Atlas Commerce

Headless storefront migration

The brief

Re-platforming a $40m DTC storefront from a monolith to a composable stack — without a dip in revenue during the cutover week.

The challenge

Atlas Commerce had built a forty-million-dollar direct-to-consumer business on a monolith that had, over six years, become a liability everyone was afraid to touch. Every deploy took nineteen minutes; every merchandising change required a developer; the product team had learned to work around the platform rather than through it. The board had signed off on a re-platform in principle two years earlier and the project had been quietly delayed twice, both times because someone senior looked at the risk of moving in a peak-trade window and said 'not this quarter'. By the time we arrived, 'not this quarter' had become the standing policy. The company was growing at thirty per cent a year on infrastructure that would not survive fifty per cent. Our brief was surgical: re-platform onto a composable stack — Shopify Hydrogen, Sanity, a bespoke checkout — inside sixteen weeks, without a merchandising freeze, without a revenue dip during the cutover, and without asking the finance team to hold their breath. In an industry where re-platforms routinely lose fifteen per cent of revenue on the go-live weekend, 'zero dip' was not just a stretch goal; it was the entire brief. The CFO had, in her own words, budgeted for a five per cent dip because that was the smallest number the industry had ever seen for a project of this size. Our first meeting was, effectively, an argument about whether zero was even possible.

Our approach

We ran a strangler-fig migration over twelve weeks. Rather than rebuild everything and swap on a single Friday night, we stood up a new storefront layer in parallel with the legacy monolith and shifted traffic gradually by product cluster. The lowest-revenue SKUs went first — bath and body, then home fragrance, then the accessories categories — with double-writes to both stacks so analytics and inventory stayed in perfect parity. By week eight, sixty per cent of transactions were flowing through the new stack while the old one continued to serve the flagship categories. Every migration cluster was preceded by a two-day soak test at production traffic volumes; two clusters were rolled back and re-migrated a week later after we caught latency regressions in the reviews module. The cutover of the last twenty per cent of traffic happened on a Tuesday morning, deliberately unremarkable. Deploy frequency went from every nineteen minutes to under three; time-to-first-byte dropped by sixty-three per cent. The finance team's quarter-close, which usually included a phone call about site issues, closed a full day early. Atlas has since added two engineers to the platform team and shipped fifteen major merchandising campaigns in the year post-launch, without a single P0 incident. The board later described the project as the first re-platform they had authorised that had returned its budget inside twelve months, and set a new internal precedent: every future infrastructure decision at Atlas would be measured against a strangler-fig plan first, before any big-bang alternative was even discussed. That single procedural change may end up being the most valuable artefact we shipped.

Outcomes

“We replatformed in Q4 and our finance team didn't notice. That's the highest compliment I can give.”— Priya Mehta, COO