Swapping the brain without stopping the body

Context

Inventory allocation ran on manually uploaded allocation factors: a human shaped the size curve, the system obeyed. The business wanted data driven size and sales curves computed from warehouse data, and new market segmentation (domestic vs. Europe ecommerce) that the manual process couldn’t express.

Problem

A hard cutover to computed curves would have bet daily operations on a brand new pipeline. If BigQuery was slow, a mapping was missing, or the data was insufficient for a confident curve, allocation work would simply stop, and nobody would be able to say why the system recommended what it did.

The decision

Treat the computed curve as an optional strategy, not a replacement. Users chose per run: BigQuery computed curve, or the legacy manual factors. The integration lived behind a request/response boundary governed by an explicit contract and validations spec: required inputs, allowed values, expected success shapes, and every failure mode (BigQuery unavailable, missing mappings, insufficient data) with its exact fallback behavior.

REJECTED: a hard cutover. Correctness of a forecasting pipeline is earned statistically, over runs; the rollout design had to let trust accumulate while operations continued.

Execution

Validations were enforced at the service boundary so the UI and downstream consumers saw consistent, predictable results instead of ad hoc errors. Every run persisted its request parameters and computed outputs, making “why did it allocate this way?” an answerable question and regressions in the market segmentation logic detectable. Adoption was measured, not assumed: fallback rate, validation failures, and BigQuery error rates tracked over time, tuning thresholds and mappings until the computed path could safely become the default.

Outcome

2 paths

Computed curve, with legacy fallback intact

Contract-first

Explicit success and failure modes

Replayable

Every run persisted for audit

Takeaway

Ship the new brain as an optional strategy. A contract, a fallback, and an audit trail turn a risky cutover into a choice.