kostanza

README

Kostanza

Kostanza is a service for emitting approximate cost metrics for services running in a kubernetes cluster. It works by:

• Periodically listing all pods.
• Determining what nodes those pods are running on.
• Using the pods cpu or memory utilization of the node to increment a cost counter.

Kostanza's pricing units are in millionths of a cent. The choice was made since the intended sink for these metrics is a prometheus counter, which is based on integers. Units of cents and thousandths of a cent were eschewed in favor of increased granularity due to truncation. Kostanza may produce inaccurate results for pods that use very little resources in conjunction with rapid polling intervals.

Configuration

Pricing

Kostanza does not make assumptions about node costs, though this may be obtained through cloud APIs in the future. Instead, it uses a user-provided configuration file that allows admins to map node labels to appropriate metadata regarding allocatable memory, cpu, and hourly instance cost. The source-ordering of the .Pricing.Entries array in the examples below is important: the first entry whose Labels collection is a subset of the labels for a given node will be used. Thus, you should generally order your table from specific to general.

Note: we may use a simple heuristic of sorting entries by the number of labels you specify.

Mapping

Kostanza does not make assumptions about the dimensions you want to use for the cost metrics it emits. It allows you to map these from resource labels and annotations. For example, organizations that use the label service may wish to use the jsonPath expression {.Pod.ObjectMeta.Labels.service}.

It's important to set to a default value since prometheus, the canonical (and to date, only) exporter will complain about the cardinality of metrics when a key is left empty.

Note: the property names are based on the CostItem struct contained within the package, which references kubernetes client-go, see https://github.com/kubernetes/kubernetes/blob/master/pkg/apis/core/types.go

{
  "Pricing": {
    "Entries": [
      {
        "Labels": {
          "beta.kubernetes.io/instance-type": "n1-standard-16"
        },
        "HourlyMemoryByteCostMicroCents": 0.00043406151235103607,
        "HourlyMilliCPUCostMicroCents": 3477.21
      },
      {
        "Labels": {
          "kubernetes.io/hostname": "minikube"
        },
        "HourlyMemoryByteCostMicroCents": 0.0002,
        "HourlyMilliCPUCostMicroCents": 2000
      }
    ]
  },
  "Mapping": {
    "Entries": [
      {
        "Destination": "service",
        "Source": "{.Pod.ObjectMeta.Labels.service}",
        "Default": "unknown"
      },
      {
        "Destination": "component",
        "Source": "{.Pod.ObjectMeta.Labels.component}",
        "Default": "unknown"
      },
      {
        "Destination": "node_instance_type",
        "Source": "{.Node.ObjectMeta.Labels['beta.kubernetes.io/instance-type']}",
        "Default": "unknown"
      }
      {
        "Destination": "kind",
        "Source": "{.Kind}",
        "Default": "unknown"
      },
      {
        "Destination": "strategy",
        "Source": "{.Strategy}",
        "Default": "unknown"
      }
    ]
  }
}

Strategies

Kostanza currently emits metrics according to two strategies by default:

• WeightedPricingStrategy
• NodePricingStrategy

WeightedPricingStrategy

The WeightedPricingStrategy strategy operates as follows:

• First, add up all cpu and memory resource requests on all nodes.
• Next, for each pod, determine the fraction of memory and cpu it has requested on its respective node. The pod's cost is based on the average of these two fractions.

This strategy does punish pods for being scheduled to fresh nodes - the choice is deliberate and is intended to capture costs incurred by services that may trigger frequent scale ups without having a chance to benefit from bin-packing additional pods.

NodePricingStrategy

The NodePricingStrategy is intended to emit baseline cost metrics for your nodes, i.e. to provide a total cost for your cluster. This can be useful in visualizations where one wishes to graph pod costs as a fraction of total cost. The mapping configuration presented earlier provides an example of how one might configure your mapper based on nascent standardized labels (e.g. beta.kubernetes.io/instance-type).

Metric Dimensions

All strategies share the same metrics and metric dimensions. This means, for example, that the NodePricingStrategy is likely to emit defaults for metrics only relevant to pods. This is a deliberate design choice at the moment - rather than having multiple measures or metrics, we have a single metric containing a superset of cost dimensions whether they apply to a particular strategy or not.

Exporters

Kostanza exports cost data in two ways: as prometheus metrics, and to optionally to google pubsub. The pubsub export is largely intended to leave the primary kostanza agent decoupled from any particular storage backend - however, a secondary aggregate command is available as part of the kostanza binary to pipe these pubsub message into an automatically-provisioned BigQuery table.

Prometheus Exporter

Intended for real time data and convenient calculation of costs as rates. Note that the cost counter will reset and it's absolute value should generally not be relied on - you'll want to take use the PromQL rate function to express costs as rates of change over time.

Pubsub Exporter and the Aggregate Subcommand

For longer term analysis, kostanza allows for publishing messages to a pubsub queue where they can be consumed and funneled into a data warehouse such as GCP. Kostanza even offers a baked-in aggregate subcommand for the purpose. The payload of these messages is default json encoding of the coster.CostData structure:

type CostData struct {
	// The kind of cost figure represented.
	Kind ResourceCostKind
	// The strategy the yielded this CostItem.
	Strategy string
	// The value in microcents that it costs.
	Value int64
	// Additional dimensions associated with the cost.
	Dimensions map[string]string
	// The interval for which this metric was created.
	EndTime time.Time
}

The aggregate subcommand will happily consume messages that adhere to this general spefication and published to subscription specified by the pubsub-subscription startup argument. This may be useful if you wish to incorporate data from systems outside of kubernetes.

Auto-provisioning

When the aggregate subcommand starts up, it will use the mapping defined in your configuration file to automatically provision a BigQuery table in the bigquery-project and bigquery-dataset, named according to the bigquery-table CLI flag.

For convenience, we parse the mapping specified in your configuration and automatically attempt to create Dimension_DestinationName columns for each of your mapping destinations. This can be incredibly convenient for bootstrapping or local development but is not a full data migration by any means: if the BigQuery table already exists, no attempt will be infer schema or data migrations on your behalf. Generally, a best practice may be to create an entirely new table if a new dimension is required.