Skip to content

Latest commit

 

History

History
429 lines (400 loc) · 22.1 KB

allocation.md

File metadata and controls

429 lines (400 loc) · 22.1 KB
Raw

Allocation API

The Allocation API is the preferred way to query for costs and resources allocated to Kubernetes workloads and optionally aggregated by Kubernetes concepts like namespace, controller, and label. Data is served from one of Kubecost's ETL pipelines. The endpoint is available at the URL:

http://<kubecost>/model/allocation

NOTE

Throughout, we use localhost:9090 as the default Kubecost URL, but your Kubecost instance may be exposed by a service or ingress. To reach Kubecost at port 9090, run: kubectl port-forward deployment/kubecost-cost-analyzer -n kubecost 9090

Quick start

Request allocation data for each 24-hour period in the last three days, aggregated by namespace:

$ curl http://localhost:9090/model/allocation \
  -d window=3d \
  -d aggregate=namespace \
  -d accumulate=false \
  -d shareIdle=false \
  -G

{
  "code": 200,
  "data": [
    {
      "__idle__": { ... },
      "default": { ... },
      "kube-system": { ... },
      "kubecost": { ... }
    },
    {
      "__idle__": { ... },
      "default": { ... },
      "kube-system": { ... },
      "kubecost": { ... }
    },
    {
      "__idle__": { ... },
      "default": { ... },
      "kube-system": { ... },
      "kubecost": { ... }
    },
    {
      "__idle__": { ... },
      "default": { ... },
      "kube-system": { ... },
      "kubecost": { ... }
    }
  ]
}

Note: querying for "3d" will likely return a range of four sets because the queried range will overlap with four precomputed 24-hour sets, each aligned to the configured timezone. For instance, querying "3d" on 2021/01/04T12:00:00 will return:

  • 2021/01/04 00:00:00 until 2021/01/04T12:00:00 (now)
  • 2021/01/03 00:00:00 until 2021/01/04 00:00:00
  • 2021/01/02 00:00:00 until 2021/01/03 00:00:00
  • 2021/01/01 00:00:00 until 2021/01/02 00:00:00

See Querying for the full list of arguments and Examples for more example queries.

Allocation schema (version 1.76)

Field Description
name Name of each relevant Kubernetes concept described by the allocation, delimited by slashes, e.g. "cluster/node/namespace/pod/container"
properties Map of name-to-value for all relevant property fields, including: cluster, node, namespace, controller, controllerKind, pod, container, labels, annotation, etc. Note: Prometheus only supports underscores (_) in label names. Dashes (-) and dots (.), while supported by Kubernetes, will be translated to underscores by Prometheus. This may cause the merging of labels, which could result in aggregated costs being charged to a single label.
window Period of time over which the allocation is defined.
start Precise starting time of the allocation. By definition must be within the window.
end Precise ending time of the allocation. By definition must be within the window.
minutes Number of minutes running; i.e. the minutes from start until end.
cpuCores Average number of CPU cores allocated while running.
cpuCoreRequestAverage Average number of CPU cores requested while running.
cpuCoreUsageAverage Average number of CPU cores used while running.
cpuCoreHours Cumulative CPU core-hours allocated.
cpuCost Cumulative cost of allocated CPU core-hours.
cpuEfficiency Ratio of cpuCoreUsageAverage-to-cpuCoreRequestAverage, meant to represent the fraction of requested resources that were used.
gpuHours Cumulative GPU-hours allocated.
gpuCost Cumulative cost of allocated GPU-hours.
networkCost Cumulative cost of network usage.
pvBytes Average number of bytes of PersistentVolumes allocated while running.
pvByteHours Cumulative PersistentVolume byte-hours allocated.
pvCost Cumulative cost of allocated PersistentVolume byte-hours.
ramBytes Average number of RAM bytes allocated while running.
ramByteRequestAverage Average number of RAM bytes allocated while running.
ramByteUsageAverage Average number of RAM bytes used while running.
ramByteHours Cumulative RAM byte-hours allocated.
ramCost Cumulative cost of allocated RAM byte-hours.
ramEfficiency Ratio of ramByteUsageAverage-to-ramByteRequestAverage, meant to represent the fraction of requested resources that were used.
sharedCost Cumulative cost of shared resources, including shared namespaces, shared labels, shared overhead.
externalCost Cumulative cost of external resources.
totalCost Total cumulative cost
totalEfficiency Cost-weighted average of cpuEfficiency and ramEfficiency. In equation form: ((cpuEfficiency * cpuCost) + (ramEfficiency * ramCost)) / (cpuCost + ramCost)
rawAllocationOnly Object with fields cpuCoreUsageMax and ramByteUsageMax, which are the maximum usages in the window for the Allocation. If the Allocation query is aggregated or accumulated, this object will be null because the meaning of maximum is ambiguous in these situations. Consider aggregating by namespace: should the maximum be the maximum of each Allocation individually, or the maximum combined usage of all Allocations (at any point in time in the window) in the namespace?

