cost-model/pkg/costmodel/allocation.go

package costmodel

import (
	"fmt"
	"math"
	"strconv"
	"strings"
	"time"

	"github.com/kubecost/cost-model/pkg/util/timeutil"

	"github.com/kubecost/cost-model/pkg/cloud"
	"github.com/kubecost/cost-model/pkg/env"
	"github.com/kubecost/cost-model/pkg/kubecost"
	"github.com/kubecost/cost-model/pkg/log"
	"github.com/kubecost/cost-model/pkg/prom"
	"k8s.io/apimachinery/pkg/labels"
	"k8s.io/klog"
)

const (
	queryFmtPods              = `avg(kube_pod_container_status_running{}) by (pod, namespace, %s)[%s:%s]%s`
	queryFmtRAMBytesAllocated = `avg(avg_over_time(container_memory_allocation_bytes{container!="", container!="POD", node!=""}[%s]%s)) by (container, pod, namespace, node, %s, provider_id)`
	queryFmtRAMRequests       = `avg(avg_over_time(kube_pod_container_resource_requests{resource="memory", unit="byte", container!="", container!="POD", node!=""}[%s]%s)) by (container, pod, namespace, node, %s)`
	queryFmtRAMUsageAvg       = `avg(avg_over_time(container_memory_working_set_bytes{container!="", container_name!="POD", container!="POD"}[%s]%s)) by (container_name, container, pod_name, pod, namespace, instance, %s)`
	queryFmtRAMUsageMax       = `max(max_over_time(container_memory_working_set_bytes{container!="", container_name!="POD", container!="POD"}[%s]%s)) by (container_name, container, pod_name, pod, namespace, instance, %s)`
	queryFmtCPUCoresAllocated = `avg(avg_over_time(container_cpu_allocation{container!="", container!="POD", node!=""}[%s]%s)) by (container, pod, namespace, node, %s)`
	queryFmtCPURequests       = `avg(avg_over_time(kube_pod_container_resource_requests{resource="cpu", unit="core", container!="", container!="POD", node!=""}[%s]%s)) by (container, pod, namespace, node, %s)`
	queryFmtCPUUsageAvg       = `avg(rate(container_cpu_usage_seconds_total{container!="", container_name!="POD", container!="POD"}[%s]%s)) by (container_name, container, pod_name, pod, namespace, instance, %s)`

