Retail Peak: Designing for Black Friday

The shape of the problem

Large retailers — Walmart, Amazon, Shopify's merchants and their peers — face a load profile that most systems never encounter. Traffic during a peak event can be an order of magnitude above a normal day, it arrives within minutes of a published start time, it is heavily concentrated on a small number of items, and the date cannot be moved.

Failure is immediately and publicly measurable in lost revenue, and there is no possibility of deferring the event to fix the system.

The engineering that follows

Pre-scaling, not autoscaling. Autoscaling has a reaction time measured in minutes and depends on a control plane that may itself be under stress. Capacity is provisioned ahead of a known event, and the scaling system is a supplement rather than the plan.

Load testing at genuine peak volume with a realistic mix, including the concentration on hot items, which is where caching and inventory contention actually break. A uniform load test at peak volume proves very little about a workload where 90% of traffic targets 1% of the catalogue.

Deliberate degradation ladders, decided in advance: which features are shed first as load rises — recommendations, reviews, personalisation — so the checkout path survives. This is a business decision made calmly beforehand rather than by whichever component fails first.

Change freezes narrowly scoped to the systems in the blast radius, for the shortest window that matters, with a practised emergency path — rather than a month-long estate-wide freeze that batches risk into January.

Inventory contention, which is the genuinely hard part: a limited-stock item with enormous concurrent demand is a hot key, a consistency problem and an overselling risk simultaneously.

The transferable lesson

A predictable peak is a gift, because it converts capacity planning from forecasting into engineering. The date is known, the shape is known from last year, and there is time to rehearse.

The practices generalise to any organisation with a known event: a marketing launch, a regulatory deadline, a seasonal cycle, an enrolment window. The failure is treating it as a normal week with more traffic, when it is a distinct operating mode that deserves its own capacity model, its own degradation plan and its own rehearsal.