migrating-features-to-chalk
Installation
SKILL.md
Migrating Existing Features to Chalk
Overview
Most ML teams already compute features — in dbt models, Prefect/Airflow-scheduled pandas jobs, Spark pipelines, or a feature store like Feast/Tecton. Migrating to Chalk is not a line-by-line code translation. It's a re-modeling: you describe what each feature is and how to compute one (a resolver), and Chalk owns the when (scheduling, materialization, caching, point-in-time replay) that your old stack hand-rolled.
The core mental shift:
| Legacy stack | Chalk |
|---|---|
| A job writes a feature table on a schedule (dbt run, Prefect flow, Spark job) | A resolver computes a feature value; Chalk schedules/materializes it |
| Online/serving features re-implemented separately from training → train-serve skew | One resolver serves both online and offline (backtest) queries |
Point-in-time correctness done by hand (asof joins, snapshot tables) |
Now + offline query replay give it for free |
| Orchestration (Prefect/Airflow DAG) is code you maintain | Dropped — Chalk is the scheduler |
This skill owns the migration-specific work: inventory, entity modeling, construct selection, and parity verification. It does not re-teach resolver mechanics — it hands off to the skill that owns each construct (see Handoffs).