Each project arrives with its purpose, milestones and first tasks written out. Here is how the first three read.
Feature inventory across production models
Most teams that start talking about a feature store cannot say how many features their live models actually consume, where each is computed, or which ones two models compute differently under the same name. A sheet listing every input of every production model, with its source tables, computation code and owner, is the starting point for every later decision in this area.
First tasks
- List every model currently serving predictions and who owns it
- Extract the input feature list from each model's saved artefact or serving code
- Record source tables, computation job and refresh frequency per feature
- Flag features that appear under different names or code in two models
- Refresh the feature inventory against the model registry @recurring(yearly)
Feature naming and definition standard
Names like user_score_v2_final tell nobody what was aggregated, over which window, or as of when. Agreeing one naming pattern (entity, measure, aggregation, window) and a short definition card for each feature lets an engineer read a training set without opening the code, and stops two people building the same thing twice.
First tasks
- Collect twenty current feature names and note what each fails to say
- Draft a naming pattern of entity, measure, aggregation and window
- Write definition cards for the ten most used features
- Get sign-off on the standard from the ML and data engineering leads
Entity keys and grain map for features
Every feature belongs to an entity at a grain: a customer per day, an order at checkout, a device per hour. When those keys are fuzzy, joins fan out silently, training rows multiply and a model learns from duplicated examples. Mapping the entities your models predict about, their join keys and the grain of each feature table removes a whole class of silent errors.
First tasks
- Write down the entities your models make predictions about
- Record the join key and grain of every feature table
- Document how merged or reissued keys are resolved
- Add a test that fails when a join changes the training row count