Architecture

Global cells & production path

DSQL owns exact scarce-right allocation inside each active-active cell. DynamoDB owns global fan-facing speed. CDC → Kinesis emits the transparency trail — not a correctness dependency.

Cell topology

                    [ Global geo router ]
                           |
        +------------------+------------------+
        |                  |                  |
   [ US Cell ]        [ EU Cell ]        [ APAC Cell ]
   DSQL A-A           DSQL A-A           DSQL A-A
   (Region set)       (Region set)       (Region set)
        |                  |                  |
   DynamoDB MREC      DynamoDB MREC      DynamoDB MREC
   DynamoDB MRSC      DynamoDB MRSC      DynamoDB MRSC

No cross-continent database cluster — global scale = multiple independent cells behind a router.

US Cell

us-east-1 + us-east-2 · witness us-west-2

Active-active DSQL within the US Region set. Owns allocation truth for US inventory slices.

EU Cell

eu-west-1 + eu-central-1 · witness eu-west-1

Independent cell — same kernel, separate cluster. No cross-continent DSQL.

APAC Cell

ap-southeast-1 + ap-southeast-2 · witness ap-southeast-2

Scale-out pattern: router picks cell by fan locale + inventory shard.

DynamoDB Global Tables — MRSC vs MREC

Mode	Use	Examples	Tradeoffs
MRSC	Global control plane	Event catalog, cell routing, cross-cell config	Strong global reads, zero RPO — no transactions, no TTL, higher write latency
MREC	Fan telemetry	Live activity feed, presence, metrics ring buffer	Async replication, single-digit-ms writes, last-writer-wins

Audit trail — DSQL CDC → Kinesis (production preview)

Aurora DSQL CDC streams committed row-level changes to Kinesis. Semantics: at-least-once, currently unordered — consumers dedupe on primary key + commit timestamp. Correctness proof comes from DSQL state + evidence capsules, not the stream.

DSQL (audit_log, confirmed_orders)
        │ CDC (preview)
        ▼
   Kinesis stream  ──► idempotent consumer ──► transparency / analytics
        │
   Evidence capsules (portable JSON) ──► dropzero-verifier (no AWS)

Vercel Edge

Geo routing, per-IP buy rate limits, bot flood defense before Node.

Aurora DSQL

Correctness core per cell — pools, holds, ledger, in-txn idempotency.

DynamoDB

Live SSE feed, metrics, non-authoritative hot path at single-digit-ms.

IaC

Terraform reference in infra/ — active-active DSQL cell + DynamoDB fan tables. Live demo cluster provisioned via Vercel Storage (submission/storage-config.png).

Thesis: proof-carrying scarcity inside each cell; global fan speed on DynamoDB; judges verify capsules locally with zero AWS credentials.