SPCS deep dive

Containers inside Snowflake's boundary, secure by default

Isolation, ingress, egress, caller identity and tunnels, as of November 2026.

Kevin Keller · kevinkeller.org

Why it matters

Data never leaves the governance boundary

  • RBAC, masking and row access policies apply automatically
  • No separate copies of data, no separate access control
  • Snowflake patches the OS, drivers, controllers and registry
  • Nodes are replaced at least monthly with hardened images

Check this first

Since May 2026, any role may publish public endpoints

BIND SERVICE ENDPOINT is now granted to PUBLIC automatically. Revoke it and grant it only to the roles that should have it.

Isolation

The owner role is the trust boundary

  • Per customer: a dedicated cluster, isolated by security groups
  • Same owner role: services can call each other by default
  • Different owner roles: only with a granted service role
  • USAGE on a service doesn't grant its endpoints, and ACCOUNTADMIN can't own one

Ingress

Every public request is authenticated and hardened

  • TLS 1.2+, network policies, authenticated Snowflake users only
  • TRACE and CONNECT blocked, bearer tokens scrubbed
  • Security headers and a baseline CSP injected automatically
  • CORS handled by the proxy, explicit origins only
  • PATs go in the header as Snowflake Token="…", not Bearer

Egress

No External Access Integration, no egress

  • Network rules are signed and installed in managed egress proxies
  • Only ports 22, 80, 443 and 1024+ are allowed
  • network.egress.denied.packets shows what was blocked
  • Stable egress IP ranges rotate: automate your allowlist

Caller identity

Run queries as the user, within a ceiling you set

With executeAsCaller, a service can query as the calling user.

Caller grants are a ceiling: a compromised service can never exceed them.

The caller token defaults to 120 seconds. Not supported through gateways or service functions.

The session token

Short-lived, scoped and refreshed in place

  • Each service gets a dedicated service user
  • A token valid up to 1 hour sits at /snowflake/session/token
  • Works only inside your SPCS cluster, for one account, role and user
  • Re-read it periodically; skip mounting it where it isn't needed

Observability

Logs and metrics land in the event table

ToolUse
SYSTEM$GET_SERVICE_LOGSReal-time tail, 100 KB
SPCS_GET_LOGSOne service, full SQL
Event tablesProduction audit, alerts, metrics

Structured JSON logging fills severity and attribute columns.

Tunnelling

Reach infrastructure that isn't on PrivateLink

ApproachBest for
Data Connectivity ProxySupported default, deny-all-inbound networks
SSH tunnelFull control, reverse access
WebSocket tunnelData sovereignty, HTTPS-only
Tailscale (userspace)Existing tailnet
ngrokWebhooks, but bypasses ingress controls

Know the limits

Plan around these

  • Maintenance runs on weekend nights: single-instance services and jobs are disrupted
  • Dropping a service with FORCE deletes its volumes; snapshots last 7 days by default
  • Images can't be dropped individually: one repository per environment
  • Traffic between your services is plain TCP/HTTP: add TLS if policy requires it

The takeaway

One thing to do today: check who holds BIND SERVICE ENDPOINT.

Read the full article

Kevin Keller · kevinkeller.org