	// This query could be written without the recording rule
	// "kubecost_savings_container_cpu_usage_seconds", but we should
	// only do that when we're ready to incur the performance tradeoffs
	// with subqueries which would probably be in the world of hourly
	// ETL.
	//
	// See PromQL subquery documentation for a rate example:
	// https://prometheus.io/blog/2019/01/28/subquery-support/#examples
	queryFmtCPUUsageMax           = `max(max_over_time(kubecost_savings_container_cpu_usage_seconds[%s]%s)) by (container_name, pod_name, namespace, instance, %s)`
	queryFmtGPUsRequested         = `avg(avg_over_time(kube_pod_container_resource_requests{resource="nvidia_com_gpu", container!="",container!="POD", node!=""}[%s]%s)) by (container, pod, namespace, node, %s)`
	queryFmtGPUsAllocated         = `avg(avg_over_time(container_gpu_allocation{container!="", container!="POD", node!=""}[%s]%s)) by (container, pod, namespace, node, %s)`
	queryFmtNodeCostPerCPUHr      = `avg(avg_over_time(node_cpu_hourly_cost[%s]%s)) by (node, %s, instance_type, provider_id)`
	queryFmtNodeCostPerRAMGiBHr   = `avg(avg_over_time(node_ram_hourly_cost[%s]%s)) by (node, %s, instance_type, provider_id)`
	queryFmtNodeCostPerGPUHr      = `avg(avg_over_time(node_gpu_hourly_cost[%s]%s)) by (node, %s, instance_type, provider_id)`
	queryFmtNodeIsSpot            = `avg_over_time(kubecost_node_is_spot[%s]%s)`
	queryFmtPVCInfo               = `avg(kube_persistentvolumeclaim_info{volumename != ""}) by (persistentvolumeclaim, storageclass, volumename, namespace, %s)[%s:%s]%s`
	queryFmtPVBytes               = `avg(avg_over_time(kube_persistentvolume_capacity_bytes[%s]%s)) by (persistentvolume, %s)`
	queryFmtPodPVCAllocation      = `avg(avg_over_time(pod_pvc_allocation[%s]%s)) by (persistentvolume, persistentvolumeclaim, pod, namespace, %s)`
	queryFmtPVCBytesRequested     = `avg(avg_over_time(kube_persistentvolumeclaim_resource_requests_storage_bytes{}[%s]%s)) by (persistentvolumeclaim, namespace, %s)`
	queryFmtPVCostPerGiBHour      = `avg(avg_over_time(pv_hourly_cost[%s]%s)) by (volumename, %s)`
	queryFmtNetZoneGiB            = `sum(increase(kubecost_pod_network_egress_bytes_total{internet="false", sameZone="false", sameRegion="true"}[%s]%s)) by (pod_name, namespace, %s) / 1024 / 1024 / 1024`
	queryFmtNetZoneCostPerGiB     = `avg(avg_over_time(kubecost_network_zone_egress_cost{}[%s]%s)) by (%s)`
	queryFmtNetRegionGiB          = `sum(increase(kubecost_pod_network_egress_bytes_total{internet="false", sameZone="false", sameRegion="false"}[%s]%s)) by (pod_name, namespace, %s) / 1024 / 1024 / 1024`
	queryFmtNetRegionCostPerGiB   = `avg(avg_over_time(kubecost_network_region_egress_cost{}[%s]%s)) by (%s)`
	queryFmtNetInternetGiB        = `sum(increase(kubecost_pod_network_egress_bytes_total{internet="true"}[%s]%s)) by (pod_name, namespace, %s) / 1024 / 1024 / 1024`
	queryFmtNetInternetCostPerGiB = `avg(avg_over_time(kubecost_network_internet_egress_cost{}[%s]%s)) by (%s)`
	queryFmtNetReceiveBytes       = `sum(increase(container_network_receive_bytes_total{pod_name!=""}[%s]%s)) by (pod_name, namespace, %s)`
	queryFmtNetTransferBytes      = `sum(increase(container_network_transmit_bytes_total{pod_name!=""}[%s]%s)) by (pod_name, namespace, %s)`
	queryFmtNamespaceLabels       = `avg_over_time(kube_namespace_labels[%s]%s)`
	queryFmtNamespaceAnnotations  = `avg_over_time(kube_namespace_annotations[%s]%s)`
	queryFmtPodLabels             = `avg_over_time(kube_pod_labels[%s]%s)`
	queryFmtPodAnnotations        = `avg_over_time(kube_pod_annotations[%s]%s)`
	queryFmtServiceLabels         = `avg_over_time(service_selector_labels[%s]%s)`
	queryFmtDeploymentLabels      = `avg_over_time(deployment_match_labels[%s]%s)`
	queryFmtStatefulSetLabels     = `avg_over_time(statefulSet_match_labels[%s]%s)`
	queryFmtDaemonSetLabels       = `sum(avg_over_time(kube_pod_owner{owner_kind="DaemonSet"}[%s]%s)) by (pod, owner_name, namespace, %s)`
	queryFmtJobLabels             = `sum(avg_over_time(kube_pod_owner{owner_kind="Job"}[%s]%s)) by (pod, owner_name, namespace ,%s)`
	queryFmtLBCostPerHr           = `avg(avg_over_time(kubecost_load_balancer_cost[%s]%s)) by (namespace, service_name, %s)`
	queryFmtLBActiveMins          = `count(kubecost_load_balancer_cost) by (namespace, service_name, %s)[%s:%s]%s`
)

