Field notes
Freshness SLAs that analysts actually trust
Uptime monitors tell you whether a cluster answers queries. They do not tell a commercial director whether yesterday's orders landed before the 08:30 briefing.
Agree the decision window first. If the briefing needs data by 08:15, the pipeline SLA must finish earlier, with buffer for retries. Publish that clock on the same page as the dashboard.
Measure arrival, not job success. A job can exit zero while writing empty partitions. Pair row-count and watermark checks with the schedule.
Communicate exceptions in plain language. 'Pipeline delayed; board pack uses Friday cut' is more useful than an opaque alert code in a shared channel.
During warehouse modernisation work we often find that tightening one critical mart's freshness does more for trust than adding dozens of low-value checks elsewhere.