Skip to content
bucker

Subject requests: access and erasure for end users

How an organization answers a request from one of its own customers' customers: "what do you hold about me" (GDPR Art. 15 and 20) or "delete me" (Art. 17). One request drives every store, with a preview, a cancel window, a deadline and a signed receipt that names no one. This page is about end users inside your telemetry. The data of a person who has a Bucker account is a different request (/compliance/dsr/export and /compliance/dsr/erasure).

The life of a request

State What it means
DRAFT Filed with its reach report. The identifiers are sealed in the draft and nowhere else.
PENDING_WINDOW A person executed the erasure. The cancel window is open (72 hours unless you chose 0 to 168).
EXECUTING The window closed and the adapters are running, store by store.
VERIFIED Every store was read again and held nothing for the person. The signed receipt exists.
CANCELLED A person cancelled a draft, or an erasure inside its window. Nothing was executed.
COMPLETED (Access requests.) The export is built and downloadable.

Every request carries its Art. 12(3) deadline: one month from receipt, clamped to the last day of a shorter month. Receipt is the moment the draft was filed unless the draft says receivedAt earlier. A request within seven days of its deadline is dueSoon in the list.

Who may do what

Act Who Notes
File a draft, run a dry run, list Anyone holding dsr:prepare, including an agent Changes no data. At most 20 unsubmitted drafts per organization; an agent files at most 10 a day.
Execute an erasure (subject.erase) A person on their own session, org:admin Step-up. Above 10,000 rows: dual control.
Execute an access request (subject.export) A person on their own session, org:admin Builds the bundle and returns a signed link.
Cancel (subject_request.cancel) A person on their own session, org:admin A draft, or an erasure inside its window.
Read the receipt, download the export A person on their own session, org:admin The link is bound to the person it was minted for.

An agent, a service principal, an OAuth client and a delegated credential are refused on every human act before anything is read. There is no MCP tool for them.

Dual control above 10,000 rows

At or below 10,000 rows the person acts at once with a step-up, and the cancel window is the second look. Above it, the request goes through the human-requester path: the requester votes with a step-up, two approvers outside the requester's chain approve (or, in an organization with fewer eligible people, the quorum is lowered to the people who exist), any human may reject during the 24-hour hold, and the requester confirms with a fresh step-up. Erasure is a legal duty, so the shipped policy row is ASK with two approvers, never the registry's DENY default.

The person names the selection fingerprint and total they were shown. A different selection (the subject, or the projects covered, such as a project added since the draft) fails closed with 409, at execution and again at the apply after any hold; that needs a new draft. A different total for the same selection (the person kept using the product) is a 409 at execution that carries the current total (currentRows) to acknowledge, and is not re-checked at the apply: the purge erases whatever the selection holds when the window closes, so binding the count across a day's hold would make an active person impossible to erase without making any erasure narrower. The receipt records what was actually removed.

What a request reaches

The dry run and the draft list every store. A store is either reached (an adapter with reach, execute, verify and export) or listed and not reached with a reason from a closed list. The stores today:

Store Reached Removal
Session replays and transcripts yes Recording deleted before its row; both gone when the purge returns
Events, release sessions, spans, logs, the query log, other free-text rows yes Rows deleted at once; a partition is dropped later by retention (housekeeping, not the removal)
Metric series and points yes Series tombstoned, points deleted; another machine may write against it for up to a day
Feedback reports yes Deleted at once, by contact address
Cold archives and cold-tier objects yes Archive rewritten in place; cold object rewritten under a new key, the old shredded
Per-subject rollups and revenue rows yes Deleted at once
Sessions yes Row deleted at once; its partition is dropped by the session sweep
Subjects, links, device keys, property history, identity records yes One transaction
Status-page email subscribers yes Subscription removed
Warehouse sink objects no Listed external_processor: a row exported before the erasure stays in the sink's files until the sink is deleted

A subject is matched by the keyed pseudonym of every identifier in every pseudonym generation the project has opened (and the frozen v1), by the scoped hash events carry, and by the literal address, IP and identifiers in the stores that scan text. A pseudonym generation opened under a previous secret cannot be computed and is named in the report.

What hiding is today

Executing an erasure holds the matched rows: a hold is recorded for each project the request reaches (erasure_ledger, readable through heldProjectIds), and a store that can hide its own rows implements hide. No store has a per-row hide column yet, so the rows stay readable until the purge and the purge, not the hold, is what removes them. This is stated in the report instead of implied.

The receipt

A verified erasure writes a subject_erasure receipt, signed (DSSE over an in-toto statement, extension kind subject_erasure) and appended to the organization's receipt log inside the transaction that marks the request verified. Per store it says: rows reached, whether the verification pass found nothing left, when removal completed or when it is expected, and why a store was not reached. It carries no identifier: the evidence schema is closed, the selection is named by a keyed digest, and the receipt is scanned for every identifier before it is stored. Verify it offline with bucker-verify verify-log.

The access export

The bundle holds every store's rows for the subject, the stores that could not be read and why, and the recipients: everyone in the organization who read person-level data, replay or survey responses in the covered projects, by surface, count and first and last time (from the read ledger). The ledger records reads of a class, not of one person, so this is an upper bound and says so.

The bundle is encrypted in the object store (per-organization key, bound to the request). The download is a signed 15-minute link bound to the organization, the request and the person it was minted for.

REST

Route
POST /orgs/:o/subject-requests File a draft. Body: kind (ACCESS or ERASURE_END_USER), identifiers, optional projectSlug, receivedAt.
GET /orgs/:o/subject-requests List: states, deadlines, receipt summaries. No identifier.
GET /orgs/:o/subject-requests/:id One request and its reach report.
POST /orgs/:o/subject-requests/:id/execute The human act. Erasure: reason, acknowledgeSelection, acknowledgeRows, optional windowHours, stepUp. 202 while it waits.
POST /orgs/:o/subject-requests/:id/cancel Cancel inside the window.
GET /orgs/:o/subject-requests/:id/receipt The receipt.
GET /orgs/:o/subject-requests/:id/export A signed link; with ?token=, the bundle.

MCP

Tool
erasure_dry_run Counts per store and the selection fingerprint. Writes a data.read ledger row.
subject_request_draft_create Files a draft (governed: ELEVATED, ALLOW). A person executes it.
subject_requests_list States, deadlines and receipt summaries.

All three need dsr:prepare and are held until the OAuth security review closes (SG-059).

Adding a store

A package that stores an end user's identifier exports a SubjectStoreRuntime from domain/src/modules/compliance/adapters/<store>.ts and joins with registerSubjectStore(...) in its wiring line; a store that cannot be reached declares unreachable with a reason and the sink or processor's retention. The registry test fails a dataset that names neither.