Skip to content

B2B SaaS · two EKS clusters on AWS · under NDA

Published · 3 min read

Making a savings number smaller

A European SaaS platform on AWS, two EKS clusters, a five-figure annual cloud bill. Automated detectors produced a headline savings figure. Reconciling them against each other removed roughly three quarters of it — and that is what made the remainder usable.

The situation

The estate was already instrumented: cost data by usage type, a metrics agent in both clusters, and a set of detectors producing findings across compute, storage, databases, networking and security. Summed naively, they produced a large and encouraging number.

It would not have survived the first technical review, for three separate reasons.

Solution

01The same money, counted from two directions

Most of the cluster nodes flagged as idle were managed by an autoscaler that sizes the fleet from pod requests. Their cost and the separately measured over-request of the workloads running on them are the same money, reached from two ends. Both appeared in the total.

We folded them into one finding, with the measured request overhead as the low end of the range and the node-level estimate as the high end — and said explicitly what the gap between the two represents.

02Advice the reader cannot carry out

The same findings recommended resizing those nodes. On an autoscaled cluster that is not an action anyone can take: the node type follows the requests, and a node deleted by hand returns within a minute.

The recommendation was rewritten to name the lever that exists — cut the requests, let consolidation remove the node — and the finding was reclassified as evidence rather than as an independent saving.

03A window that does not match its label

One family of findings reported a multi-month total under a monthly label, inflating those lines several-fold. It was caught by the oldest check there is: every figure in the report has to reconcile to a line on the bill for a named month.

Those findings were rebuilt from the billing data directly, and the defect was fixed at the source rather than worked around.

Outcome

−73%
removed from the claimed figure, after review
2
EKS clusters, every figure traced to a bill line
2
findings worth doing that save nothing
What was delivered
A range rather than a maximum, with the basis for each end stated
Traceability
Every figure to a bill line and a named month, checkable without us
Sequencing
What to do first, what it depends on, what to leave until the fleet settles
Gaps
An explicit list of what we could not evaluate and what data would close each

Roughly a quarter of the naive total, and defensible line by line. Two findings were marked as costing nothing and worth doing anyway, because the risk was worse than the spend.

A large savings figure is easy to produce and expensive to defend. The moment someone senior asks an engineer to check it, every double-count and every unexecutable recommendation becomes a reason to distrust the whole exercise — including the findings that were real. We would rather hand over a smaller number that holds. It is worth more, because it gets acted on.

Core tech

  • AWS
  • Amazon EKS
  • Karpenter
  • AWS Cost Explorer
  • Prometheus
  • ClickHouse
  • Parquet
  • MetrixForge

Method

What transfers to other estates

Nothing here is proprietary. This is the part that transfers to your estate whether or not you ever call us.

Count each dollar once

Node-level and workload-level findings that name the same money are one finding, not two. On an autoscaled cluster that overlap is the rule rather than the exception.

Discard advice the reader cannot execute

Resizing a node an autoscaler owns is not an action. Rewrite it as the lever that exists, and reclassify the original as evidence.

Reconcile every figure to a bill line and a month

It is the cheapest check available and it catches the expensive mistakes — including a multi-month total wearing a monthly label.

Publish the gap list

What could not be evaluated, and what data would close it. A reviewer who hides that has moved the uncertainty somewhere the reader cannot see it.

More

Other work

All case studies

Gambling · 500+ applications across several data centres · under NDA

Moving more than 500 applications to AWS, one at a time

An on-premise estate spread across several data centres, grown past what it could carry. More than 500 applications moved to AWS one at a time, least critical first, with the strategy chosen per application — and about one in seven retired instead of moved.

E-commerce · production EKS on AWS · under NDA

A storage bill grew 20× on a filesystem holding 8 GB

Daily spend started climbing in the same week a dozen database instances were shut down — so the bill should have been falling. The billing line said EFS. The cause was in node provisioning, two layers away.

E-commerce · 18 RDS PostgreSQL instances on AWS · under NDA

Consolidating 18 production databases into 5

A legacy account with one RDS instance per service, most idling at 2–10% CPU. The business would grant two maintenance windows of fifteen to twenty minutes, on separate days, for a payment-adjacent platform. Everything else had to happen under live traffic.

Gambling · several production databases on one server · under NDA

Splitting one database server into three

Several production PostgreSQL databases shared a single server, so one busy neighbour was everybody's problem. Split onto three dedicated RDS instances, grouped by what actually contended rather than by size.

Illustrative · managed Kubernetes on AWS

Three EKS versions upgraded without a maintenance window

A cluster two versions past its support window, moved forward across three minor releases with no scheduled downtime — rehearsed first, with the disruption budgets fixed before anything moved.

Illustrative · CI on a hosted runner fleet

A CI pipeline from 38 minutes to 9

Build and test times measured per step before anything was touched, then caches that actually hit, parallelism where the dependency graph allowed it, and the work that had been running on every commit for no reason.