Databricks DE · Topic 9.1

Delta Sharing

An open protocol for sharing live data with other organizations — zero-copy, read-only, governed, and revocable. Share to Databricks (D2D) or any client (open sharing). This deep dive walks both sides with the code.

zero-copyD2D vs opendeltaSharing readerMarketplace · Clean Rooms

The one-liner

Delta Sharing gives recipients read-only, live access to your data without copying it. D2D shares to another Databricks org (no tokens); open sharing shares tabular data to any client via a token profile. Objects: provider · share · recipient.

📖 Analogy: a library card for your data, not photocopies. You don't mail copies of the book (no data movement) — you issue a read-only card (the recipient) to read the live book where it sits, and you can cancel the card anytime.

Zero-copy sharing — and instant revoke

▸ no data moves; access is a grant you can pull back

Provider

owns the data (stays here)

— share (no copy) →

Recipient

read-only live access

Access granted — recipient reads live data in place.

1 · Provider side — build & grant a share

CREATE SHARE sales_share COMMENT 'Curated sales for partners';
ALTER SHARE sales_share ADD TABLE prod.sales.orders;

-- D2D recipient: bind to the partner's Databricks sharing identifier (no tokens)
CREATE RECIPIENT partner_db USING ID 'aws:us-east-1:<sharing-identifier>';
-- Open recipient: omit USING ID → Databricks generates an activation link/token
CREATE RECIPIENT partner_bi;

GRANT SELECT ON SHARE sales_share TO RECIPIENT partner_db;
GRANT SELECT ON SHARE sales_share TO RECIPIENT partner_bi;

2 · Recipient side — consume the share

D2D — appears as a read-only catalog

SELECT * FROM partner_inbound.sales.orders;   -- query like any UC table

Open — credential profile + deltaSharing reader (any client)

df = (spark.read.format("deltaSharing")
        .load("/path/to/config.share#sales_share.sales.orders"))   # profile#share.schema.table

3 · D2D vs Open sharing — click each

Databricks-to-Databricks

Open sharing

D2D — recipient is another UC-enabled Databricks account.

No token management — identity flows through the platform
Can share tables, notebooks, volumes, models
Consumed as a read-only catalog; use when the partner is on Databricks

Open sharing — recipient is any client (pandas, Power BI, Spark…).

Auth via bearer token / OIDC profile
Tabular data, read with the deltaSharing connector
Use when the partner is NOT on Databricks

	D2D	Open
Recipient	Databricks org	Any client
Auth	No tokens	Bearer token / OIDC
Consumed as	read-only catalog	`deltaSharing` reader
Shareable	Tables + notebooks/volumes/models	Tabular data

4 · Governance & ecosystem

Fine-grained limits: a share grants whole tables — layer dynamic views (8.3) for row/column restrictions before sharing.
Revoke instantly (REVOKE / drop recipient); rotate open-sharing tokens.
Databricks Marketplace is built on Delta Sharing — publish/consume data products. Clean Rooms let two parties run joint analysis without either seeing the other's raw rows.

Why it matters

Traditional sharing = FTP/export copies → stale, ungoverned, costly. Delta Sharing is zero-copy, live, governed, revocable.
"Share with a non-Databricks partner?" → open sharing; "with another Databricks org?" → D2D.

🤝 Real-world use case: a provider shares a curated sales table to a partner's BI tool via open sharing and to a sister BU's workspace via D2D — both read the same live data, no copies, revocable instantly.

Uses, edge cases & limitations

Uses: B2B exchange, cross-region/cloud sharing, internal cross-org sharing, distributing datasets to non-Databricks tools, Marketplace publishing, Clean Rooms collaboration.
Edge cases: open-sharing tokens must be rotated/secured; row/column limits need dynamic views; the D2D sharing identifier must be exchanged correctly.
Limitation: recipients are read-only; notebooks/volumes/models are D2D only; it's sharing, not Lakehouse Federation (querying external sources — 9.2).

Common gotchas

❌ Thinking it copies data — it's zero-copy, live access in place.
❌ Using open sharing for a Databricks recipient when D2D (no tokens) is simpler.
❌ Leaking / never-rotating open-sharing tokens.
❌ Expecting fine-grained row/column control from the share alone — add dynamic views.
❌ Confusing Delta Sharing (share out) with Lakehouse Federation (query external in) — opposite directions.