Production / Safe correlation

Logging

Log enough to reproduce a failure without turning your log store into a second copy of user content or credentials.

Inspector

Inspect what your event actually stores

Toggle fields to see the event you would persist. The goal is an event that makes failures reproducible while keeping credentials and user content out of your log store entirely.

Operation namekeepLow-cardinality and always useful.
Status and durationkeepThe core of any latency or error investigation.
Correlation identifierskeepTurns a user report into a specific request.
Hashed end-user referencekeepGroups events without storing the identity itself.
API key or Authorization headerneverA credential in logs is a security incident.
Raw query textneverUser content, frequently personal.
Returned memory contentneverDuplicates user data into a second store.

Event preview

{
  "operation": "memory.query",
  "status": 200,
  "duration_ms": 412,
  "request_id": "<X-Request-ID response header>",
  "end_user_token": "hmac(user-123)",
}

Safe to ship. Reproducible without storing credentials or user content.

Policy

Field policy

Allow-list: Serialize an explicit set of fields. Anything not on the list never reaches the logger.
Hash identities: Store a keyed hash of the end-user reference so events group without carrying the identity.
Describe, don’t copy: Record length, category, or a content hash instead of the text itself.
Test the serializer: Keep fixtures for nested and malformed payloads so redaction cannot regress silently.

Governance

Retention and access

Logs inherit the sensitivity of what they contain. Give them an owner, a retention period that matches your data classification, restricted access, and verbose-debug settings that expire on their own rather than living in production forever.

Was this page helpful?

← Previous

Monitoring & Observability

Incident Response