I use a bottom-up approach: I first identify which downstream component is failing, then reason about which upstream components feed into it. From there, I trace back step by step, checking logs, data outputs, and intermediate results in each upstream branch until I pinpoint the stage that introduced the error.