A fintech client needed to handle 10,000 concurrent users for a Friday product launch — on a $400/month infrastructure budget. We hit the target with room to spare. Here's exactly how, with the trade-offs we accepted.
Edge caching does 70% of the work
Before optimizing anything, we audited which routes were truly user-specific. Turns out only the dashboard and checkout flow were. Every marketing page, pricing card, and product image got pushed to Cloudflare's edge with aggressive cache headers. Origin traffic dropped 92% overnight.
Queues for everything async
Sign-up emails, webhook deliveries, analytics events, PDF generation — all of it moved to a queue with a single worker process. Latency on the request path collapsed from 800ms to 40ms. The queue handled bursts the database never had to see.
One read replica, used correctly
We added a single Postgres read replica and routed every analytics query, every dashboard widget, and every list view through it. Writes stayed on primary. The replica cost $60/month and absorbed 80% of database load.
What we deliberately didn't do
No Kubernetes. No service mesh. No multi-region active-active. These add complexity that small teams cannot operate. Boring infrastructure run well beats clever infrastructure run poorly — every single time.