// This is a bit of a hack to work around garbage data from cadvisor
// The maximum CPU count on AWS is 96. Standard_M192idms_v2 on azure allows 192. GCP offers 224 in n2ds. Many improvements here are possible to get lower bounnds, but 96 seems reasonable now.
// Ideally you cap each pod to the max CPU on its node, but that involves a bit more complexity, as it it would need to be done when allocations joins with asset data.
const MAX_CPU_CAP = 96

// CanCompute should return true if CostModel can act as a valid source for the
// given time range. In the case of CostModel we want to attempt to compute as
// long as the range starts in the past. If the CostModel ends up not having
// data to match, that's okay, and should be communicated with an error
// response from ComputeAllocation.
func (cm *CostModel) CanCompute(start, end time.Time) bool {
	return start.Before(time.Now())
}

// Name returns the name of the Source
func (cm *CostModel) Name() string {
	return "CostModel"
}

// ComputeAllocation uses the CostModel instance to compute an AllocationSet
// for the window defined by the given start and end times. The Allocations
// returned are unaggregated (i.e. down to the container level).
func (cm *CostModel) ComputeAllocation(start, end time.Time, resolution time.Duration) (*kubecost.AllocationSet, error) {
	// 1. Build out Pod map from resolution-tuned, batched Pod start/end query
	// 2. Run and apply the results of the remaining queries to
	// 3. Build out AllocationSet from completed Pod map

	// Create a window spanning the requested query
	window := kubecost.NewWindow(&start, &end)

	// Create an empty AllocationSet. For safety, in the case of an error, we
	// should prefer to return this empty set with the error. (In the case of
	// no error, of course we populate the set and return it.)
	allocSet := kubecost.NewAllocationSet(start, end)

	// (1) Build out Pod map

	// Build out a map of Allocations as a mapping from pod-to-container-to-
	// underlying-Allocation instance, starting with (start, end) so that we
	// begin with minutes, from which we compute resource allocation and cost
	// totals from measured rate data.
	podMap := map[podKey]*Pod{}

	// clusterStarts and clusterEnds record the earliest start and latest end
	// times, respectively, on a cluster-basis. These are used for unmounted
	// PVs and other "virtual" Allocations so that minutes are maximally
	// accurate during start-up or spin-down of a cluster
	clusterStart := map[string]time.Time{}
	clusterEnd := map[string]time.Time{}

	cm.buildPodMap(window, resolution, env.GetETLMaxBatchDuration(), podMap, clusterStart, clusterEnd)

	// (2) Run and apply remaining queries

	// Convert window (start, end) to (duration, offset) for querying Prometheus,
	// including handling Thanos offset
	durStr, offStr, err := window.DurationOffsetForPrometheus()
	if err != nil {
		// Negative duration, so return empty set
		return allocSet, nil
	}

	// Convert resolution duration to a query-ready string
	resStr := timeutil.DurationString(resolution)

	ctx := prom.NewContext(cm.PrometheusClient)

	queryRAMBytesAllocated := fmt.Sprintf(queryFmtRAMBytesAllocated, durStr, offStr, env.GetPromClusterLabel())
	resChRAMBytesAllocated := ctx.Query(queryRAMBytesAllocated)

	queryRAMRequests := fmt.Sprintf(queryFmtRAMRequests, durStr, offStr, env.GetPromClusterLabel())
	resChRAMRequests := ctx.Query(queryRAMRequests)

	queryRAMUsageAvg := fmt.Sprintf(queryFmtRAMUsageAvg, durStr, offStr, env.GetPromClusterLabel())
	resChRAMUsageAvg := ctx.Query(queryRAMUsageAvg)

	queryRAMUsageMax := fmt.Sprintf(queryFmtRAMUsageMax, durStr, offStr, env.GetPromClusterLabel())
	resChRAMUsageMax := ctx.Query(queryRAMUsageMax)

