Preparation for model pipeline deployment
Exporting a model as a MOJO for final model deployment is trivial, for instance, consider the following:
final.download_MOJO("final_MOJO.zip")
Deployment of the MOJO in various architectures via multiple recipes is covered, in detail, in Chapter 9, Production Scoring and the H2O MOJO. In general, there is a significant amount of effort that must be assigned to productionizing data for model scoring. The key is that data used in production must have a schema identical to that of the training data used in modeling. In our case, that means all the data wrangling and feature engineering tasks must be productionized before scoring in production can occur. In other words, the process is simply as follows:
- Transform raw data into the training data format.
- Score the model using the MOJO on the transformed data.
It is a best practice to work with your DevOps or equivalent production team well in advance of model delivery...