Example allocation

Here is an example allocation for the cost-model container in a pod in Kubecost's kubecost-cost-analyzer deployment, deployed into the kubecost namespace. The properties object describes that, as well as the cluster, node, services, labels, and annotations related to this allocation. Notice that this allocation ran for 10 hours within the given window, using the resources described by their respective values at a cost dictated by the node on which it ran.

{
  "name": "cluster-one/gke-niko-pool-2-9182dfa7-okb2/kubecost/kubecost-cost-analyzer-94dc86fc-lwvrm/cost-model",
  "properties": {
    "annotation": {},
    "cluster": "cluster-one",
    "container": "cost-model",
    "controller": "kubecost-cost-analyzer",
    "controllerKind": "deployment",
    "label": {
      "app": "cost-analyzer",
      "app_kubernetes_io_instance": "kubecost",
      "app_kubernetes_io_name": "cost-analyzer",
      "name": "kubecost",
      "pod_template_hash": "94dc86fc"
    },
    "namespace": "kubecost",
    "node": "gke-niko-pool-2-9182dfa7-okb2",
    "pod": "kubecost-cost-analyzer-94dc86fc-lwvrm",
    "service": [
      "kubecost-frontend",
      "kubecost-cost-analyzer"
    ]
  },
  "window": {
    "start": "2021-03-11T00:00:00-07:00",
    "end": "2021-03-12T00:00:00-07:00"
  },
  "start": "2021-03-11T07:00:00Z",
  "end": "2021-03-11T17:00:00Z",
  "minutes": 600,
  "cpuCores": 0.200399,
  "cpuCoreRequestAverage": 0.2,
  "cpuCoreUsageAverage": 0.004317,
  "cpuCoreHours": 2.00399,
  "cpuCost": 0.044344,
  "cpuEfficiency": 0.021583,
  "gpuHours": 0,
  "gpuCost": 0,
  "networkCost": 1.3e-05,
  "pvBytes": 11453246122.666668,
  "pvByteHours": 114532461226.66667,
  "pvCost": 0.005845,
  "ramBytes": 59791087.387687,
  "ramByteRequestAverage": 57671680,
  "ramByteUsageAverage": 52687168.319468,
  "ramByteHours": 597910873.876871,
  "ramCost": 0.001652,
  "ramEfficiency": 0.913571,
  "sharedCost": 0,
  "externalCost": 0,
  "totalCost": 0.051853,
  "totalEfficiency": 0.053611
}

Special types of allocation

  • __idle__ refers to resources on a cluster that were not dedicated to a Kubernetes object (e.g. unused CPU core-hours on a node). An idle resource can be shared (proportionally or evenly) with the other allocations from the same cluster. (See the argument shareIdle.)
  • __unallocated__ refers to aggregated allocations without the selected aggregate field; e.g. aggregating by label:app might produce an __unallocated__ allocation composed of allocations without the app label.
  • __unmounted__ (or "Unmounted PVs") refers to the resources used by PersistentVolumes that aren't mounted to a Pod using a PVC, and thus cannot be allocated to a Pod.

