Personal Data and GDPR Erasure
An event store keeps everything by design. A request under GDPR Article 17, the right to erasure, asks for the opposite. The usual answer in an event-sourced system is crypto-shredding:
- personal data is encrypted with a key that belongs to one data subject, before it is stored;
- erasing that person means deleting their key;
- the ciphertext can stay where it is, because nothing can read it any more.
Events stay immutable and streams stay intact. The data is unreadable everywhere a copy of it went, including projections, messages and backups, without anything having to find those copies.
Tayra is the Critter Stack's answer
Tayra is the officially sanctioned crypto-shredding solution for the Critter Stack, from Radarleaf Technologies and Babu Annamalai, a Critter Stack co-maintainer. It provides:
- field-level encryption of personal data, marked declaratively on your types;
- HMAC blind indexes, so encrypted fields can still be queried by exact value;
- GDPR Article 17 erasure by deleting a data subject's key;
- pluggable key stores, so keys live somewhere other than beside the data they protect.
Its documentation is at docs.tayra.dev.
WARNING
Today Tayra's Critter Stack integration covers Marten and Wolverine, not Fisher.
- Tayra.Marten covers documents and events, inline, live and async projections, and multi-tenancy.
- Tayra.Wolverine covers message bodies, the outbox, the inbox and dead letters.
So a Wolverine application on Fisher gets Tayra's coverage for its messages. Fisher's own documents and events have no Tayra integration yet. Check docs.tayra.dev for the current list before planning around it.
Fisher does not ship a crypto-shredding mechanism of its own, and will not. Tayra is the answer, and a second, Fisher-only mechanism would be one that application code could not carry to Marten or Polecat.
What Fisher has built in, and why none of it is erasure
Fisher has several operations that remove or hide event data. Each answers a different question, and none of them reaches every copy of the data.
| Operation | What it does | What it does not reach |
|---|---|---|
| Archiving | Marks the stream no longer live | Anything. The events are all still there. |
| Stream compacting | Replaces the history with one Compacted<T> event | The snapshot itself, which carries the aggregate's state and so any personal data in it. Also anything an Archiver copied out first. |
| Event data masking | Rewrites protected fields in stored event bodies | Projections, snapshots and flat tables already built from the old bodies, and the aggregate write cache |
TombstoneStream | Hard-deletes the stream and its events | Projected documents built from those events |
Two more gaps apply to every row in the table:
- Backups. A copy of the database file is a copy of the data. Nothing done to the live file reaches it.
- The file itself. SQLite does not necessarily overwrite a page when a row on it is deleted or rewritten. The old content can stay in the file's free pages, and in the write-ahead log, until that space is reused. SQLite's
secure_deletePRAGMA andVACUUMreduce this. Neither one reaches a backup.
Crypto-shredding avoids both gaps, because it removes the key rather than chasing every copy of the data.
If you mask instead
Event data masking is still the right tool for a narrower job: removing information you should never have stored, from the live file, as one atomic batch. If you use it as part of an erasure process:
- rebuild every projection that read the masked events, because masking leaves derived state alone;
- leave the affected aggregate types out of the aggregate write cache, because a node's cached baseline was built from the unmasked bodies;
- decide what happens to backups separately, since masking cannot reach them.

JasperFx provides formal support for Fisher and other Critter Stack libraries. Please check our