Skip to content
This repository has been archived by the owner on Feb 2, 2024. It is now read-only.

Latest commit

 

History

History
135 lines (100 loc) · 8.36 KB

allocation-api.md

File metadata and controls

135 lines (100 loc) · 8.36 KB

Kubecost APIs

⚠️ DEPRECATED ⚠️ These APIs are outdated and should not be used. Please refer to the Allocation API docs

Kubecost exposes multiple APIs to obtain cost, resource allocation, and utilization data. Below is documentation on two options: the cost model API and aggregated cost model API.

Cost model API

The full cost model API exposes pricing model inputs at the individual container/workload level and is available at:

http://<kubecost-address>/model/costDataModel

Here's an example use:

http://localhost:9090/model/costDataModel?timeWindow=7d&offset=7d

API parameters include the following:

  • timeWindow dictates the applicable window for measuring cost metrics. Supported units are d, h, and m.
  • offset shifts timeWindow backward relative to the current time. Supported units are d, h, and m.

This API returns a set of JSON elements in this format:

{
  cpuallocated: [{timestamp: 1567531940, value: 0.01}]
  cpureq: [{timestamp: 1567531940, value: 0.01}]
  cpuused: [{timestamp: 1567531940, value: 0.006}]
  deployments: ["cost-model"]
  gpureq: [{timestamp: 0, value: 0}]
  labels: {app: "cost-model", pod-template-hash: "1576869057"}
  name: "cost-model"
  namespace: "cost-model"
  node: {hourlyCost: "", CPU: "2", CPUHourlyCost: "0.031611", RAM: "13335256Ki",…}
  nodeName: "gke-kc-demo-highmem-pool-b1faa4fc-fs6g"
  podName: "cost-model-59cbdbf49c-rbr2t"
  pvcData: [{class: "standard", claim: "kubecost-model", namespace: "kubecost",…}]
  ramallocated: [{timestamp: 1567531940, value: 55000000}]
  ramreq: [{timestamp: 1567531940, value: 55000000}]
  ramused: [{timestamp: 1567531940, value: 19463457.32}]
  services: ["cost-model"]
}


Optional request parameters include the following:

Field Description
filterFields Blacklist of fields to be filtered from the response. For example, appending &filterFields=cpuused,cpureq,ramreq,ramused will remove request and usage data.
namespace Filter results by namespace. For example, appending &namespace=kubecost only returns data for the kubecost namespace

Aggregated cost model API

NOTE: this API is actively being replaced by the Kubecost Allocation API. That is the recommended API for querying historical and run-rate cost allocation metrics.

The aggregated cost model API retrieves data similar to the Kubecost Allocation frontend view (e.g. cost by namespace, label, deployment, and more) and is available at the following endpoint:

http://<kubecost-address>/model/aggregatedCostModel

Here are example uses:

  • http://localhost:9090/model/aggregatedCostModel?window=1d&aggregation=namespace
  • http://localhost:9090/model/aggregatedCostModel?window=1d&aggregation=label&aggregationSubfield=product
  • http://localhost:9090/model/aggregatedCostModel?window=1d&aggregation=namespace&sharedNamespaces=kube-system

API parameters include the following:

  • window dictates the applicable window for measuring cost metrics. Current support options:
    • "15m", "24h", "7d", "48h", etc.
    • "today", "yesterday", "week", "month", "lastweek", "lastmonth"
    • "1586822400,1586908800", etc. (start and end unix timestamps)
    • "2020-04-01T00:00:00Z,2020-04-03T00:00:00Z", etc. (start and end UTC RFC3339 pairs)
  • offset (optional) shifts window backward from current time. Supported units are d, h, m, and s.
  • aggregation is the field used to consolidate cost model data. Supported types are cluster, namespace, controller, deployment, service, label, pod and container.
  • aggregationSubfield used for aggregation types that require subfields, e.g. aggregation type equals label and the value of the label (aggregationSubfield) equals app. Comma-separated list of values supported.
  • allocateIdle (optional) when set to true, applies the cost of all idle compute resources to tenants, default false.
  • sharedNamespaces (optional) provide a comma-separated list of namespaces (e.g. kube-system) to be allocated to other tenants. These resources are evenly allocated to other tenants as sharedCost.
  • sharedLabelNames (optional) provide a comma-separated list of Kubernetes labels (e.g. app) to be allocated to other tenants. Must provide the corresponding set of label values in sharedLabelValues.
  • sharedLabelValues (optional) label value (e.g. prometheus) associated with sharedLabelNames parameter.
  • sharedSplit (optional) Shared costs are split evenly across tenants unless weighted is passed for this request parameter. When allocating shared costs on a weighted basis, these costs are distributed based on the percentage of in-cluster resource costs of the individual pods in the particular aggregation, e.g. namespace.
  • disableCache this API caches recently fetched data by default. Set this variable to false to avoid cache entirely.
  • etl setting this variable to true forces a request to be served by the ETL pipeline. More info on this feature is in the Caching Overview section below.


Optional filter parameters include the following:

Filter Description
cluster Filter results by cluster ID. For example, appending &cluster=cluster-one will restrict data only to the cluster-one cluster. Note: cluster ID is generated from cluster_id provided during installation.
namespace Filter results by namespace. For example, appending &namespace=kubecost only returns data for the kubecost namespace.
labels Filter results by label(s). For example, appending &labels=app%3Dcost-analyzer only returns data for pods with label app=cost-analyzer. CSV list of label values supported. Note that parameters must be URL encoded.

This API returns a set of JSON objects in this format:

{
  aggregation: "namespace"        // value of aggregation type parameter
  cpuAllocationAverage: 0.01      // average number of cores allocated over time window, max(request,usage)
  cpuCost: 0.053106479            // total cost of CPU allocated
  cpuEfficiency: 0.0469166        // ratio of cost-weighted CPUs being utilized
  efficiency: 0.1476976           // efficiency of both CPU and RAM provisioned
  environment: "ingress-nginx"    // instance of aggregation 
  gpuAllocationAverage: 0         // average number of cores allocated over time window, based on request
  gpuCost: 0                      // total cost of GPU allocated
  networkCost: 0                  // cost of network egress
  pvAllocationAverage: 0          // average GB allocated, based on amount provisioned
  pvCost: 0                       // total cost of persistent volumes allocated
  ramAllocationAverage: 0.0880    // average number of RAM GB allocated over time window, max(request,usage)
  ramCost: 0.006268486            // total cost of RAM allocated
  ramEfficiency: 1.00             // ratio of cost-weighted RAM being utilized
  sharedCost: 0                   // value of costs allocated via sharedOverhead, sharedNamespaces, or sharedLabelNames
  totalCost: 0.059374966          // sum of all costs
}

Caching Overview

Kubecost implements a two-layer caching system for cost allocation metrics.

First, the unaggregated cost model is pre-cached for commonly used time windows, 1 and 2 days by default. This data is refreshed every ~5 minutes.

Longer time windows, 120 days by default, are part of an ETL pipeline that stores cost by day for each workload. This pipeline is updated approximately ~10 mins. On update, only the latest day is rebuilt to reduce the load on the underlying data store. Currently, this ETL pipeline is stored in memory and is built any time the pod restarts. ETL is built with daily granularity for UI improved performance. Daily aggregations default to UTC but timezones can be configured with the utcOffset within values.

Returning cached data from either caching layer typically takes < 300ms on medium-sized clusters. To guarantee you bypass both caches, you can set etl=false and disableCache=false.

Have questions? Email us at [email protected].

Edit this doc on GitHub