What Phrontis does
The report and recommendations are free. Automation is Pro, starts in dry-run, and changes only what you turn on.
See the bill
Cost per namespace and workload
Each container is charged its request plus its usage above the request. On an overcommitted node, requests are paid first and bursts share what is left.
Whatever a node has left over is idle, so namespaces plus idle equal the cluster.
Real node prices
AWS and Azure Linux nodes are priced from the vendor's public list price, found from the instance type, region and spot labels. Prices are USD list prices, before discounts or savings plans.
Other nodes use flat per-vCPU-hour and per-GiB-hour rates that you set.
Storage and load balancers
Persistent volumes and LoadBalancer services get their own sections, apart from compute. Volumes with no bound claim show up as unclaimed, the storage waste to look at.
Right-size requests
The recommended request is the peak usage plus 20% for CPU and 30% for memory. Changes under 20% are not shown.
OOM guard
A container whose last exit was an OOM kill never gets a memory request below its current request, or its memory limit plus 30%.
Throttling flag
A container whose CPU peak reaches 90% of its limit is flagged as throttled and always listed. Where a new request would pass the limit, the limit is raised with it.
Cheaper nodes
Karpenter NodePool
On AWS, the report bin-packs your pods onto c, m and r instance types with the rightsizing applied, picks the cheapest fleet, and prints a ready Karpenter NodePool.
Spot candidates
The free report lists which workloads can move to spot. Deployments with two or more pods are spot-safe; the rest are named so you can pin them to on-demand.
Automation Pro
Each one starts as a dry-run that shows what would change and what it would save.
Workload autoscaler
Label a namespace and its new pods start with the recommended requests. Your manifests in git stay as they are. After an OOM kill, the memory recommendation goes up for the next pod.
Node autoscaler
Set the families, architecture, spot and a vCPU limit. Phrontis shows the diff and applies the Karpenter objects only after you confirm, draining at most the share of nodes you allow, inside your maintenance window.
Spot with on-demand fallback
Spot workloads prefer spot nodes. When spot capacity runs out, pods land on on-demand nodes instead.
Journal and kill switch
Every change is logged with the spec before and after. One command stops automation and leaves the cluster as it is.
Try it on your laptop first
The CLI reads your current kubeconfig context and prints the same report as text.
# one usage sample from the current context phrontis # another context, a longer window, your own flat rates phrontis --context prod --window 10m --cpu-hour 0.04 --gib-hour 0.005