Skip to content

Audit

Every decision the ZStrike Policy Decision Point makes — allow or deny — is captured as an audit event. The audit service ingests these events, streams them live to the Console, and can archive them durably to per-organization object storage.

Each event records the full context of one authorization decision:

Field Description
principal The Cedar entity that made the request, e.g. App::User::"alice"
action The action that was attempted, e.g. App::Action::"read"
resource The resource it targeted, e.g. App::Doc::"contract-42"
decision Allow or Deny
policy The Cedar policy ID(s) that determined the decision
timestamp When the decision was made (ISO 8601)
project The project the decision belongs to
orgId The organization the project belongs to
eventId A unique ID for the event
context The request context as sent (including any inline resource entities), when present

Events are validated on ingestion — principal, action, and resource must be well-formed Cedar entity references — so your audit trail only ever contains structurally valid records.

Audit events are streamed in real time to the Console’s audit views, scoped per project, so you see decisions as they happen: who asked, what they asked for, and which policy answered. Only events that have been durably written are streamed.

Historical decisions are queryable through the Control Plane API and the Console’s Audit Logs view:

  • GET /audit/events: filter by principal, action, resource, decision (allow/deny), policy, and time range (start_date, end_date). Page with pageSize (max 100) and offset, and sort with sortBy and sortOrder.
  • GET /audit/summary: aggregate counts for the same filters, as totals (total/allow/deny) plus breakdowns byPrincipal, byAction, and byPolicy, and lastActivityAt. start_date/end_date are full ISO timestamps, so “the last 24 hours” works as expected.

Both endpoints need an app token with the audit: read grant.

When archival is enabled for your organization, events are additionally written as immutable Parquet files to organization-scoped object storage, partitioned by date. Batches are written idempotently — redelivered batches overwrite their own keys rather than creating duplicates — so the archive stays clean even across retries.

Parquet is an open columnar format, so archived audit data is queryable with standard tooling (DuckDB, Spark, Athena, and similar) in addition to the platform’s own APIs.

  • Ingestion is synchronous: an event is only acknowledged after it has been validated (and stored, when archival is enabled). Producers redeliver on failure.
  • Batch redelivery is safe: batches carry an ID and overwrite themselves on retry.
  • Ingestion is authenticated: producers present short-lived, environment-scoped service credentials, so a token from one environment can never write into another.

The audit pipeline is source-agnostic by design. Today the Policy Decision Point is the primary producer, but the event contract is not tied to it — additional event sources plug into the same validated, streamed, and archived pipeline.