	queryCPUCoresAllocated := fmt.Sprintf(queryFmtCPUCoresAllocated, durStr, offStr, env.GetPromClusterLabel())
	resChCPUCoresAllocated := ctx.Query(queryCPUCoresAllocated)

	queryCPURequests := fmt.Sprintf(queryFmtCPURequests, durStr, offStr, env.GetPromClusterLabel())
	resChCPURequests := ctx.Query(queryCPURequests)

	queryCPUUsageAvg := fmt.Sprintf(queryFmtCPUUsageAvg, durStr, offStr, env.GetPromClusterLabel())
	resChCPUUsageAvg := ctx.Query(queryCPUUsageAvg)

	queryCPUUsageMax := fmt.Sprintf(queryFmtCPUUsageMax, durStr, offStr, env.GetPromClusterLabel())
	resChCPUUsageMax := ctx.Query(queryCPUUsageMax)

	queryGPUsRequested := fmt.Sprintf(queryFmtGPUsRequested, durStr, offStr, env.GetPromClusterLabel())
	resChGPUsRequested := ctx.Query(queryGPUsRequested)

	queryGPUsAllocated := fmt.Sprintf(queryFmtGPUsAllocated, durStr, offStr, env.GetPromClusterLabel())
	resChGPUsAllocated := ctx.Query(queryGPUsAllocated)

	queryNodeCostPerCPUHr := fmt.Sprintf(queryFmtNodeCostPerCPUHr, durStr, offStr, env.GetPromClusterLabel())
	resChNodeCostPerCPUHr := ctx.Query(queryNodeCostPerCPUHr)

	queryNodeCostPerRAMGiBHr := fmt.Sprintf(queryFmtNodeCostPerRAMGiBHr, durStr, offStr, env.GetPromClusterLabel())
	resChNodeCostPerRAMGiBHr := ctx.Query(queryNodeCostPerRAMGiBHr)

	queryNodeCostPerGPUHr := fmt.Sprintf(queryFmtNodeCostPerGPUHr, durStr, offStr, env.GetPromClusterLabel())
	resChNodeCostPerGPUHr := ctx.Query(queryNodeCostPerGPUHr)

	queryNodeIsSpot := fmt.Sprintf(queryFmtNodeIsSpot, durStr, offStr)
	resChNodeIsSpot := ctx.Query(queryNodeIsSpot)

	queryPVCInfo := fmt.Sprintf(queryFmtPVCInfo, env.GetPromClusterLabel(), durStr, resStr, offStr)
	resChPVCInfo := ctx.Query(queryPVCInfo)

	queryPVBytes := fmt.Sprintf(queryFmtPVBytes, durStr, offStr, env.GetPromClusterLabel())
	resChPVBytes := ctx.Query(queryPVBytes)

	queryPodPVCAllocation := fmt.Sprintf(queryFmtPodPVCAllocation, durStr, offStr, env.GetPromClusterLabel())
	resChPodPVCAllocation := ctx.Query(queryPodPVCAllocation)

	queryPVCBytesRequested := fmt.Sprintf(queryFmtPVCBytesRequested, durStr, offStr, env.GetPromClusterLabel())
	resChPVCBytesRequested := ctx.Query(queryPVCBytesRequested)

	queryPVCostPerGiBHour := fmt.Sprintf(queryFmtPVCostPerGiBHour, durStr, offStr, env.GetPromClusterLabel())
	resChPVCostPerGiBHour := ctx.Query(queryPVCostPerGiBHour)

	queryNetTransferBytes := fmt.Sprintf(queryFmtNetTransferBytes, durStr, offStr, env.GetPromClusterLabel())
	resChNetTransferBytes := ctx.Query(queryNetTransferBytes)

