A Week of Double-WritesPipeline · hard · ~15 minDuring last week's worker instability (evictions, retries), the fact_payments loader — a plain INSERT-append task — re-ran partially or fully on SEVERAL days. Finance reconciliation is off by +2.3% for the week. Nobody knows exactly which of the 7 daily intervals are duplicated, or by how much.This is a blast-radius incident: bound it, fix the mechanism, rebuild the data, and prove it's clean.Evidence

Evidence

Config

Checkpoints

First move in a blast-radius incident like this?

The rebuild plan, in the right order?

Design the verification + the go-forward guarantees, and narrate the incident end-to-end for a postmortem. Interview-style.