WooCloud All articles
Cloud Economics

May Day, Black Friday and the Tuesday After: Modelling UK Seasonal Cloud Peaks

Callum MacLeod
May Day, Black Friday and the Tuesday After: Modelling UK Seasonal Cloud Peaks

Photo: Josie Campbell , CC BY-SA 2.0, via Wikimedia Commons

Every UK business has a rhythm. Retail lives and dies by the last week of November. Anything touching finance spikes on 31 January. Travel and hospitality brace for the May bank holidays, then enjoy a soft August. Your cloud bill, if you read it carefully, has the same pulse — just delayed by a few weeks and buried under a dozen service lines.

Most cost forecasts treat demand as a gentle upward slope with a bit of noise. Then November happens, autoscaling does something unexpected, and finance asks a question nobody can answer until the invoice lands in December.

Traffic has a calendar, not just a trend line

Seasonality is the cheapest forecasting advantage available to a UK team, because the peaks are not secret. They are in the retail calendar, the payroll calendar and the public holidays list. What is missing is the mapping between those events and the specific services that pay for them.

Start by annotating twelve months of billing data with UK events and looking for correlation. Some of it will be obvious. Some of it will surprise you — the load on your customer service platform during a bank holiday weekend, for instance, which nobody put in the capacity plan because the office was shut.

The peaks worth putting in the model

Baseline, burst, and the Tuesday after

The common mistake is modelling the peak and stopping there. The expensive part is often the shoulder. A retail platform that triples capacity on Friday and holds it through the weekend will pay for that capacity on a Monday when demand has gone home. Idle compute at peak rates is how a good trading week becomes a mediocre one.

Model three numbers per event: the floor your estate will not go below, the burst ceiling, and the decay curve back to normal. That third number is where most of the waste hides.

Match the buying model to the shape

Commitments should cover the floor, not the ceiling. Reserved capacity and savings plans are excellent value for the baseline that runs all year — the load balancers, the databases, the logging pipeline. They are a poor fit for a three-week retail surge, because a one-year commitment bought for November will still be charging you in July.

So split the estate deliberately. Baseline on a commitment, burst on on-demand, and batch or non-urgent work on spot where the failure mode is tolerable. Then check the split annually: most teams find the committed percentage drifting, usually downwards, as new services get launched without anyone attaching them to the plan.

Autoscaling you have not tested is a rumour

A scaling policy that has never met real load is a hope, not a control. Well before your peak, run a load test that exercises the full chain: scaling limits and service quotas, warm pools, database connection ceilings, and the rate limits of every third party you call. Provider quotas in particular need raising in advance, and that request is not instant.

The other classic is a replica that scales beautifully while the database behind it quietly becomes the bottleneck. Add capacity, get a slower system, and pay more for the privilege.

Alerts that fire before the invoice

Daily cost granularity, anomaly detection per service, and a named human who receives the alert. Monthly budget reviews are a post-mortem, not a control. Set thresholds against your forecast rather than against last month, and route the alerts to whoever is on call during the peak — because that is when someone might actually act.

The change freeze is a cost lever too

Most UK teams freeze changes through December. That is sensible for stability and quietly useful for cost: no migrations, no re-architecture, no new services arriving three days before the busiest week. The corollary is that January is where the backlog lands, so plan the re-architecture for the quiet weeks rather than pretending it will happen in the run-up to Christmas.

A one-page model you can defend

Keep it simple enough to explain in a meeting: traffic index per event, multiplied by unit cost per unit of traffic, multiplied by a seasonality multiplier you have calibrated against last year. Show the forecast, then show actuals against it after each event. Within two cycles you will have a seasonality file that is genuinely yours rather than a vendor's rule of thumb — and a finance conversation that starts with what you predicted instead of what arrived.

All Articles

Related Articles

The Bill That Grows When Nothing Breaks: A UK Guide to Observability Costs

The Bill That Grows When Nothing Breaks: A UK Guide to Observability Costs

The Premium You Didn't Budget For: Running a Second Cloud Estate in the UK

The Premium You Didn't Budget For: Running a Second Cloud Estate in the UK

The Side Door on Your Cloud Bill: Reining In Marketplace Spend

The Side Door on Your Cloud Bill: Reining In Marketplace Spend