	queryNetReceiveBytes := fmt.Sprintf(queryFmtNetReceiveBytes, durStr, offStr, env.GetPromClusterLabel())
	resChNetReceiveBytes := ctx.Query(queryNetReceiveBytes)

	queryNetZoneGiB := fmt.Sprintf(queryFmtNetZoneGiB, durStr, offStr, env.GetPromClusterLabel())
	resChNetZoneGiB := ctx.Query(queryNetZoneGiB)

	queryNetZoneCostPerGiB := fmt.Sprintf(queryFmtNetZoneCostPerGiB, durStr, offStr, env.GetPromClusterLabel())
	resChNetZoneCostPerGiB := ctx.Query(queryNetZoneCostPerGiB)

	queryNetRegionGiB := fmt.Sprintf(queryFmtNetRegionGiB, durStr, offStr, env.GetPromClusterLabel())
	resChNetRegionGiB := ctx.Query(queryNetRegionGiB)

	queryNetRegionCostPerGiB := fmt.Sprintf(queryFmtNetRegionCostPerGiB, durStr, offStr, env.GetPromClusterLabel())
	resChNetRegionCostPerGiB := ctx.Query(queryNetRegionCostPerGiB)

	queryNetInternetGiB := fmt.Sprintf(queryFmtNetInternetGiB, durStr, offStr, env.GetPromClusterLabel())
	resChNetInternetGiB := ctx.Query(queryNetInternetGiB)

	queryNetInternetCostPerGiB := fmt.Sprintf(queryFmtNetInternetCostPerGiB, durStr, offStr, env.GetPromClusterLabel())
	resChNetInternetCostPerGiB := ctx.Query(queryNetInternetCostPerGiB)

	queryNamespaceLabels := fmt.Sprintf(queryFmtNamespaceLabels, durStr, offStr)
	resChNamespaceLabels := ctx.Query(queryNamespaceLabels)

	queryNamespaceAnnotations := fmt.Sprintf(queryFmtNamespaceAnnotations, durStr, offStr)
	resChNamespaceAnnotations := ctx.Query(queryNamespaceAnnotations)

	queryPodLabels := fmt.Sprintf(queryFmtPodLabels, durStr, offStr)
	resChPodLabels := ctx.Query(queryPodLabels)

	queryPodAnnotations := fmt.Sprintf(queryFmtPodAnnotations, durStr, offStr)
	resChPodAnnotations := ctx.Query(queryPodAnnotations)

	queryServiceLabels := fmt.Sprintf(queryFmtServiceLabels, durStr, offStr)
	resChServiceLabels := ctx.Query(queryServiceLabels)

	queryDeploymentLabels := fmt.Sprintf(queryFmtDeploymentLabels, durStr, offStr)
	resChDeploymentLabels := ctx.Query(queryDeploymentLabels)

	queryStatefulSetLabels := fmt.Sprintf(queryFmtStatefulSetLabels, durStr, offStr)
	resChStatefulSetLabels := ctx.Query(queryStatefulSetLabels)

	queryDaemonSetLabels := fmt.Sprintf(queryFmtDaemonSetLabels, durStr, offStr, env.GetPromClusterLabel())
	resChDaemonSetLabels := ctx.Query(queryDaemonSetLabels)

	queryJobLabels := fmt.Sprintf(queryFmtJobLabels, durStr, offStr, env.GetPromClusterLabel())
	resChJobLabels := ctx.Query(queryJobLabels)

	queryLBCostPerHr := fmt.Sprintf(queryFmtLBCostPerHr, durStr, offStr, env.GetPromClusterLabel())
	resChLBCostPerHr := ctx.Query(queryLBCostPerHr)

	queryLBActiveMins := fmt.Sprintf(queryFmtLBActiveMins, env.GetPromClusterLabel(), durStr, resStr, offStr)
	resChLBActiveMins := ctx.Query(queryLBActiveMins)

