A resource request is a reservation, not a measurement. Set it high to be safe and the scheduler buys node capacity for consumption that never arrives, which is why the writeup names requests as the biggest lever and node count as the wrong place to start [5]. The failure mode carries no signal: nothing pages, and the waste spreads thinly across hundreds of pods instead of parking itself in one line item you can point at [12][4].
The headline number needs care. The 30-50% is a share of wasted spend, not of the invoice [3]. Take the article's own illustration, a bill 40% higher than the workload needs, and the waste is 28.6% of what you actually pay [13]. Clearing 30-50% of that returns between 8.6% and 14.3% of the bill [14]. That is real money on a large cluster and it is nothing like what the headline implies, so the distinction is worth settling before a finance conversation rather than during one.
The prescribed method also carries a clock. You are told not to guess: pull two to four weeks of real usage, or run the Vertical Pod Autoscaler in recommendation-only mode for a couple of weeks, before editing anything [10]. So the largest lever is the slowest one to pull, roughly a month out from the day you start watching [16]. What moves this month is disks and namespaces. Released PersistentVolumes are the cleanest of those: deleting a PVC does not always take the underlying cloud disk with it, depending on reclaimPolicy, and a single jq query separates the volumes with a claim from the ones billing you for nothing [6][7].
The autoscaler floor belongs in the same bucket. Teams tune the maximum node count and leave the minimum at whatever felt safe when the cluster was provisioned, which bills every night and every weekend at that floor [9]. It is one integer in one config, and unlike requests it needs no observation period.
Two things in the source deserve pushback. The sample manifest sets limits at exactly twice requests on both CPU and memory [17][18], which is a convention rather than a finding; the whole argument of the piece is that your own usage data sets those numbers. And the four-figure monthly recovery from released volumes is offered without a named cluster or an audit trail [7], so treat it as an existence proof for looking, not a forecast for your own bill.
The ordering is also off. Labels appear last, after the cleanup, yet the taxonomy of team, application, environment and cost-center enforced at admission by Kyverno or Gatekeeper is what makes rightsizing by team, chargeback and showback possible at all [11]. It is also what lets you demonstrate that Phase 1 worked. And it is where the promise of no new tooling budget [3] gets thin: admission controllers and VPA are free to license, not free to run, and enforcement means somebody's deploy starts failing for a missing label.
