Keeping track of what produced what.
WHAT MUST BE VERSIONED
Code Data Model artefacts Environment Configuration
WHY DATA IS THE DIFFICULT ONE
It is too large for ordinary version control.
WHAT THE APPROACHES ARE
Tools storing data externally with references in version control Immutable, dated storage locations Table formats supporting reading as of a point in time
WHAT IMMUTABLE DATED STORAGE PROVIDES
Simplicity, and a reference that never changes.
WHY THAT IS FREQUENTLY ENOUGH
Most teams need reproducibility, not full data version control.
WHAT TO RECORD FOR EVERY MODEL
The exact data reference The code commit The configuration The environment The metrics achieved
WHAT THAT ENABLES
Rebuilding any deployed model.
WHY THAT MATTERS
Questions arise months later, about decisions the model made.
WHAT TO STORE ARTEFACTS IN
Object storage, with a registry recording the metadata.
WHAT TO NEVER DO
Overwrite a model artefact.
WHY
Any prediction attributed to it becomes unexplainable.
WHAT TO RETAIN
Every deployed version, for as long as its decisions matter.