Querying

GET /model/allocation
Argument Default Description
window (required) — Duration of time over which to query. Accepts: words like today, week, month, yesterday, lastweek, lastmonth; durations like 30m, 12h, 7d; comma-separated RFC3339 date pairs like 2021-01-02T15:04:05Z,2021-02-02T15:04:05Z; comma-separated unix timestamp (seconds) pairs like 1578002645,1580681045.
aggregate Field by which to aggregate the results. Accepts: cluster, namespace, controllerKind, controller, service, node, pod, label:<name>, and annotation:<name>. Also accepts comma-separated lists for multi-aggregation, like namespace,label:app.
accumulate false If true, sum the entire range of sets into a single set.
idle true If true, include idle cost (i.e. the cost of the un-allocated assets) as its own allocation. (See special types of allocation.)
external false If true, include external costs in each allocation.
filterClusters Comma-separated list of clusters to match; e.g. cluster-one,cluster-two will return results from only those two clusters.
filterNodes Comma-separated list of nodes to match; e.g. node-one,node-two will return results from only those two nodes.
filterNamespaces Comma-separated list of namespaces to match; e.g. namespace-one,namespace-two will return results from only those two namespaces.
filterControllerKinds Comma-separated list of controller kinds to match; e.g. deployment,job will return results with only those two controller kinds.
filterControllers Comma-separated list of controllers to match; e.g. deployment-one,statefulset-two will return results from only those two controllers.
filterPods Comma-separated list of pods to match; e.g. pod-one,pod-two will return results from only those two pods.
filterAnnotations Comma-separated list of annotations to match; e.g. name:annotation-one,name:annotation-two will return results with either of those two annotation key-value-pairs.
filterLabels Comma-separated list of annotations to match; e.g. app:cost-analyzer, app:prometheus will return results with either of those two label key-value-pairs.
filterServices Comma-separated list of services to match; e.g. frontend-one,frontend-two will return results with either of those two services.
format Set to csv to download an accumulated version of the allocation results in CSV format. By default, results will be in JSON format.
shareIdle false If true, idle cost is allocated proportionally across all non-idle allocations, per-resource. That is, idle CPU cost is shared with each non-idle allocation's CPU cost, according to the percentage of the total CPU cost represented.
splitIdle false If true, and shareIdle == false Idle Allocations are created on a per cluster or per node basis rather than being aggregated into a single "_idle_" allocation.
idleByNode false If true, idle allocations are created on a per node basis. Which will result in different values when shared and more idle allocations when split.
reconcile true If true pulls data from the Assets cache and corrects prices of Allocations according to their related Assets. The corrections from this process are stored in each cost categories cost adjustment field. If the integration with your cloud provider's billing data has been set up, this will result in the most accurate costs for Allocations.
shareTenancyCosts true If true, share the cost of cluster overhead assets such as cluster management costs and node attached volumes across tenants of those resources. Results are added to the sharedCost field.
shareNamespaces Comma-separated list of namespaces to share; e.g. kube-system, kubecost will share the costs of those two namespaces with the remaining non-idle, unshared allocations.
shareLabels Comma-separated list of labels to share; e.g. env:staging, app:test will share the costs of those two label values with the remaining non-idle, unshared allocations.
shareCost 0.0 Floating-point value representing a monthly cost to share with the remaining non-idle, unshared allocations; e.g. 30.42 ($1.00/day == $30.42/month) for the query yesterday (1 day) will split and distribute exactly $1.00 across the allocations.
shareSplit weighted Determines how to split shared costs among non-idle, unshared allocations. By default, the split will be weighted; i.e. proportional to the cost of the allocation, relative to the total. The other option is even; i.e. each allocation gets an equal portion of the shared cost.

Query examples

Allocation data for today, unaggregated:

$ curl http://localhost:9090/model/allocation \
  -d window=today \
  -G

{
  "code": 200,
  "data": [
    {
      "__idle__": { ... },
      "cluster-one/gke-niko-pool-2-9182dfa7-okb2/kubecost/kubecost-cost-analyzer-94dc86fc-lwvrm/cost-model": { ... },
      "cluster-one/gke-niko-pool-2-9182dfa7-okb2/kubecost/kubecost-cost-analyzer-94dc86fc-lwvrm/cost-analyzer-frontend": { ... },
      "cluster-one/gke-niko-pool-2-9182dfa7-okb2/kubecost/kubecost-grafana-6df5cc66b6-dzszt/grafana": { ... }
    }
  ]
}

