Persistables
Persistables are the fundamental blocks of SimpleML objects. They are defined as the abstract base that details the expected methods and attributes of all inheriting subclasses.
SQLAlchemy
All persistables are backed by database records. This provides the seamless transition between development and production. The other major benefit is a full record of all iterations allowing for powerful querying and improvement tracking. The database orchestration is conducted via sqlalchemy for compatibility with all supported databases.
Registry
The gatekeeper syncing the database backend to the defined classes is the SimpleML-Registry. As the name suggests, it gets injected automatically into all classes that inherit from the base persistable and registers them on import. This means that references can be stored in the database (not the actual classes) and training can be scheduled using deferred execution and graph construction. The only limitation here is that any referenced class needs to be imported into the global namespace (and the registry) before it is invoked.
Standard practice to ensure all classes are imported is to source them in their respective package __init__ files and import the top level module before training or scoring.
Hashing
SimpleML exposes an optional top-level attribute for each persistable called the hash. The intent is to unambiguously capture the instance in order to compare references. The downstream use cases extend from conditional training (only train if new) to graph analytics (which models share the same datasets or pipelines).
A custom hashing method is included in the base persistable that is configured to handle all data types consistently across installations. Each inheriting class can then define which attributes to include in the hash. For use with conditional training, it is essential that the hash only consist of properties that exist upon initialization. If the hash is dependent on values that get modified during the training process, it will be impossible to compare new instances without first training them and comparing the resulting hashes.