The Problem
A fast-growing SaaS platform hit a hard wall at 800,000 concurrent users. Their PostgreSQL monolith — sharded manually across 3 regions — was exhibiting write amplification, cross-shard join latency, and replication lag that breached SLA thresholds during APAC business hours.
The ops team was spending 40% of their sprint capacity on incident response and manual database maintenance, with no clear path to 10× growth.
The Solution
We re-architected the data layer with a multi-tier distributed strategy:
- DynamoDB Global Tables for low-latency, strongly consistent entity storage with automatic multi-region replication
- ElastiCache (Redis Cluster) for session data and hot read paths — reducing database read load by 73%
- Amazon Aurora Serverless v2 retained for complex relational reporting queries, auto-pausing during off-peak hours
- Kubernetes (EKS) replaced the legacy EC2 fleet — HPA-driven autoscaling with custom KEDA metrics tied to queue depth
- Istio service mesh for mTLS between services, distributed tracing, and intelligent traffic management
The Result
The migration was completed in 10 weeks using feature flags and progressive traffic shifting:
- Scaled from 800K to 5M+ concurrent users with no further architecture changes
- 99.99% uptime maintained through 3 major traffic spikes, including a viral product launch
- Sub-40ms P99 latency globally (down from 380ms peak)
- 60% reduction in infrastructure costs through serverless auto-pause and right-sizing
- Ops incident response time dropped from 45 minutes to under 5 minutes