Case study · Retail & Commerce
Eleven weeks to a new storefront.
Atlas needed a faster storefront before peak, and had one quarter. We pulled the front end off the suite and left the back office alone.
- 11-week delivery
- 8x peak traffic absorbed
- 34% conversion uplift
- Sector
- Multi-brand retail
- Duration
- 11 weeks
- Squad
- 6 engineers, 2 designers
- Peak
- 8x baseline traffic
Overview
The situation
Atlas ran four brands on one commerce suite. Page loads averaged 4.6 seconds on mobile, and the vendor's next release — which included the performance work — was scheduled for the following March.
Peak trading was eleven weeks away and accounted for 31% of annual revenue.
Challenge
The challenge
A full replatform was impossible in the time available, and attempting one before peak would have been reckless. The suite handled merchandising, pricing, order management and finance integration adequately.
Only the customer-facing layer was genuinely failing.
- 4.6s median mobile load, against a 2.5s target
- Eleven weeks to the start of peak trading
- Four brands sharing one catalogue and pricing model
- No appetite for back-office disruption before peak
Approach
What we did
We took the narrowest possible scope: the storefront and the read path that feeds it. Everything else stayed exactly where it was.
01
Draw the boundary
Week one established what would not change. The suite kept merchandising, pricing, orders and finance; we took rendering, search and the catalogue read path.
02
Build the read model
A denormalised catalogue projection updated from suite events, so storefront rendering never waits on the suite's API — which was the actual source of the latency.
03
Storefront on a performance budget
A Next.js storefront with a budget enforced in CI. A pull request that pushed largest-contentful-paint over target failed the build, including marketing tag additions.
04
Model peak before it arrived
Load modelling from the previous two years' traffic, then game days at 8x baseline. Two failure modes surfaced — inventory contention and a payment provider timeout — and both were fixed with weeks to spare.
05
Migrate brand by brand
The smallest brand went first at 5% of traffic, ramping over nine days. The largest went last, six days before peak, with the previous storefront one flag away throughout.
Results
The outcome
All four brands were on the new storefront before peak. Traffic reached 8.2x baseline on the busiest hour with no degradation and no war room, and the conversion uplift held after the trading period ended.
Conversion uplift
Sustained post-peak
Median mobile load
Down from 4.6 seconds
Peak traffic absorbed
No degradation, no incident
Concept to full rollout
Across four brands
“The discipline was in what they refused to touch. Everyone else who pitched wanted to replace the whole suite. They took the twenty percent that was actually broken.”
Industries
Technologies
Published 20 November 2025
More work
Other systems we have shipped
Have a problem shaped like this one?
Bring us the constraint your last partner called impossible. We will tell you on the call whether we believe it.