06-reference/research

snowflake semantic view enforcement contract vs convenience

2026-11-14·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
semantic-layersnowflakegovernancerbacagents

Nothing in the semantic view enforces its metric on a bypassing consumer: base-table grants are the only fence, and Snowflake's own agent path requires those grants

The question

"What actually enforces a Snowflake semantic view's metric definition for a third-party BI or agent consumer that does not query through the semantic layer?"

Context: open follow-up from [[2026-10-02-snowflake-semantic-views-cortex-gateway-agent-identity-audit]], and the load-bearing question under the founder's draft "Agents don't need your data warehouse. They need your definitions." ([[2026-10-02-agents-need-your-definitions-v1]]). A definition is a governance control only if something stops a consumer from recomputing the metric off base tables.

Headline: For a consumer that does not query through the semantic view, nothing enforces the definition. The only fence is ordinary role-based access control (RBAC) on the base tables. Semantic views do run with owner's rights, so you can build a role that sees only the semantic view (the "secure view" pattern). But Snowflake's own docs say Cortex Agents need SELECT on both the semantic view and its underlying tables, and the Cortex Analyst access table lists SELECT on "the tables mentioned in the semantic model." So the agent path Snowflake sells cannot be sealed this way. Even a sealed role can build its own ad-hoc aggregates from public facts and dimensions, and dimensions cannot be made private. Verdict: by default a semantic view is a convenience. It is a partial contract only for SQL and BI roles that are denied the base tables, and even then it controls vocabulary, joins and row policy, not the arithmetic.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The honest verdict is the negative one: a Snowflake semantic view is a convenience that can be partly hardened into a contract, and the hardening is unavailable on the agent path. The single fact the backlog expected to decide this, whether a role can get the view without the base tables, resolves "yes" for SQL and BI roles. Owner's rights work like a standard view. But Snowflake's own docs then exempt the case the thesis is about. Cortex Agents and Cortex Analyst need the executing role to hold SELECT on the base tables. Any principal who can ask the agent can also run SELECT ... FROM base_table and compute readmission rate their own way. Even a sealed BI role can build ad-hoc aggregates from public facts and dimensions inside the view. PRIVATE cannot hide dimensions. The exports, Ossie YAML and Tableau TDS, are previews that carry text, not policy. The access modifier rides in an extension that consuming tools are told they may ignore.

This changes the draft's argument rather than killing it, and the change makes it sharper. "They need your definitions" survives as a claim about consistency of intent. One written definition that agents and dashboards default to beats one definition per prompt, and the readmission example still proves that. It does not survive as a claim about control. The accurate three-layer version, each layer a GA primitive: (1) the semantic view publishes the definition and constrains join paths. (2) Base-table grants plus row and masking policies fence who can bypass it and what they see. Today that fence cannot exclude agent roles from base tables. (3) ACCESS_HISTORY with agents_info detects bypass after the fact, by showing which base columns a principal read directly versus through the view. So for the agent consumer, governance of meaning on Snowflake today is audit, not prevention. A post that says so plainly is more credible to a platform team than one that implies the semantic view is a fence. It also answers the bear case the draft already concedes.

For phData and CAF delivery, this converts into a concrete design pattern and an honest caveat. Pattern: issue BI and service consumers semantic-view-only roles, with no base-table grants. Mark component facts PRIVATE where a raw numerator or denominator should not be re-aggregated. Run a recurring ACCESS_HISTORY query that flags any principal reading the base columns of a governed metric outside the view. Caveat to state up front: agent users keep base-table access by design, so the governed-metric guarantee for agents is "default plus audit," and the client should know that before treating a semantic view as a compliance control for a regulated metric. That caveat is a differentiator: competitor content that sells semantic views as "the single source of truth" does not mention it.

One inference flagged as unverified. Wrapping the agent's data access in an owner's-rights stored procedure or custom tool, rather than the built-in Cortex Analyst tool, might restore a sealed agent path. No doc reviewed here states this. It is a follow-up, not a claim.

Why this is in the vault

This decides how [[2026-10-02-agents-need-your-definitions-v1]] (a founder-pinned piece under [[2026-09-29-nov30-credibility-targets]]) may describe semantic views. It can say "publish," not "enforce," and the agent-path caveat must sit beside the owner's-rights fact. It also gives phData semantic-layer engagements a specific seal-plus-audit design pattern, with a documented limit to disclose to regulated (healthcare) clients.

Open follow-ups

Related

Sources

Vault:

Web (all fetched 2026-11-14):