	resCPUCoresAllocated, _ := resChCPUCoresAllocated.Await()
	resCPURequests, _ := resChCPURequests.Await()
	resCPUUsageAvg, _ := resChCPUUsageAvg.Await()
	resCPUUsageMax, _ := resChCPUUsageMax.Await()
	resRAMBytesAllocated, _ := resChRAMBytesAllocated.Await()
	resRAMRequests, _ := resChRAMRequests.Await()
	resRAMUsageAvg, _ := resChRAMUsageAvg.Await()
	resRAMUsageMax, _ := resChRAMUsageMax.Await()
	resGPUsRequested, _ := resChGPUsRequested.Await()
	resGPUsAllocated, _ := resChGPUsAllocated.Await()

	resNodeCostPerCPUHr, _ := resChNodeCostPerCPUHr.Await()
	resNodeCostPerRAMGiBHr, _ := resChNodeCostPerRAMGiBHr.Await()
	resNodeCostPerGPUHr, _ := resChNodeCostPerGPUHr.Await()
	resNodeIsSpot, _ := resChNodeIsSpot.Await()

	resPVBytes, _ := resChPVBytes.Await()
	resPVCostPerGiBHour, _ := resChPVCostPerGiBHour.Await()

	resPVCInfo, _ := resChPVCInfo.Await()
	resPVCBytesRequested, _ := resChPVCBytesRequested.Await()
	resPodPVCAllocation, _ := resChPodPVCAllocation.Await()

	resNetTransferBytes, _ := resChNetTransferBytes.Await()
	resNetReceiveBytes, _ := resChNetReceiveBytes.Await()
	resNetZoneGiB, _ := resChNetZoneGiB.Await()
	resNetZoneCostPerGiB, _ := resChNetZoneCostPerGiB.Await()
	resNetRegionGiB, _ := resChNetRegionGiB.Await()
	resNetRegionCostPerGiB, _ := resChNetRegionCostPerGiB.Await()
	resNetInternetGiB, _ := resChNetInternetGiB.Await()
	resNetInternetCostPerGiB, _ := resChNetInternetCostPerGiB.Await()

	resNamespaceLabels, _ := resChNamespaceLabels.Await()
	resNamespaceAnnotations, _ := resChNamespaceAnnotations.Await()
	resPodLabels, _ := resChPodLabels.Await()
	resPodAnnotations, _ := resChPodAnnotations.Await()
	resServiceLabels, _ := resChServiceLabels.Await()
	resDeploymentLabels, _ := resChDeploymentLabels.Await()
	resStatefulSetLabels, _ := resChStatefulSetLabels.Await()
	resDaemonSetLabels, _ := resChDaemonSetLabels.Await()
	resJobLabels, _ := resChJobLabels.Await()
	resLBCostPerHr, _ := resChLBCostPerHr.Await()
	resLBActiveMins, _ := resChLBActiveMins.Await()

	if ctx.HasErrors() {
		for _, err := range ctx.Errors() {
			log.Errorf("CostModel.ComputeAllocation: %s", err)
		}

		return allocSet, ctx.ErrorCollection()
	}

	// We choose to apply allocation before requests in the cases of RAM and
	// CPU so that we can assert that allocation should always be greater than
	// or equal to request.
	applyCPUCoresAllocated(podMap, resCPUCoresAllocated)
	applyCPUCoresRequested(podMap, resCPURequests)
	applyCPUCoresUsedAvg(podMap, resCPUUsageAvg)
	applyCPUCoresUsedMax(podMap, resCPUUsageMax)
	applyRAMBytesAllocated(podMap, resRAMBytesAllocated)
	applyRAMBytesRequested(podMap, resRAMRequests)
	applyRAMBytesUsedAvg(podMap, resRAMUsageAvg)
	applyRAMBytesUsedMax(podMap, resRAMUsageMax)
	applyGPUsAllocated(podMap, resGPUsRequested, resGPUsAllocated)
	applyNetworkTotals(podMap, resNetTransferBytes, resNetReceiveBytes)
	applyNetworkAllocation(podMap, resNetZoneGiB, resNetZoneCostPerGiB)
	applyNetworkAllocation(podMap, resNetRegionGiB, resNetRegionCostPerGiB)
	applyNetworkAllocation(podMap, resNetInternetGiB, resNetInternetCostPerGiB)

