eCommerce scalability & performance engineering

Use a layered, stateless architecture: CDN and edge caching for static assets, API gateway in front of horizontally scalable app services, aggressive caching for product/catalog data, and asynchronous queues for non‑critical workflows (emails, inventory sync). Combine auto‑scaling, rate limiting, and pre‑event load tests that simulate 10–20x traffic to verify your bottlenecks before Black Friday.

Microservices are not mandatory; a well‑tuned modular monolith with caching, a strong database, and horizontal scaling can handle significant load. Microservices help when you need independent scaling, rapid team autonomy, and resilience per domain, but they add operational complexity, so many teams evolve from monolith → “modulith” → selectively extracted services.

For DB performance, design good schemas, add the right composite indexes on common filters and joins, and avoid N+1 patterns via careful query design. Offload reads with caching and read replicas, keep writes on the primary, and regularly profile slow queries, dropping unused indexes to reduce overhead.

Key KPIs: TTFB (aim <300 ms for key pages), full page load for PDP/PLP, checkout API latency (often <500–800 ms end‑to‑end), error rate (<1% under expected peak), and throughput (orders/second at target latency). Also track resource metrics (CPU, DB connections) but treat user‑centric timings as your primary success criteria.

Set auto‑scaling on meaningful metrics (CPU, request rate, queue depth) with conservative minimum capacity, fast scale‑out, and slower, stepwise scale‑in. Use instance right‑sizing, spot/preemptible capacity for non‑critical workers, scheduled or predictive scaling around known events, and per‑service limits to avoid runaway cost.​

Common mistakes: “lift‑and‑shift” without redesigning for statelessness or autoscaling, underestimating data migration and network latency, ignoring security/compliance, and failing to load‑test the new setup. Others include over‑provisioning for worst case (wasting money) or misconfigured scaling/observability, leaving teams blind during real traffic spikes.