Loading…
Case study · Retail
A top-100 D2C retailer's Magento monolith was at the limit of what could be patched into Black Friday survival. We replatformed onto a headless commerce architecture in 7 months with a peak-season-tested cutover, then walked the operations team through a Black Friday that delivered 8x peak-traffic capacity with no degradation.
The challenge
The customer's Magento 2 estate had been patched, partially extended, and operationally stretched across four years. Page-load times were degrading; peak-season scaling had become a quarterly engineering sprint; and the third-party extensions in production had divergent fates as Magento's roadmap evolved. The retailer's BFCM week represented 18% of annual revenue — an outage during peak was an existential question for the engineering org.
Our approach
We replatformed onto a headless architecture: Next.js storefront on Vercel, commercetools for the commerce platform, Algolia for product search, Stripe for payments, and a custom read-model layer on Postgres + Redis that decoupled storefront reads from the commerce platform's transactional load.
The read-model layer was the architectural bet. Storefront pages render from a denormalized Postgres + Redis cache that's continuously updated from commercetools events. Commerce platform load is bounded by writes (carts, checkouts) — read traffic doesn't touch it. This is what made 8x peak capacity defensible.
Six weeks before BFCM, we ran chaos-engineering exercises against the staging environment scaled to peak load: random pod kills, AZ failover, CDN edge degradation, payment provider latency injection, third-party extension timeouts. Every failure mode the post-mortem corpus documented got rehearsed.
The solution
Phased cutover by category and traffic source, with documented rollback per phase. New architecture in front of CDN edge cache (Vercel + Cloudflare); legacy Magento served as fallback during cutover windows. Twelve cutover phases over 6 weeks; each gated on conversion-rate parity vs the legacy baseline.
Results
BFCM week shipped on the new platform with zero P0 / P1 incidents. Peak traffic was 8x the prior year; conversion rate held parity through peak; median page load fell 47%. Engineering on-call received zero pages during the peak Saturday window — the war room sat in calm tension and watched the dashboards do their job.
Stack
Talk to us
A senior engineer plus the relevant department lead joins the first call. No discovery gauntlet, no junior reps.