Skip to content
SQ

Data engineering24 March 20264 min

One canonical schema, many features

The least glamorous decision in Fidras, and the reason its features agree with each other.

An ecommerce intelligence platform accumulates features quickly: segmentation, CLV, churn, forecasting, recommendations, cart recovery, ad reporting. The fastest way to ship each one is a bespoke query against raw store data. The result is a product where churn and segmentation quietly disagree about who a customer is.

What the canonical layer buys

  • Every feature reads one tenant-scoped definition of customer, order and event.
  • A fix to the definition propagates everywhere instead of being patched per feature.
  • Model features and dashboard numbers come from the same place, so a merchant can reconcile them.
  • Multi-tenancy is enforced in one layer rather than remembered in every query.

The cost

It is slower at the start, and it feels like overhead while the product has two features. The payoff arrives the first time a merchant asks why the churn number and the segment count disagree — and there is a single answer instead of an investigation.

Interpretability as a product requirement

The same instinct drove the modelling choices. RFM and gradient boosting won over heavier models not on score but because an operator will only act on a prediction they can interrogate. A segment that cannot be described in one sentence does not get campaigned to.