The SQL analytics endpoint is a separate layer that maintains its own catalog and cache over the Delta tables in OneLake. Keeping it in sync with the underlying data is a background process with real latency, so data can land in the lakehouse while the endpoint still returns stale (or missing) rows.
Two details make it worse than a fixed delay. The endpoint is serverless, so after a period of inactivity it spins down and the background refresh stops — the classic symptom is the first query of the morning returning yesterday’s data. And the refresh can simply stall: no error, no warning, just stale results. (Dennes Torres documented this behavior in detail, and the long-running “SQL endpoint lags behind lakehouse” thread on the Fabric community shows how widely it is felt.)
In a Direct Query model feeding a live report, this lag is visible directly to your users — which is exactly what the architecture was supposed to prevent.
Automated refresh from notebook is fine for batch pipelines where you can fire a refresh when ingestion finishes. However, it falls apart for continuous mirroring: there is no natural “end of ingestion” to trigger from.
