Skip to content

Retail · 2024 · Client anonymised at their request

Surviving a campaign that had taken the old platform down twice

Commerce replatform sized against a measured load profile rather than a guess.

A national retail chain

Diagram: a measured campaign load curve rising to fourteen times baseline traffic, with the real failure ceiling marked at connection pool exhaustion, well below the compute ceiling everyone assumed. Alongside it, the four changes made — cached read paths with explicit invalidation, checkout isolated in its own connection pool, admission control, and progressive rollout with automated rollback.

14×

Peak traffic absorbed

99.98%

Campaign-period availability

1.2s

Largest contentful paint at peak

01

The constraint

Two previous seasonal campaigns had ended in outages, and the business had no reliable model of what its actual peak looked like. We were given eleven weeks before the next campaign.

02

The architecture decision

We characterised load against production-like data with k6 before changing anything, which established that the failure was database connection exhaustion rather than the assumed compute shortfall. Read paths moved behind a cache with explicit invalidation, checkout was isolated into its own connection pool with a queue and admission control, and progressive rollout gates were wired into the deployment pipeline.

03

What shipped

  • 01

    Load characterisation and capacity model

  • 02

    Cache layer with explicit invalidation contracts

  • 03

    Isolated checkout path with admission control

  • 04

    Progressive delivery pipeline with automated rollback

The campaign ran at fourteen times baseline traffic without incident. The capacity model is now reviewed before every seasonal event.

Figures verified with the client’s own instrumentation, not ours.

04

What it moved

Practices involved

  • DevOps & scaling
  • Cloud infrastructure
  • Product engineering