Insight types
CostGraph checks for several kinds of findings. The available insights depend on the providers you connect and the usage data you collect.
Waste and savings findings appear only when they are worth at least $10 per
month. Smaller findings stay hidden so the list shows changes worth making.
Reliability and data quality findings appear whatever their cost.
They have no savings figure.
What CostGraph checks
CostGraph groups related findings about one resource into a single insight. For example, one insight per workload covers everything wrong with its resources and health. The insight takes the severity of its most serious finding and lists each finding with its measurements and cost. The following tables list the checks by area.Kubernetes workloads
Each workload has one resources and health insight. It covers these findings:
CostGraph AutoPilot can fix memory and CPU
sizing, and raises memory after an out-of-memory kill. For CPU throttling,
raise or remove the workload’s CPU limit. Findings about a workload that is not
running need a fix in the workload or its configuration.
Kubernetes clusters
To find node pool savings, CostGraph sizes each workload at its recommended
requests and plans the pool two ways. One plan packs the largest workloads
first. The other pairs workloads whose CPU and memory needs complement each
other. CostGraph recommends the plan that fits more workloads, then the cheaper
one. Pools that mix CPU-heavy and memory-heavy workloads often fit on fewer or
smaller nodes this way.
Autoscaler choice
The autoscaler choice insight compares what each node autoscaler would cost for a cluster’s workloads. It covers the following options on each cloud, as well as keeping the current setup:
CostGraph prices every option on the same workloads at their recommended
requests, using the node prices on the cluster’s bill. Keeping the current
setup counts what tuning its settings can reach. Options that scale node groups
use on-demand capacity only. Karpenter, auto-provisioning, and compute class
options can also use spot capacity and split a pool in two.
The insight appears when the cheapest option saves more than tuning the current
setup, and its saving is only that extra amount. It charts the monthly cost of
each option, splits the saving into tuning, tighter packing, and cheaper
capacity, and lists the checks the recommended option fails. Every compared
option except keeping the current setup checks for pods whose disruption budget
allows no evictions and pods marked as not safe to evict. Karpenter also needs
spot interruption handling. GKE and AKS node auto-provisioning need a cluster
with no Windows nodes. Each failed check raises the migration effort, which
breaks ties between options that save the same amount.
CostGraph does not price the management fee of EKS Auto Mode or GKE Autopilot
yet. The insight lists them without comparing them, and does not run on a
cluster where any pool already uses one. Refer to
Supported node autoscalers for how
CostGraph recognizes each autoscaler.
Volumes
Each volume has one health insight. It covers these findings:GPUs and databases
Each GPU and each database has one health insight.Commitments
See Commitments for coverage and utilization.
Waste on your bill
Network and AI
Sandboxes
The cheaper provider and idle time findings read the per-sandbox rows on your
provider bill, so they need the provider’s cost integration. Idle time needs a
provider that reports it, such as Daytona.