Privacy Law Keeps Blaming the Policy. The Bug Is in the Schema.

Archi4 min read

privacylawdata

Contents
  1. The violation was designed in a schema review
  2. Consent without a purpose field
  3. Retention is a cron job or it is a promise
  4. The join key is the whole argument
  5. What regulators are starting to ask for
  6. Why this keeps being framed as a legal problem

The violation was designed in a schema review#

Most privacy enforcement actions read like failures of judgment. A company collected too much, kept it too long, or shared it with someone it should not have. Regulators write it up as a governance problem, and the company responds with a new policy, a training module, and a revised notice.

The root cause is usually duller and more technical. The system that stored the data had no way to record why it was collected, who agreed to what, or when it should be deleted. Those facts were never in the schema. Everything downstream, including the privacy notice, became a description of behavior the code could not enforce.

My view: until privacy obligations show up as columns, constraints, and scheduled jobs, privacy law will keep regulating intentions while breaches keep coming from structure.

A consent record that stores a user ID, a timestamp, and a boolean tells you almost nothing. It cannot answer the question regulators actually ask: consent to what, for which use, for how long?

Many systems built between 2015 and 2021 have a single consent_given flag on the user table. Marketing, analytics, and product personalization all read that flag. When a user withdraws consent for marketing, the flag flips to false and everything stops, including functions the user wanted. Or the team adds a second flag, then a third, and the meaning of each drifts until nobody can trace which service checks which one.

The fix is not a policy update. It is a normalized consent table with a purpose code, a scope, a grant time, an expiry, and a link to the specific notice version shown. Without that, withdrawal is a guess.

Retention is a cron job or it is a promise#

Privacy notices say data is kept "as long as necessary." Databases rarely know what necessary means. In most production systems I have seen described publicly, deletion happens when someone remembers, when storage costs bite, or never.

Retention has to be an attribute of the record, not a paragraph in a document. A retain_until column plus a scheduled purge is unglamorous and testable. Auditors can query it. Regulators can sample it. A legal team cannot.

The same logic applies to deletion requests. If a user's rows are spread across twelve services, three caches, and a search index, honoring erasure means each of those stores has a documented deletion path. Companies that treat erasure as a ticket queue discover during an incident that two backups and a analytics warehouse were never in scope.

The join key is the whole argument#

Purpose limitation sounds abstract until you look at how tables are linked. If a single customer ID ties billing, support transcripts, and browsing events into one row, then every internal tool can see everything. The limitation exists only in the acceptable use policy.

Splitting identifiers by purpose, or at least restricting which services can join on which key, is an architectural decision. It is also the decision most likely to be skipped because it complicates dashboards. The teams that do it tend to have fewer findings, not because they are more ethical, but because the data needed for a violation is harder to assemble.

What regulators are starting to ask for#

Enforcement letters increasingly request records of processing, data maps, and evidence of deletion. Those requests are awkward for companies whose only source of truth is a Word document last edited eighteen months ago.

The practical response is to treat the data catalog as a regulated artifact. Every table that holds personal data should carry an owner, a purpose, a retention rule, and a lawful basis. That is four fields. It is not a platform project. It is a migration and a habit.

Lawyers write the notices. Engineers build the systems. The gap between them is where violations live, and it persists because the two groups measure success differently. Legal counts completed assessments. Engineering counts shipped features. Nobody counts how many tables have a purpose column populated.

My opinion is that privacy programs should be judged on schema coverage the way security programs are judged on patching. The number that matters is not how many policies were approved. It is what percentage of personal data tables have enforceable retention and purpose metadata attached.

That number is usually low. It is also fixable without new legislation, which is why it is a better target than another round of notice revisions. The law will keep asking who consented. The database will keep answering with a boolean. Until someone changes the schema, the argument repeats.

More articles