Import mode loads compressed data into the dataset. It fails operationally when everyone publishes 5 GB pilots onto one gateway and schedules them all at 05:30.
Adoption sequence
- Start with one certified Import model per domain; block duplicate shadow PBIX copies in My Workspace for the same facts.
- Stagger Scheduled refresh (e.g. finance 04:30, ops 05:15) and check Refresh history for overlapping Failed due to gateway congestion.
- Cap pilot model size early — use VertiPaq Analyzer; drop high-cardinality text, split date/time, remove unused columns before promoting.
Settings worth enforcing
In dataset settings, set a clear time zone for schedules (UTC vs local causes “missing morning data” tickets). Prefer large-date dimensions as Import even if some facts stay Dual/DirectQuery. When refreshing from SQL via gateway, keep Max parallel engine threads / container awareness with your DBA — saturating the source creates blocking that looks like a Power BI outage.
Only move to incremental refresh after Import nightly Success is boring for two weeks — incremental on a broken watermark multiplies partition Failures.