Open sharing
Sharing makes platform-managed tables consumable by engines and organizations outside System9, over the protocols they already speak, under one governance model. A share is not a hole in the policy: the security plane is enforced on the sharing path, so the policy that governed the data at rest still governs it as it leaves.
A share is not an export. The policy that governed the data at rest is compiled and travels with it, and every read is logged against the recipient.
Two open protocols, governed the same way
| Protocol | For | How it is governed |
|---|---|---|
| Delta Sharing | Sharing with anyone, including non-System9 consumers | A full protocol server and client: recipient management, share/schema/table registration, profile issuance, credential auth, versioned and change-feed reads, an admin surface, and an audit log: with the security plane enforced on the path. |
| Iceberg REST catalog | Iceberg-native consumers (Spark, PyIceberg) | An operator-session admin API manages recipients and per-namespace and per-table grants, and credential issue / rotate / revoke, behind the same role-based access control as the Delta Sharing admin surface. |
Row- and column-level security is enforced on the served files themselves. Each recipient reads a filtered snapshot they cannot bypass; a revoked recipient's snapshots are dropped and are excluded from disaster-recovery replication. That is the same in-engine enforcement described in governed reads, applied to the sharing path.
The two admin surfaces are at functional parity, not byte-identical parity, Delta Sharing is organized around shares and recipients, the Iceberg REST catalog around recipients and namespace/table grants. Each covers the same ground in its own protocol's vocabulary, so a capability present on one is reachable on the other, under a different name and shape.
Sharing inside an estate
Beyond the open protocols, workspace shares move governed data inside one estate: share grants between workspaces, incoming and outgoing views, accept and decline, revoke, a share access log, and shares of a model, not only a table. A Direct Lake share builder supports BI consumption without a copy, and public datasets support open publication with access requests and approval.
Cross-organization sharing
When a share crosses an organization boundary, the compiled security policy travels with the data in a signed envelope rather than being re-implemented by the recipient, governed by trust-edge validation with classification limits. The next boundary out, a marketplace between organizations that do not share a secret, is the Data Exchange.
Clean rooms
A clean room lets two parties compute over shared data without either side reading the other's inputs, and a joint result can be promoted into a governed data product. The isolation model is a partition, not a filter on a shared table. Treat clean rooms as an early, scoped capability. See the caution on the concepts page.
Across the deployment matrix
- Single-cell and multi-cell: a recipient reads a filtered snapshot from the serving cell; grants and revokes hold across cells.
- Sovereign / air-gapped: sharing runs from the cell over the open protocols; nothing depends on a vendor control plane.
- Cloud-enabled: the same protocols reach cloud consumers and tools.