Skip to main content

The virtual graph

The virtual graph is a zero-copy property graph you query in GQL / openCypher (the ISO GQL query language), projected over the lakehouse and over warehouses you already run, copying nothing. It is not a new database: it is another access method of the same data engine, alongside SQL and the existing graph store.

A third access method, not a third store​

The sharp point is that the graph is a view of the engine, not a copy of the data. The same tables answer a SQL query, a native-graph traversal, or a GQL/openCypher query. Nothing is exported into a graph database to make the graph work.

*Figure C7: the virtual graph as a third access method. GQL/openCypher (green) sits beside SQL and the existing graph on one engine, and projects zero-copy over native lakehouse tables and external warehouses read in place (MS Fabric and Databricks (Delta), Snowflake (Iceberg)) over your own object store. Write-back is governed (create/update/delete, each with undo) through the same deny-by-default governance and metering as the reads. If a credential cannot be resolved the read is denied (red), never silently downgraded to a local or unauthenticated fetch.

Table equivalent

ElementRoleState (meaning)
SQLAn access method of the engineneutral
Existing graphAn access method of the engineneutral
Virtual graph (GQL/openCypher)A third access method: a projection, not a new storeverified (green)
Native lakehouse tablesProjected zero-copyneutral
MS Fabric / Databricks (Delta), Snowflake (Iceberg)External warehouses read in place over your object storeneutral
Governed write-back (via ontology Actions)create / update / delete on the source tables, each with undo: the projection itself stays read-onlyverified (green, governed + metered)
Credential unresolvedDenied, never downgraded to a local/unauthenticated fetchdeny (red)

Governed write-back, with undo​

Reads are virtual (a zero-copy projection over your lakehouse rows) and the projection itself is read-only: a direct write to the virtual graph is refused, never silently applied. Mutation goes through ontology Actions, which commit to the source tables under the same deny-by-default governance and metering the reads pass through: create, update, and delete, each with undo. A write is an authorized, metered, reversible action against the underlying data, not a side door that mutates the projection.

Inside a clean room​

Within a clean room, cross-party graph gluing is deny-by-default, and every denial is recorded. Joining two parties' graphs is a governed act under the room's mutual-distrust policy, not an automatic merge.

The honest ceiling: narrows, not matches​

This is the honest answer to the one place rivals concede ground. Palantir's ontology. A governed semantic layer over your own lakehouse, including the warehouses you already bought, with no copy and a governed write-back, narrows that gap. It does not close it, and the claim to make is "narrows," not "matches."

The edge, said precisely

The cell owns the resident lakehouse readers and projects the graph; the Edge Agent reaches them over the governed fabric, degrading fail-closed to scan-and-carry when no path exists. A full cell can be deployed at the edge for sovereign or air-gapped use, but the edge device itself does not run a lakehouse reader of its own.

Across the deployment matrix​

  • Single-cell, the graph projects over the cell's own lakehouse and any external warehouse it can reach.
  • Multi-cell: the cell that owns the readers projects the graph; peers reach it over the governed fabric.
  • Sovereign / air-gapped: a cell at the site projects the graph locally with no vendor control plane; an edge device reaches a cell's readers over the fabric and fails closed when it cannot.
  • Cloud-enabled, external warehouses (Fabric, Databricks, Snowflake) are read in place; a failed credential is denied, never downgraded.