Authorization Policy (Cedar)

Spice.ai Enterprise enforces fine-grained authorization using Cedar, the open-source policy language developed by AWS. Policies decide, on every request, whether an authenticated principal may perform a given action on a Spice resource.

Authorization Model

Policy evaluation is the standard Cedar (principal, action, resource, context) decision. Spice.ai supplies an embedded Cedar schema that defines the entity types, actions, and attributes available in policies.

Entity types

Actions

The Cedar schema is fixed by the runtime; policies reference these types and actions directly.

Configuration

Policies are configured under runtime.authorization in spicepod.yaml. Authentication must be configured separately under runtime.auth — see Authentication.

runtime:
  auth:
    oidc:
      issuer_url: https://auth.example.com/
      audience:
        - spice-runtime

  authorization:
    enabled: true              # default: true
    default: allow             # allow | deny  — decision when no policy matches
    provider: local            # local | operator | cloud
    policies:
      - name: analysts-read-only
        cedar: |
          permit(
            principal in Spice::Role::"analyst",
            action == Spice::Action::"query",
            resource
          );
      - name: block-pii
        path: ./policies/block-pii.cedar

runtime.authorization fields

PolicyDefinition

Policy Examples

Default-deny baseline

runtime:
  authorization:
    default: deny
    policies:
      - name: admins-everything
        cedar: |
          permit(
            principal in Spice::Role::"admin",
            action,
            resource
          );

Read-only analysts

permit(
  principal in Spice::Role::"analyst",
  action == Spice::Action::"query",
  resource
);

forbid(
  principal in Spice::Role::"analyst",
  action in [
    Spice::Action::"insert",
    Spice::Action::"update",
    Spice::Action::"delete",
    Spice::Action::"ddl"
  ],
  resource
);

Restrict a dataset to a single role

forbid(
  principal,
  action,
  resource == Spice::Dataset::"sales.public.orders"
)
unless { principal in Spice::Role::"sales-ops" };

Block a PII schema from non-privileged roles

To mask individual columns rather than block access entirely, use column masks.

forbid(
  principal,
  action == Spice::Action::"query",
  resource
)
when {
  resource.catalog == "customers" &&
  resource.schema == "pii" &&
  !(principal in Spice::Role::"compliance")
};

Limit model invocation to a paid tier

permit(
  principal in Spice::Role::"premium",
  action == Spice::Action::"invoke",
  resource == Spice::Model::"gpt-4o"
);

Endpoint-level access (e.g. lock down /v1/sql)

forbid(
  principal,
  action == Spice::Action::"access",
  resource == Spice::Endpoint::"sql"
)
unless { principal in Spice::Role::"engineer" };

Policy Providers

Policy reloads are atomic: in-flight requests complete against the previous policy set, and subsequent requests evaluate against the new set. Empty policy fetches do not silently disable enforcement — combined with default: deny, an empty set denies everything.

Row Filters and Column Masks

Beyond coarse allow/deny, policies enforce fine-grained access — row-level filtering and column-level masking — on the read action. The runtime compiles these from Cedar policy annotations into SQL that is applied to the table scan, so filtered rows and masked values are never materialized for the request: they are enforced before any downstream operator (or the client) observes them.

Annotations attach to a permit policy for Spice::Action::"read". A read permit also implicitly authorizes query, so a single read policy can grant masked, row-filtered access in one rule.

Annotations

Row filters and mask expressions are evaluated with the request's identity, so they can reference the identity SQL functions (current_user_id(), current_org_id(), current_user_has_role('...')) for per-user or per-tenant access.

Example

// Physicians: full access to patients in their own organization.
@id("physician_read")
@row_filter("org = current_org_id()")
permit(
  principal in Spice::Role::"physician",
  action == Spice::Action::"read",
  resource == Spice::Dataset::"patients"
);

// Analysts: same org scoping, but SSN and any PHI-tagged column are masked.
@id("analyst_read")
@row_filter("org = current_org_id()")
@mask_ssn("'XXX-XX-XXXX'")
@mask_tag_phi("'REDACTED'")
permit(
  principal in Spice::Role::"analyst",
  action == Spice::Action::"read",
  resource == Spice::Dataset::"patients"
);

With these policies, an analyst running SELECT * FROM patients sees only their organization's rows with ssn and any PHI-tagged column replaced, while a physician sees their organization's rows unmasked — from the same query.

Tagging columns

Tag-based masks (@mask_tag_* / @column_mask_tag) target columns by the dataset's column metadata:

datasets:
  - from: postgres:patients
    name: patients
    columns:
      - name: diagnosis
        metadata:
          tags:
            - phi

Rules

• Fine-grained annotations are honored only on permit policies. Attaching one to a forbid is a load-time error.
• A row filter must evaluate to a Boolean; a column mask expression must return the column's data type. Type mismatches fail closed — the query errors rather than returning unmasked data.
• If two matching policies define conflicting masks for the same column or tag, policy load fails.
• To deny access outright, use a forbid on read/query rather than a mask.

Distributed Cluster Behavior

In a SpicepodCluster, the scheduler is the source of truth for policy. Executors pull policy as part of the GetAppDefinition bootstrap RPC and re-evaluate when policy changes are pushed; the same Cedar decision applies regardless of which scheduler routes a request. Policy diagnostics (matched policy IDs, decision) are emitted as structured logs and surface in runtime.task_history.

Production Checklist

• runtime.authorization.enabled: true and default: deny for production deployments.
• Roles are sourced from a trusted identity provider (OIDC groups or API key tags) — see Authentication.
• Policy files are stored in Git and rolled out via GitOps (Argo CD / Flux) rather than ad-hoc edits on the cluster.
• An "admin escape hatch" permit rule exists and is restricted to a small Spice::Role::"admin" group with audited membership.
• Policy decision logs are forwarded to the SIEM together with authentication logs.
• Cedar policies are validated in CI with cedar validate against the embedded Spice schema before merge.

See also

• Authentication
• Security
• Distributed Query
• Authorization recipe — a runnable example covering multi-tenancy, row-level security, PII masking, and role-based access control
• Cedar policy language reference