Allocation data for last week, per day, aggregated by cluster:

$ curl http://localhost:9090/model/allocation \
  -d window=lastweek \
  -d aggregate=cluster \
  -G

{
  "code": 200,
  "data": [
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "__idle__": { ... },
      "cluster-one": { ... },
      "cluster-two": { ... }
    }
  ]
}

Allocation data for the last 30 days, aggregated by the "app" label, sharing idle allocation, sharing allocations from two namespaces, sharing $100/mo in overhead, and accumulated into one allocation for the entire window:

$ curl http://localhost:9090/model/allocation \
  -d window=30d \
  -d aggregate=label:app \
  -d accumulate=true \
  -d shareIdle=weighted \
  -d shareNamespaces=kube-system,kubecost
  -d shareCost=100
  -G

{
  "code": 200,
  "data": [
    {
      "__unallocated__": { ... },
      "app=redis": { ... },
      "app=cost-analyzer": { ... },
      "app=prometheus": { ... },
      "app=grafana": { ... },
      "app=nginx": { ... },
      "app=helm": { ... }
    }
  ]
}

Allocation data for 2021-03-10T00:00:00 to 2021-03-11T00:00:00 (i.e. 24h), multi-aggregated by namespace and the "app" label, filtering by properties.cluster == "cluster-one", and accumulated into one allocation for the entire window.

$ curl http://localhost:9090/model/allocation \
  -d window=2021-03-10T00:00:00Z,2021-03-11T00:00:00Z \
  -d aggregate=namespace,label:app \
  -d accumulate=true \
  -d filterClusters=cluster-one \
  -G

{
  "code": 200,
  "data": [
    {
      "default/app=redis": { ... },
      "kubecost/app=cost-analyzer": { ... },
      "kubecost/app=prometheus": { ... },
      "kubecost/app=grafana": { ... },
      "kubecost/app=prometheus": { ... },
      "kube-system/app=helm": { ... }
    }
  ]
}

Allocation of Asset Costs:

Both the reconcile and shareTenancyCosts flags start query-time processes that distribute the costs of Assets to Allocations related to them. For the reconcile flag, these connections can be straightforward like the connection between a node Asset and an Allocation where the CPU, GPU, and RAM usage can be used to distribute a proportion of the node's cost to the Allocations that run on it. For Assets and Allocations where the connection is less well-defined, such as network Assets we have opted for a method of distributing the cost that we call Distribution by Usage Hours.

Distribution by Usage Hours takes the usage of the windows (start time and end time) of an Asset and all the Allocations connected to it and finds the number of hours that both the Allocation and Asset were running. The number of hours for each Allocation related to an Asset is called Alloc_Usage_Hours. The sum of all Alloc_Usage_Hours for a single Assets is Total_Usage_Hours. With these values, an Assets cost is distributed to each connected Allocation using the formula Asset_Cost * Alloc_Usage_Hours/Total_Usage_Hours. Depending on the Asset type an Allocation can receive proportions of multiple Asset Costs.

Asset types that use this distribution method include:

  • Network (reconcile): When the network pod is not enabled cost is distributed by usage hours. If the network pod is enabled cost is distributed to Allocations proportionally to usage.
  • Load Balancer (reconcile)
  • Cluster Management (shareTenancyCosts)
  • Attached disks (shareTenancyCosts): Does not include PVs, which are handled by reconcile

Querying on-demand (experimental)

⚠️ WARNING

Querying on-demand with high resolution for long windows can cause serious Prometheus performance issues, including OOM errors. Start with short windows (1d or less) and proceed with caution.

Computing allocation data on-demand allows for greater flexibility with respect to step size and accuracy-versus-performance. (See resolution and error bounds for details.) Unlike the standard endpoint, which can only serve results from precomputed sets with predefined step sizes (e.g. 24h aligned to the UTC timezone), asking for a "7d" query will almost certainly result in 8 sets, including "today" and the final set, which might span 6.5d-7.5d ago. With this endpoint, however, you will be computing everything on-demand, so "7d" will return exactly seven days of data, starting at the moment the query is received. (You can still use window keywords like "today" and "lastweek", of course, which should align perfectly with the same queries of the standard ETL-driven endpoint.)

