↓ Skip to main content

The Data, the Key, or the Switch

A framework for customer-held control over the most sensitive data in a cloud data platform. By Kevin Keller, November 2026.

Every data sovereignty conversation ends at the same question:

“If we put this in your platform, what is still in our hands?”

For most data, native platform controls answer it: encryption at rest, customer-managed keys, role-based access, masking, private connectivity, regional deployment. But almost every regulated organisation has a small slice of data where the answer has to be stronger: national IDs, bank details, salaries, patient identifiers, or datasets a regulator says may not be stored in a foreign-operated cloud at all.

For that slice, the useful question is not “trust us” versus “don’t use the cloud”. It is which one thing the data owner keeps.

The Data

The most sensitive records never live in the platform. It reaches them only when you allow it, or never.

  • On-demand tunnel to your premises
  • Keep it separate and join locally
  • Synthetic data instead of the real thing

The Key

The data sits in the platform, encrypted. You hold the key and decide, query by query and person by person, when it is used.

  • External key broker with step-up
  • Approve on your phone, per query
  • Remote decrypt: the key never enters
  • Homomorphic encryption: never decrypt

The Switch

You can cut access instantly, with a time limit and a full audit trail.

  • Tunnels that exist only on demand
  • Time-limited step-up grants you can revoke
  • A named person, second factor and reason for every release

Key custody is not data custody
#

“The key stays with us” and “the data stays with us” are different requirements, and they lead to different patterns. Requirements documents often blur them. Decrypting upstream before loading also keeps the key out of the platform, but it puts plaintext in the platform from the first table on. Settle which requirement you actually have before choosing a pattern.

Two questions decide most cases
#

  1. May the sensitive data be stored in the platform at all, even encrypted? If not, use an on-demand tunnel when users need the platform experience, or keep the data separate when they don’t.
  2. If it may be stored encrypted, do people need to read values, or only compute on them? Arithmetic only points to homomorphic encryption. Reading values inside the platform points to a key broker or phone approval, or to remote decryption if the key may never be in the platform. Reading them in an on-premises application points to unlocking only there.
Decision guide

Layer them
#

The patterns combine. A realistic setup uses native controls for the bulk of the data, a key pattern for identifying columns, homomorphic encryption for values you only aggregate, and a data pattern for the few datasets that may not be stored in the cloud at all. In the benchmarks, the extra controls cost between a fraction of a second and a few seconds per query, and only where the risk is.

What none of this solves
#

  • No pattern stops an authorised person from copying what they are allowed to see. That stays a matter of roles, unload restrictions and monitoring.
  • Tunnel and key patterns control when and for whom plaintext exists in compute. Only keeping data separate, unlocking in your own application and homomorphic encryption keep plaintext out of the platform entirely.
  • Revocation is not instant everywhere: cached data keys live until they are renewed.

Go deeper
#

  • The full article: all eight patterns on Snowflake, with measured latency, the threat model and the trade-offs.
  • The code: runnable reference implementations, a white paper, a slide deck and a single worksheet that walks through every option.

Cite this: Keller, Kevin. "The Data, the Key, or the Switch: a framework for customer-held control." kevinkeller.org, November 2026. https://kevinkeller.org/data-key-switch/

Personal framework and "art of the possible" project. Not an official, reviewed or supported Snowflake solution. Review any pattern with your security and legal teams before adapting it.

Share this article