Multi-cloud shortcuts

S3, ADLS, GCS, and Dataverse shortcuts — caching behavior, egress cost, and consistency gotchas.

A shortcut is a pointer. It makes data that lives in another storage account or cloud appear as a folder or table inside a Fabric lakehouse, with no copy and no pipeline. Every Fabric engine can then read it as if it were local OneLake.

CU impact

Shortcuts themselves cost nothing to create. Reads through a shortcut consume the same CUs as a native read plus:

Cloud egress billed by the source provider (AWS/GCP) on cache misses.
Slightly higher latency on the first read of each file before it is cached.

Caching

Fabric caches shortcut data on read. Cache entries live up to 24 hours and are evicted on an LRU basis per capacity.

The cache is not invalidated when the source file changes. If an S3 object is overwritten in place, Fabric may serve the stale cached copy for up to 24h. Design upstream jobs to write new files (immutable naming) rather than overwrite, or accept the staleness window.

Source types and gotchas

Auth via access key + secret, stored in a connection. Use a scoped IAM policy (s3:GetObject, s3:ListBucket on one prefix).
s3compat endpoint type covers MinIO, Cloudflare R2, Wasabi — set the custom endpoint URL and keep path-style addressing in mind.
Egress: every cache miss is a GET billed by AWS. Co-locate the Fabric capacity region with the bucket region.

Delta tables through shortcuts

If the source folder is a valid Delta table (_delta_log present), create the shortcut under Tables/ and Fabric registers it in the SQL endpoint. Under Files/ it stays raw. Shortcut Delta tables are read-only from Fabric — writes must go back through the owning engine.

When to copy instead

The source is in a different region and read-heavy — egress + latency beat the storage cost of a mirrored copy.
You need V-Order or Fabric-side OPTIMIZE — you cannot rewrite files you do not own.
Compliance requires the data to physically reside in your tenant.