	namespaceLabels := resToNamespaceLabels(resNamespaceLabels)
	podLabels := resToPodLabels(resPodLabels)
	namespaceAnnotations := resToNamespaceAnnotations(resNamespaceAnnotations)
	podAnnotations := resToPodAnnotations(resPodAnnotations)
	applyLabels(podMap, namespaceLabels, podLabels)
	applyAnnotations(podMap, namespaceAnnotations, podAnnotations)

	serviceLabels := getServiceLabels(resServiceLabels)
	allocsByService := map[serviceKey][]*kubecost.Allocation{}
	applyServicesToPods(podMap, podLabels, allocsByService, serviceLabels)

	podDeploymentMap := labelsToPodControllerMap(podLabels, resToDeploymentLabels(resDeploymentLabels))
	podStatefulSetMap := labelsToPodControllerMap(podLabels, resToStatefulSetLabels(resStatefulSetLabels))
	podDaemonSetMap := resToPodDaemonSetMap(resDaemonSetLabels)
	podJobMap := resToPodJobMap(resJobLabels)
	applyControllersToPods(podMap, podDeploymentMap)
	applyControllersToPods(podMap, podStatefulSetMap)
	applyControllersToPods(podMap, podDaemonSetMap)
	applyControllersToPods(podMap, podJobMap)

	// TODO breakdown network costs?

	// Build out a map of Nodes with resource costs, discounts, and node types
	// for converting resource allocation data to cumulative costs.
	nodeMap := map[nodeKey]*NodePricing{}

	applyNodeCostPerCPUHr(nodeMap, resNodeCostPerCPUHr)
	applyNodeCostPerRAMGiBHr(nodeMap, resNodeCostPerRAMGiBHr)
	applyNodeCostPerGPUHr(nodeMap, resNodeCostPerGPUHr)
	applyNodeSpot(nodeMap, resNodeIsSpot)
	applyNodeDiscount(nodeMap, cm)

	// Build out the map of all PVs with class, size and cost-per-hour.
	// Note: this does not record time running, which we may want to
	// include later for increased PV precision. (As long as the PV has
	// a PVC, we get time running there, so this is only inaccurate
	// for short-lived, unmounted PVs.)
	pvMap := map[pvKey]*PV{}
	buildPVMap(pvMap, resPVCostPerGiBHour)
	applyPVBytes(pvMap, resPVBytes)

	// Build out the map of all PVCs with time running, bytes requested,
	// and connect to the correct PV from pvMap. (If no PV exists, that
	// is noted, but does not result in any allocation/cost.)
	pvcMap := map[pvcKey]*PVC{}
	buildPVCMap(window, pvcMap, pvMap, resPVCInfo)
	applyPVCBytesRequested(pvcMap, resPVCBytesRequested)

	// Build out the relationships of pods to their PVCs. This step
	// populates the PVC.Count field so that PVC allocation can be
	// split appropriately among each pod's container allocation.
	podPVCMap := map[podKey][]*PVC{}
	buildPodPVCMap(podPVCMap, pvMap, pvcMap, podMap, resPodPVCAllocation)

	// Identify unmounted PVs (PVs without PVCs) and add one Allocation per
	// cluster representing each cluster's unmounted PVs (if necessary).
	applyUnmountedPVs(window, podMap, pvMap, pvcMap)

	lbMap := getLoadBalancerCosts(resLBCostPerHr, resLBActiveMins, resolution)
	applyLoadBalancersToPods(lbMap, allocsByService)

