▲ DROPZERO

Architecture — globally consistent drops that never oversell

Clients
🌐 Buyers worldwide
  • Limited drop, huge concurrent demand
  • 8 languages incl. RTL (Arabic)
→
Frontend · Vercel
Next.js 16 (v0-style UI)
  • App Router · real-time stock UI
  • API routes: /api/inventory, /api/purchase
  • Deployed on Vercel
→
Auth · no static secrets
Vercel OIDC → AWS IAM
  • OIDC federation identifies the project
  • Assume AWS IAM role (STS)
  • Generate DSQL auth token per connection
↓
Database · Amazon Aurora DSQL
Serverless distributed SQL · multi-Region · active-active · strongly consistent
Tokyo
ap-northeast-1
Active peer
⇄
Seoul
ap-northeast-2
Active peer
+
Osaka
ap-northeast-3
Witness (quorum)
Single logical database · two regional endpoints · synchronous commit quorum · stock row never goes negative
Data model · in Aurora DSQL
Deliberate, DSQL-native schema
products
  • id · uuid PK
  • name
  • drop_name
→
inventory
  • id · uuid PK
  • product_id
  • stock ← single row
every buyer races for this one row
→
orders
  • id · uuid PK
  • product_id
  • user_ref
  • status
  • region
No foreign keys · no sequences (DSQL) · UUID PKs · purchase = check + decrement in one transaction · OCC retry → never oversell
Never oversell — by design

Each purchase runs check + decrement in one transaction. Aurora DSQL uses optimistic concurrency control; conflicting commits are rejected (SQLSTATE 40001) and retried with exponential backoff. The loser re-reads the latest stock and returns “sold out”.

Proven under load (k6)

100 units · 3,000 concurrent purchases across the Tokyo & Seoul endpoints → exactly 100 confirmed, 2,900 sold-out, 0 errors, final stock 0. Zero oversell at scale.

Frontend: Vercel · Database: Amazon Aurora DSQL (APAC region set) · Auth: Vercel OIDC federation