The 08:00 Dashboard Finance No Longer Trusts

Extending to six hours treats a capacity symptom and leaves three structural problems untouched. The window is already consuming most of the night, and source systems will keep growing; a one-hour buffer buys perhaps two quarters. More importantly, a longer window does not fix the $2.3M definition gap or the 1.8% late-arriving lines, so finance still sees a late and disputed number at 08:00. A better sequence is to (1) partition and incrementally load the largest facts instead of full rebuilds, (2) move the 14 full aggregate rebuilds to incremental or materialized views, (3) handle late-arriving data with a revision layer and a clearly labeled as-of timestamp, and (4) negotiate a freshness SLA by consumer tier rather than one 08:00 deadline. Extending the window may be acceptable as a two-week tactical bridge if the source team agrees and the business accepts a temporary 07:00 availability, but it must not be the plan of record. The judgment is to separate freshness from correctness and cost.

Sequence by dependency and risk, not by loudest complaint. First, freeze metric changes and charter a small revenue-definition working group with finance as accountable owner, sales ops, and the warehouse team; produce one documented definition of net revenue and the reconciliation rules to GL. In parallel, instrument the pipeline with row counts, timestamps, and freshness alerts so failures are visible before 08:00. Next, fix the highest-volume full rebuilds and late-arriving handling, because they cause both latency and restatement. Only then rebuild the sales mart against the agreed definition and retire the shadow spreadsheet by giving finance a reconciled, certified view. A common mistake is to rebuild aggregates first; that hardens the wrong definition into more objects and increases rework. Publish a semantic layer or certified dataset so dashboards consume one metric, and require certification for finance-facing reports. Set a realistic date for the one-number goal after two clean closes, and use the first close to compare old and new side by side rather than switching cold.

Resolve the variance by decomposing it, not by averaging. Reconcile at the order-line level: shipped-not-invoiced timing, $1.1M credits, $1.2M returns, and any currency or intercompany effects. Each component needs an owner and a rule. Decide explicitly whether the enterprise revenue metric is gross or net and whether it is recognized on shipment or invoice; the GL usually wins for statutory reporting, but sales may legitimately need a bookable pipeline view, which should be labeled differently. Prevent recurrence with a certified metric definition, a semantic layer, automated reconciliation thresholds that block publication when variance exceeds tolerance, and change control for source schema changes such as the 40 new columns. Add a data quality dashboard that finance reviews at close, and a monthly sign-off by the revenue metric owner. The judgment is that two numbers can coexist only if they have different names and purposes; two dashboards both called revenue cannot.