Managing ML Models From Iteration To Production With MLOps In Snowflake

Quick iteration with models running in production is the key to success. Snowflake’s Model Registry and MLOps features turn this process into a robust workflow for pushing models from development to production. This video shows how Snowflake can be used to manage lifecycle, access, and inferences of models.
Subscribe for more! www.snowflake.c...
Explore sample code, download tools, and connect with peers: developers.sno...

[@emanueol]

Great demo 👍
1. whats table size limit to train/infer/predict? I imagine Snowflake behind the scenes splits table rows and computes distributes to learn and to predict? Probably there's no hard limit in table size, but a doft limit to train multiple batches of rows, and repeat till full table trained right? Regarding prediction I suppose compute threads pick/predict row by row basis? or it depends on model type with some able to work in batches of rows?
2. finally you mentioned data set versioning? how does this physically maps to tables? is a version just a time travel timestamp (so we can reuse/access same data?)
Thanks

[@SumitDas-snow]

1. Train limit is based on your size of warehouse as data need to fit in the memory. No limit on infer/predict
2. Dataset is meant to be an immutable copy for the data for provenance / reproducibility. Table is mutable.

[@emanueol]

ok thanks, and how about the meaning of dataset version shown in 26m 33s in videov
name='AI307_DS.DATA.TRAINING_DATASET
version='v2'

[@christiansetzkorn6241]

Let us say I have a model (e.g. xgb) fitted using an external process and oly have a file (e.g. pickle). Can I add this to the registry and ultimately invoke from streamlit deployed in snowflake? Any pointers very much welcome! Thanks.

[@SumitDas-snow]

Today log_model() expect a python object instead of a pickle file. You can load the model from pickle file and then call log_model().