Survived 14x traffic on launch day.
Consumer platform pre-launch. We modelled the scaling cliff, sharded the hot table, and stood up read-replicas before TechCrunch published.
What we walked into.
Consumer platform had a TechCrunch feature scheduled and a pre-launch following of 80k waitlist signups. The current architecture was a single-region Postgres deployment that had handled their beta of 2k DAU comfortably. And absolutely no further.
Six weeks to launch. Two backend engineers. The CEO wanted answers, not slideware.
How we shipped it.
Load model with real production traffic
Built a load test that replayed two weeks of production traffic at increasing multiples. Identified the exact request volume where p99 latency crossed 1s. The scaling cliff. Then designed for 4x that.
Hot-table sharding behind a read API
The `activity` table was 80% of read load. Sharded by user ID across four logical shards behind a read API. Application code unchanged.
Read-replicas with geo-routing
Stood up read replicas in us-west-2 for the western US user base, routed via Route 53 latency-based records. Cut p99 for half the user base in half.
Pre-warmed cache and CDN
Pre-warmed CloudFront for all hot URLs the morning of the TechCrunch publish. ElastiCache pre-loaded with predicted hot keys.
AWS services in this engagement
EC2 / ECS
RDS (sharded)
ALB
CloudFront
S3
LambdaWhat shipped.
TechCrunch traffic peaked at 14x baseline. Platform held at p99 <200ms throughout the launch week. No customer-facing incidents, no engineer woken up.
The architecture has since handled three further traffic spikes from press coverage without further engineering work.
Hardened 47 IAM roles in 11 days.
Series B fintech needed audit-ready IAM before SOC 2. We refactored every role into least-privilege Terraform modules.
PlatformCut cloud spend 38% in 6 weeks.
Multi-region platform was burning compute. We re-sized the fleet, moved to Graviton, and added a savings plan model that paid back in 21 days.
ArchitectureMigrated from Heroku to a cloud platform in 9 weeks.
B2B platform hit Heroku ceilings. We landed them on ECS with managed Postgres, zero downtime, and a runbook the team actually uses.
Tell us what you're trying to ship.
A 30-minute scoping call with the engineers who would do the work.