Home › Learn › Databricks compute and cost › Sharing and regions
Databricks compute and cost · data engineering
Delta Sharing inside, and how to avoid cross-region cost on Databricks
How Delta Sharing works internally (signed file links read straight from storage), when it helps with cost, why cross-region transfer is charged and the seven places it hides in Databricks.
Prices, defaults, limits and preview status change often, and dollar figures here are illustrations. Check the current Databricks documentation and price list before quoting a number.
Sharing and regions
Delta Sharing: does it help, and how it works inside
How it works
Delta Sharing is an open protocol. The data owner (provider) publishes a share that lists tables. A reader (recipient) gets a credential, or for Databricks to Databricks, a trust link inside Unity Catalog. The key idea: the sharing server hands out short-lived signed links to the Parquet files, and the reader downloads straight from cloud storage. The server never carries the data.
- Table versions. The server reads the Delta log to know which files make up the current version, or a past version, and which files can be skipped using min/max stats.
- Two flavours. Databricks to Databricks uses Unity Catalog trust and needs no token. Open sharing gives other tools a token or federated identity.
- Limits you set. You can share a table, a view or a partition slice, and use dynamic views for row and column filtering.
Does it help with system table cost?
It helps indirectly, and only in one pattern. It does not make the original system tables cheaper to read. It helps when you build one central FinOps workspace, create the small rollup table there, and share the rollup to other workspaces, regions or accounts. Each BU then reads the small shared table instead of running its own heavy system table queries.
| Use of Delta Sharing | Helps with cost? | Verdict |
|---|---|---|
| Share your small rollup table with BU workspaces or other accounts | Yes. One heavy read, many small reads. | Pursue |
| Share the raw system tables onward | No. Reads are still big, and the cross-region data transfer may cost more. | Drop |
| Share a table across regions or clouds repeatedly | Costs egress. See next section. | Pursue with a replica, not repeated reads |
Sharing and regions
Cross-region cost, and how to avoid it on Databricks
You are right that moving data between regions costs money. The cloud providers charge for data leaving a region (and sometimes between availability zones), while data moving inside a region between compute and storage is normally free. Databricks does not add this fee. It is a cloud provider charge that appears on a cloud invoice (or on the Databricks one for serverless networking). Check your provider's current rates, but the order of magnitude is a few cents per gigabyte.
| Where it appears in Databricks | Why | How to avoid |
|---|---|---|
| A cluster or warehouse in region A reads storage in region B | Every read crosses the region boundary | Keep compute and the data it reads in the same region. Create the workspace where the data is. |
| Delta Sharing to a recipient in another region or cloud | The provider's storage sends files out of the region | Share small aggregates, not raw tables. Replicate once to the recipient's region and read the replica. Use a zero-egress replica location such as Cloudflare R2 if your setup allows. |
| Disaster recovery copy in another region | One-time and ongoing transfer | Replicate only what matters, use incremental deep clone, compress, and test the cost before turning it on. |
| Metastore is regional | A workspace in another region needs its own metastore, so people copy or share data across | Plan one metastore per region from the start. Put catalogs where their data lives. |
| Multiple availability zones in one region | Shuffle traffic between nodes in different zones can be charged | Where supported, keep a cluster in one zone for shuffle-heavy work. |
| Repeated reads of the same remote data | You pay every time | Copy once to local storage, or rely on the worker SSD disk cache so a second read stays local. |
| Serverless compute reading from a different region | Serverless runs in the workspace's region. Storage elsewhere means crossing regions. | Put the storage in the workspace's region. |
The rule: move compute to data, or move a small summary of the data, never the raw bulk, and never repeatedly.
FAQ
Questions people ask
How does Delta Sharing work internally?
The recipient asks a sharing server, which checks permission and returns short-lived signed links to the Parquet files. The recipient downloads straight from the provider's cloud storage. The server never carries the data.
How do I avoid cross-region cost?
Keep compute and storage in the same region, share small aggregates instead of raw tables, replicate once to the recipient's region, and never read big remote tables repeatedly.
Next
Where to go next
Related on this site: Cheapest way to move data in and out · UniForm
All Databricks compute and cost pages · All data engineering · Databricks on this site