How Vorboss Boosted Kubernetes Reliability (and Cut Costs) with Predictive Autoscaling

About Vorboss

Vorboss isn't your typical ISP. They built over 700km of their own dedicated fibre network across Central London from scratch — no legacy infrastructure, no third-party dependencies, no contractors. Every engineer on their team is a Vorboss employee. They provide business internet up to 100Gbps, managed IT, and cybersecurity to thousands of organisations, from SMEs to enterprises and public sector. When they promise reliability, they mean it: they back it with one of the most aggressive SLAs in the industry. It's no surprise they've been named the UK's Best Enterprise ISP.

That same obsession with reliability extends to their internal engineering. Vorboss runs on Kubernetes, and their platform team manages hundreds of workloads across multiple clusters. But like most teams operating at that scale, they were running into a familiar set of problems.

The Classic Kubernetes Scaling Dilemma

If you've ever managed Kubernetes workloads in production, you know the drill. Setting resource requests and limits correctly is tedious, easy to get wrong, and rarely prioritised until something breaks. Most teams end up in one of two camps: overprovisioned workloads that burn money, or underprovisioned workloads that risk service degradation. Often, it's both at the same time — across different services in the same cluster.

The Vorboss engineering team was navigating the same reality. Some workloads were sitting at around 15% utilization — safe, sure, but significantly inefficient. Others were running hot, edging toward resource exhaustion and creating real risk. The platform engineering team was stuck in a reactive loop: manually right-sizing workloads, educating developers on the intricacies of Kubernetes pod resource request semantics, and constantly playing catch-up. It's the kind of operational toil that eats engineering hours without moving the product forward.

Traditional autoscalers don't solve this problem. Kubernetes' built-in HPA is reactive by design — it waits for a spike to happen, then scrambles to respond. That lag means you either keep a big resource buffer around (expensive) or accept that some spikes will cause degradation (risky). Pick your poison.

Enter Predictive Autoscaling

This is where Thoras comes in. Instead of reacting to demand after the fact, Thoras analyzes historical resource usage, real-time metrics, and custom signals to predict what your workloads will need — and scales them before the demand arrives. It deploys directly into your cluster via Helm, integrates with Prometheus, and runs entirely in-cluster with no data leaving your environment. No manual threshold tuning, no guesswork.

For Vorboss, the impact was immediate and tangible across hundreds of workloads:

Overprovisioned workloads went from ~15% utilization to 85%. That's not a rounding error but rather a fundamental shift in cost efficiency. Thoras identified workloads that were sitting on far more resources than they needed and safely tightened them to optimal levels, all without introducing risk.

Underprovisioned workloads were brought out of the danger zone. Services that had been running at risk of degradation were brought into safe resource allocation levels around 80% utilization. Instead of waiting for an incident to reveal a problem, Thoras surfaced the risk proactively and resolved it.

Engineering hours came back. The platform team no longer needs to spend cycles manually right-sizing workloads or running education sessions on the finer points of Kubernetes resource management. Thoras runs continuously, so the clusters essentially tune themselves. That's time the engineering team can now redirect toward building features and improving the platform — work that actually moves the business forward.

Reliability and Cost Savings — Not a Trade-off

Here's the part that usually gets a raised eyebrow: Vorboss came out ahead on both reliability and cost. Conventional wisdom says you pick one. You over-provision for safety and accept the waste, or you optimize for cost and accept some risk.

Predictive scaling breaks that trade-off. Because Thoras forecasts demand and scales in advance, workloads don't need a massive buffer of idle resources "just in case." You can target higher utilization safely because the system is already preparing for what's coming next. The result is that Vorboss fixed their at-risk workloads and their wasteful ones, and still saved money doing it.

A Partnership, Not Just a Product

One thing worth highlighting is how the engagement itself worked. Getting predictive autoscaling running across hundreds of workloads and multiple clusters isn't something you toss over the wall. Thoras and Vorboss worked together as a genuine partnership — collaborating closely to get everything configured, validated, and tuned to Vorboss' specific environment and workload patterns. The Thoras team embedded alongside Vorboss' platform engineers, building trust through transparency and shared ownership of the outcome. Now the clusters scale themselves proactively, around the clock, without ongoing manual intervention.

For a company like Vorboss — one that prides itself on operational excellence and keeping London businesses connected — having infrastructure that anticipates demand rather than chasing it is a natural fit. Their customers expect zero downtime. Now their Kubernetes platform is engineered to deliver exactly that.

“We were looking for a platform that could help us improve efficiency without compromising reliability. We provide critical national infrastructure, and our production workloads are mission critical, so trust was non-negotiable. A lot of optimization tools focus heavily on cost, but they don’t always put the same emphasis on uptime that Thoras does. We’re running our Kubernetes environments more efficiently, and just as importantly, we know scaling is being handled in a way that protects production. That balance of efficiency and reliability is exactly what we needed.

Aaron Rice, CIO”