Usage Metering & Billing Pipeline  ·  View 10 of 22  ·  Data

Storage Zones

Five zones, sorted by one question: if this store were deleted, could the platform rebuild it?

Editable source SVG draw.io All views
Data estate Immutable record — the only truth Event ledger S3 Iceberg, 400 d hot Audit tail Glacier Object Lock, 7 y Privileged-action log CloudTrail, locked Financial ledger — append-only, corrections are new entries Invoice ledger Aurora PostgreSQL Balance ledger grants, drawdown Credit memos and supersessions Versioned configuration — immutable per version Plan registry Aurora + S3 versions Currency rates dated records Schema catalogue Glue, all versions Derived — droppable and rebuildable by replay Aggregate store DynamoDB Flink checkpoints S3 Usage read model DynamoDB Transient — no recovery value Event stream Kinesis, 24 h Outbox EventBridge Entitlement cache DAX Quarantine S3, replayable Data warehouse rebuild lifecycle Storage Zones — Ownership and Rebuildability Data store Queue / topic External / third party batch The test for a zone is one question: if this store were deleted, could the platform reconstruct it? Only the top two zones answer no. Quarantined events replay into the ledger once the defect that rejected them is fixed. v 1.0 · owner Architecture Practice · date 2026-10

The test

  • Only the top two zones answer no. The immutable record cannot be rebuilt because it is the record; the financial ledger cannot be rebuilt because an issued invoice is a published fact, not a derivation.
  • Everything in the derived zone may be dropped and rebuilt by replay, which is why a corrupt aggregate, a bad deployment and an operator error all have the same remedy.
  • Versioned configuration sits in its own zone because it is the one small, read-heavy store whose corruption is applied uniformly to every affected account at once.

Retention, as assumed

  • Raw events queryable for 400 days, then cold for 7 years (stated assumptions) — the first covers a full annual cycle plus the 60-day dispute window, the second statutory audit.
  • Event stream retained 24 hours: it is transport, and the ledger behind it is the durable copy.
  • Entitlement cache has no recovery value by design; a cold cache costs latency, not correctness.

Risks

  • Cold-tier retrieval cost and latency are a constraint on the dispute process, not a storage detail. A dispute outside 400 days is answerable but slow and chargeable.
  • Object Lock on the audit tail means a retention mistake is not reversible. That is the point, and it needs to be a deliberate decision rather than a default.