This is placeholder text for a retrospective on rebuilding a batch data pipeline from scratch. It will open with the reasons for the rebuild — mainly accumulated complexity in the scheduling layer rather than any single outage.
A middle section will cover idempotency: how reruns were made safe by keying every job on a deterministic watermark instead of wall-clock time.
The closing section will likely touch on observability — what got logged, what got alerted on, and which of those turned out to be noise in practice.