Servers ready fifteen minutes before the crowd, switched off when it leaves

In short

Instead of waiting for servers to saturate, a model forecasts traffic peaks (flash sales, weekend evenings, campaigns) and starts the capacity just before, then scales it down when things calm down.

Client: Halvora Retail
Sector: Fashion e-commerce
Size: 300 staff, flash sales every week

15 minahead of traffic peaks
−31%on the infrastructure bill
−88%latency during flash sales

The context

Halvora Retail’s flash sales push traffic up tenfold in minutes. Standard Kubernetes autoscaling reacts too late: the site slows down at the worst moment. To compensate, the team kept spare servers running all week.

The challenge

01
Reactive autoscaling only adds servers after saturation: the first minutes are lost.
02
Permanent over-provisioning is expensive, for a few peak hours a week.
03
Peaks follow the marketing calendar, not just the clock.
04
A wrong forecast must never take the site down.

Our approach

We train a forecasting model (Prophet, with anomaly detection) on the Prometheus history, enriched with the campaign calendar and public holidays. Every five minutes it forecasts the next quarter hour of traffic and drives KEDA to pre-warm pods. Reactive autoscaling stays as a safety net, and every saving is measured.

The solution, step by step

01
History and calendar
Months of metrics, plus marketing campaigns and holidays.
02
A 15-minute forecast
The model anticipates the next quarter hour of traffic.
03
Servers pre-warmed
KEDA starts the pods before the crowd arrives.
04
A flash sale, no slowdown
Latency stays flat, where reactive autoscaling fell behind.
05
Calm, and savings
Capacity drops right away, and the saving is reported monthly.

Results

−31%on the infrastructure bill
210 msp95 latency at the height of a flash sale, vs 1.8 s
15 minaverage anticipation
100%of flash sales without incident

What we deliver

Data pipeline (Prometheus, marketing calendar)
Forecasting model and anomaly detection
KEDA control and autoscaling guardrails
FinOps savings dashboard
Monthly performance report

Technologies

Kubernetes
KEDA
Prometheus
Python
Prophet
scikit-learn

Timeline

8 to 12 weeks

Team

1 DevOps engineer, 1 data scientist

A similar project? Let’s talk

Next case · Talvane RHA security review that understands the architecture, not just the syntax

Illustrative cases: clients and figures are fictional, representative of the solutions we deliver.