AI
What OpenCost Measures and Why Clusters Overprovision
OpenCost is a vendor-neutral CNCF incubating project. It models cost allocation and asset usage across Kubernetes clusters with multi-cloud support. On-premises teams can supply a custom pricing sheet. Attapattu’s core observation: clusters are grossly overprovisioned because workloads are not right-sized. Tools like Karpenter exist, but without cost visibility, engineers don’t know where to cut. OpenCost surfaces that gap directly.
The Three-Zone Cost Model Inside a Cluster
Every cluster has three cost zones. Overhead covers control plane expenses. Allocation costs cover CPU, memory, GPU, and disk, billed from the moment a node provisions. Resource usage costs track activity: egress, cross-zone, and cross-region traffic. Cross-zone chatter inside a single cloud vendor is a common blind spot. OpenCost captures these network traffic data structures even though the open source version does not yet surface full network cost attribution.
Right-Sizing Containers to Shrink Cluster Spend
Right-sizing starts at the container level. Reducing a container’s resource request lowers the cluster’s total demand. Autoscaling and Karpenter then shrink the underlying nodes automatically. Without autoscaling, OpenCost still shows idle capacity so teams can resize manually. Attapattu also pointed to unit economics as a second output: the same allocation data builds the foundation for internal chargeback, which larger organizations use to attribute spend accurately across teams.
Prometheus Mode, Promless Alpha, and the Plugin Architecture
OpenCost currently runs on Prometheus. Large clusters with many workloads hit performance limits there, so the community is building a promless mode, currently in alpha. The plugin architecture extends coverage beyond Kubernetes: a Datadog plugin already exists, and a Snowflake plugin is working through pull requests. Both let teams pull non-Kubernetes cost models into OpenCost alongside their cluster data.
MCP Server Demo: Querying Costs with Natural Language
An intern built an MCP server for OpenCost through the LFX mentorship program. Attapattu queried it live, asking for top namespaces and the top five deployments by CPU and memory. The MCP server calls the OpenCost REST APIs directly. A UI is available at opencost.io under the demo link, though Attapattu flagged it is not polished. The APIs and data model are what matter most for downstream integrations.
Notable Quotes
you’d be surprised to see how many clusters are grossly overprovisioned, mostly because the workloads are not sort of right sized. Rajith Attapattu · ▶ 02:32
what’s important with Opencast is the model, the APIs, and then obviously things like the MCP server. Rajith Attapattu · ▶ 07:01
chatter between those zones, that’s going to cost money. Rajith Attapattu · ▶ 03:42
Key Takeaways
- OpenCost splits cluster spend into overhead, allocation, and resource usage zones.
- Containers right-sized first reduce node demand, letting Karpenter shrink clusters automatically.
- A new promless alpha mode targets performance limits in large, high-workload clusters.
About the Speaker(s)
Rajith Attapattu is CTO of Randoli and a member of the Apache Software Foundation. He contributes to OpenTelemetry and open source observability tooling, and is building an OTel-native observability platform at Randoli. He also maintains OpenCost as a CNCF project contributor.