The replatform we recommended against

The situation

A direct-to-consumer retailer had approved a fourteen-month rebuild of their commerce platform. The stated reasons were performance, the difficulty of hiring for the existing stack, and a checkout that fell over during promotions.

We were engaged for three weeks to validate the plan before contracts were signed with the implementation partner.

What we found

The performance problem was real and severe — p95 checkout latency of 4.1 seconds, with timeouts under load. But it was not distributed across the platform. It was concentrated in two places.

The product catalogue was querying inventory per line item rather than in batch, producing between forty and three hundred database round trips per checkout depending on basket size. And the session store had been placed on the same database instance as the transactional workload, so promotional traffic caused the two to contend for the same connection pool.

Neither problem was architectural. Both were fixable in weeks. Neither would have been fixed by a rebuild — the proposed design reproduced the second one.

The hiring argument was more genuine, but the rebuild did not solve it either: the target stack was scarcer in their market than the one they had.

What we recommended

A written assessment with three options priced against each other: the rebuild as proposed, a targeted remediation, and a staged extraction of the catalogue service only if growth later justified it.

We recommended the second, at an estimated $96,000 — roughly seven percent of the approved budget. We also named what the client would not get for that money, because the rebuild had been sold partly on benefits the remediation genuinely would not deliver.

The board took some convincing. The implementation partner disagreed with our read, in detail and in good faith, and we published their response alongside ours in the final document.

Where it landed

The remediation was carried out by the client’s own team, with us reviewing design and pull requests one day a week for two months. p95 checkout latency is now 0.9 seconds and has held through two promotional peaks.

The rebuild has not been revisited. We would still support it if the catalogue extraction becomes necessary — the assessment says so, and says under what conditions.

We include this engagement here because a due diligence practice that never recommends against the plan is not doing diligence.