's a reason why a specific choice like that is made. Thanks.