Observability That Thinks: AI for Pipeline Monitoring

Median/MAD over mean plus stddev is the right call for pipeline metrics, since one giant batch or a retry storm wrecks a z-score but barely moves the median. The same-weekday baseline is the detail I'd steal, because a Monday reprocessing job looks nothing like a Sunday and averaging across the week hides both. How do you handle holidays and one-off backfills, do they poison the trailing-8 window, or do you exclude them explicitly?