How do you get visibility across sites when every warehouse runs a different WMS?

• 01Conformance belongs in the semantic layer, not in the source systems. Every WMS keeps running and is accessed read-only; the mapping from its fields to the shared metric definition lives in one versioned place you can audit, instead of being re-implemented inside a dozen report-level formulas that drift apart.
• 02The real deliverable is a metric dictionary, one row per KPI. For each metric it fixes the grain (line, unit, carton or order), which WMS event counts as the timestamp (pick confirm, task close or wave complete), the inclusion rules for cancelled lines, kitting, cross-dock and returns, the unit-of-measure conversion between eaches, cases and pallets, and the site calendar, shift boundary and time zone. Sites disagree on those five things far more often than the SQL is wrong.
• 03Extraction surface varies by site and sets the sequencing. A cloud WMS generally exposes a REST API and a scheduled flat-file extract; an on-premises or vendor-hosted WMS is normally read from a SQL replica or a reporting schema rather than from production, because report queries against a live pick-path database are felt on the floor. Firewalled sources are reached through an on-premises data gateway, which opens an outbound connection only.
• 04Incremental refresh keys off a row-level watermark, usually a last-modified timestamp, so only changed records move. Two WMS-specific failure modes break it: tables updated in place without a trustworthy modified column, where the transaction or history table has to be read instead, and sites in different time zones writing local time, where a naive watermark silently skips late-arriving rows. Both surface during the first site, not at go-live.
• 05When one site changes WMS, only that site’s pipeline and field mapping are rebuilt — the conformed model, the KPI definitions and every dashboard stay as they are. Record the mapping with an effective date and keep the already-conformed history in OneLake, or the cutover appears as an unexplained step change in the trend.
• 06The honest limit: if a site does not capture the event at all, no semantic layer can manufacture it. A paper-pick or scan-at-pack operation has no line-level pick timestamp, so it can be compared at order level and no further until the process on the floor changes. That is a warehouse decision, not a reporting one.

For the full explainer on this topic rather than this specific question, see the detailed guide on the blog.