Skip to main content

Identity Resolution and Privacy Boundaries

Identity resolution is a governed mapping problem, not a license to identify everyone. The same person can have many devices and accounts; one device or account can represent many people. Deterministic authenticated evidence can reduce uncertainty within a declared context, but it does not prove personhood or authorize cross-context linkage.

Prerequisites​

Complete instrumentation validation and data quality. Review the actor and entity grains from the second lesson.

Learning goals​

  • Use a dated, versioned identity snapshot instead of mutable current links.
  • Preserve unresolved and possibly shared identities as uncertainty.
  • Make purpose, consent, retention, deletion, and access boundaries executable.
  • Explain why identity resolution changes historical metrics.

An explicit identity policy​

The fixture's snapshot is dated 2026-10-02 and contains only anon-a → u1 and anon-d → u2. Its method is “deterministic authenticated link only.” Even though another event carries both an anonymous and account identifier, the teaching snapshot does not automatically authorize historical stitching. A producer field can be wrong, imported, or outside the approved linkage process.

For each event, the worked metrics use the observed user_id when present, then the approved snapshot link, otherwise the anonymous ID. This is a measurement convention, not a claim that unresolved IDs are distinct people. Results must carry the snapshot date because adding a later link can merge actors and revise historical counts.

Privacy controls in the data flow​

  • Purpose: product_measurement and essential_operations are separate. Data collected for one is not silently reused for the other.
  • Consent or other authorization: the example requires analytics consent for product measurement and quarantines a denied event. Real legal bases vary by jurisdiction and purpose; consult qualified privacy/legal owners.
  • Data minimization: the event contains no name, email, message content, precise location, or advertising identifier.
  • Retention: raw measurement events need a documented expiry; aggregate retention must also be justified. Quarantine should have a short incident-handling retention, not become a shadow warehouse.
  • Deletion and access: both anonymous_id and user_id are indexed deletion keys so requests can reach pre-login and authenticated records. Identity maps, raw events, derived tables, caches, and backups need documented propagation and verification.
  • Uncertainty: shared devices, account sharing, merges, offline events, and deleted links remain visible limitations.

Deleting a mapping while retaining linkable raw identifiers may not satisfy the intended deletion outcome. Conversely, deleting raw evidence can make an old aggregate irreproducible. The policy must declare whether aggregates are recomputed, tombstoned, or retained under a justified rule and must not promise irreversible anonymization without evidence.

Worked example​

import json
from pathlib import Path

root = Path("tests/fixtures/product-analytics")
plan = json.loads((root / "tracking-plan-v1.json").read_text())
rows = [json.loads(line) for line in (root / "events-v1.jsonl").read_text().splitlines()]
links = plan["identity_snapshot"]["links"]
eligible_anon = {row["anonymous_id"] for row in rows if row["analytics_consent"]}
unresolved = sorted(identifier for identifier in eligible_anon if identifier not in links)
denied = sorted({row["anonymous_id"] for row in rows if not row["analytics_consent"]})
print(f"known_links={links} unresolved={unresolved}")
print(f"analytics_denied={denied} deletion_keys={['anonymous_id', 'user_id']} infer_household=False")

Expected output:

known_links={'anon-a': 'u1', 'anon-d': 'u2'} unresolved=['anon-b', 'anon-e']
analytics_denied=['anon-c'] deletion_keys=['anonymous_id', 'user_id'] infer_household=False

The unresolved list is not an error to “fix” with probabilistic matching. It is evidence that the available, authorized identity model has limits.

Limitations and common mistakes​

  • Email equality is not automatically a durable or authorized person-level key.
  • Probabilistic similarity can create false merges that expose one person's behavior to another.
  • Backfilling links changes denominators and must carry snapshot/version evidence.
  • Consent is purpose-specific and revocable; one boolean is a simplified teaching boundary, not a complete consent-management system.
  • Hashing stable identifiers does not by itself anonymize linkable event histories.
  • Retention and deletion promises require tested propagation, exception handling, and accountable owners.

Practice​

Write a deletion runbook for u1: list identity-map, raw-event, quarantine, derived-metric, cache, and backup handling; define completion evidence and exceptions. Then explain how deleting anon-a history changes the two worked metric values.

Next steps​

This measurement foundation is complete. Ticket 62 begins with funnel analysis, reusing the frozen event schema, identity snapshot, UTC boundary, duplicate rule, consent filter, and late-data policy.

Primary sources​

Checked 2026-10-19: W3C, Privacy Principles and Data Privacy Vocabulary; European Union, Regulation (EU) 2016/679 (GDPR), Articles 5, 15, 17, and 25; IETF, RFC 6973: Privacy Considerations for Internet Protocols.