An e-commerce site will sell 5,000 units of a phone in a flash sale with 2 million users waiting. How do you prevent overselling and keep the site up?
Short answer
Protect the system with a virtual waiting room and edge caching, keep stock correct with atomic conditional decrements (or pre-split Redis counters reconciled to the database), give reservations a short TTL, and process orders asynchronously behind a queue with idempotency.
Detailed answer
- Admission control: a virtual waiting room at the edge (CDN-based queueing) admits users at a rate the backend can handle; bot protection and per-account purchase limits.
- Static content at the edge: product page cached; stock badge updated every few seconds.
- Atomic stock reservation:
UPDATE stock SET reserved = reserved + 1 WHERE sku = ? AND on_hand - reserved >= 1or DynamoDB conditional writes. For extreme contention, split 5,000 units into e.g. 50 Redis counters of 100, decrement atomically (DECRwith check via Lua), and persist reservations to the database asynchronously; the database remains source of truth for orders. - Reservation TTL (e.g. 10 minutes) released if payment does not complete.
- Order intake queue: accept "buy" requests into a queue, process at a controlled rate, return status via polling/push.
- Idempotency keys to avoid double orders on retries.
- Pre-scaling and load tests at 2–3× expected peak.
Architecture diagram
flowchart LR U[2M users] --> WR[Waiting room + bot protection] WR -->|admitted rate| API[Checkout API] API --> RC[(Redis stock buckets<br/>atomic decrement)] API --> Q[[Order queue]] Q --> W[Order workers] --> DB[(Orders DB)] W --> PAY[Payment] W -- timeout/decline --> RC
Real-world example
Ticketing platforms and sneaker drops commonly use queue-based waiting rooms (for example Queue-it, Cloudflare Waiting Room) combined with atomic inventory counters.
Follow-up questions
- What if Redis fails mid-sale? (Fall back to DB conditional updates at lower rate; Redis with replication; reconcile.)
- How do you handle payment failures? (Release reservation; next user in queue gets a chance.)
What the interviewer expects
Both protection of the platform and correctness of stock, with clear atomicity mechanisms and failure handling.
Common mistakes to avoid
Read-then-write stock checks (race conditions), distributed locks per request, synchronous calls to slow payment providers inside the inventory transaction.