GET /model/allocation/compute
Argument Default Description
window (required) — Duration of time over which to query. Accepts: words like today, week, month, yesterday, lastweek, lastmonth; durations like 30m, 12h, 7d; RFC3339 date pairs like 2021-01-02T15:04:05Z,2021-02-02T15:04:05Z; unix timestamps like 1578002645,1580681045.
resolution 1m Duration to use as resolution in Prometheus queries. Smaller values (i.e. higher resolutions) will provide better accuracy, but worse performance (i.e. slower query time, higher memory use). Larger values (i.e. lower resolutions) will perform better, but at the expense of lower accuracy for short-running workloads. (See error bounds for details.)
step window Duration of a single allocation set. If unspecified, this defaults to the window, so that you receive exactly one set for the entire window. If specified, it works chronologically backward, querying in durations of step until the full window is covered.
aggregate Field by which to aggregate the results. Accepts: cluster, namespace, controllerKind, controller, service, label:<name>, and annotation:<name>. Also accepts comma-separated lists for multi-aggregation, like namespace,label:app.
accumulate false If true, sum the entire range of sets into a single set.

On-demand query examples

Allocation data for the last 60m, in steps of 10m, with resolution 1m, aggregated by cluster.

$ curl http://localhost:9090/model/allocation/compute \
  -d window=60m \
  -d step=10m \
  -d resolution=1m \
  -d aggregate=cluster \
  -d accumulate=false \
  -G

{
  "code": 200,
  "data": [
    {
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "cluster-one": { ... },
      "cluster-two": { ... }
    },
    {
      "cluster-one": { ... },
      "cluster-two": { ... }
    }
  ]
}

Allocation data for the last 9d, in steps of 3d, with a 10m resolution, aggregated by namespace.

$ curl http://localhost:9090/model/allocation/compute \
  -d window=9d \
  -d step=3d \
  -d resolution=10m
  -d aggregate=namespace \
  -d accumulate=false \
  -G

{
  "code": 200,
  "data": [
    {
      "default": { ... },
      "kubecost": { ... },
      "kube-system": { ... }
    },
    {
      "default": { ... },
      "kubecost": { ... },
      "kube-system": { ... }
    },
    {
      "default": { ... },
      "kubecost": { ... },
      "kube-system": { ... }
    }
  ]
}

Theoretical error bounds

Tuning the resolution parameter allows the querier to make tradeoffs between accuracy and performance. For long-running pods (>1d) resolution can be tuned aggressively low (>10m) with relatively little effect on accuracy. However, even modestly low resolutions (5m) can result in significant accuracy degradation for short-running pods (<1h).

Here, we provide theoretical error bounds for different resolution values given pods of differing running durations. The tuple represents lower- and upper-bounds for accuracy as a percentage of the actual value. For example:

  • 1.00, 1.00 means that results should always be accurate to less than 0.5% error
  • 0.83, 1.00 means that results should never be high by more than 0.5% error, but could be low by as much as 17% error
  • -1.00, 10.00 means that the result could be as high as 1000% error (e.g. 30s pod being counted for 5m) or the pod could be missed altogether, i.e. -100% error.
resolution 30s pod 5m pod 1h pod 1d pod 7d pod
1m -1.00, 2.00 0.80, 1.00 0.98, 1.00 1.00, 1.00 1.00, 1.00
2m -1.00, 4.00 0.80, 1.20 0.97, 1.00 1.00, 1.00 1.00, 1.00
5m -1.00, 10.00 -1.00, 1.00 0.92, 1.00 1.00, 1.00 1.00, 1.00
10m -1.00, 20.00 -1.00, 2.00 0.83, 1.00 0.99, 1.00 1.00, 1.00
30m -1.00, 60.00 -1.00, 6.00 0.50, 1.00 0.98, 1.00 1.00, 1.00
60m -1.00, 120.00 -1.00, 12.00 -1.00, 1.00 0.96, 1.00 0.99, 1.00

Edit this doc on Github