	// (3) Build out AllocationSet from Pod map

	for _, pod := range podMap {
		for _, alloc := range pod.Allocations {
			cluster := alloc.Properties.Cluster
			nodeName := alloc.Properties.Node
			namespace := alloc.Properties.Namespace
			pod := alloc.Properties.Pod
			container := alloc.Properties.Container

			podKey := newPodKey(cluster, namespace, pod)
			nodeKey := newNodeKey(cluster, nodeName)

			node := cm.getNodePricing(nodeMap, nodeKey)
			alloc.Properties.ProviderID = node.ProviderID
			alloc.CPUCost = alloc.CPUCoreHours * node.CostPerCPUHr
			alloc.RAMCost = (alloc.RAMByteHours / 1024 / 1024 / 1024) * node.CostPerRAMGiBHr
			alloc.GPUCost = alloc.GPUHours * node.CostPerGPUHr
			if pvcs, ok := podPVCMap[podKey]; ok {
				for _, pvc := range pvcs {
					// Determine the (start, end) of the relationship between the
					// given PVC and the associated Allocation so that a precise
					// number of hours can be used to compute cumulative cost.
					s, e := alloc.Start, alloc.End
					if pvc.Start.After(alloc.Start) {
						s = pvc.Start
					}
					if pvc.End.Before(alloc.End) {
						e = pvc.End
					}
					minutes := e.Sub(s).Minutes()
					hrs := minutes / 60.0

					count := float64(pvc.Count)
					if pvc.Count < 1 {
						count = 1
					}

					gib := pvc.Bytes / 1024 / 1024 / 1024
					cost := pvc.Volume.CostPerGiBHour * gib * hrs

					// Apply the size and cost of the PV to the allocation, each
					// weighted by count (i.e. the number of containers in the pod)
					// record the amount of total PVBytes Hours attributable to a given PV
					if alloc.PVs == nil {
						alloc.PVs = kubecost.PVAllocations{}
					}
					pvKey := kubecost.PVKey{
						Cluster: pvc.Cluster,
						Name:    pvc.Volume.Name,
					}
					alloc.PVs[pvKey] = &kubecost.PVAllocation{
						ByteHours: pvc.Bytes * hrs / count,
						Cost:      cost / count,
					}
				}
			}

			// Make sure that the name is correct (node may not be present at this
			// point due to it missing from queryMinutes) then insert.
			alloc.Name = fmt.Sprintf("%s/%s/%s/%s/%s", cluster, nodeName, namespace, pod, container)
			allocSet.Set(alloc)
		}
	}

	return allocSet, nil
}

func (cm *CostModel) buildPodMap(window kubecost.Window, resolution, maxBatchSize time.Duration, podMap map[podKey]*Pod, clusterStart, clusterEnd map[string]time.Time) error {
	// Assumes that window is positive and closed
	start, end := *window.Start(), *window.End()

	// Convert resolution duration to a query-ready string
	resStr := timeutil.DurationString(resolution)

	ctx := prom.NewContext(cm.PrometheusClient)

	// Query for (start, end) by (pod, namespace, cluster) over the given
	// window, using the given resolution, and if necessary in batches no
	// larger than the given maximum batch size. If working in batches, track
	// overall progress by starting with (window.start, window.start) and
	// querying in batches no larger than maxBatchSize from start-to-end,
	// folding each result set into podMap as the results come back.
	coverage := kubecost.NewWindow(&start, &start)

	numQuery := 1
	for coverage.End().Before(end) {
		// Determine the (start, end) of the current batch
		batchStart := *coverage.End()
		batchEnd := coverage.End().Add(maxBatchSize)
		if batchEnd.After(end) {
			batchEnd = end
		}
		batchWindow := kubecost.NewWindow(&batchStart, &batchEnd)

		var resPods []*prom.QueryResult
		var err error
		maxTries := 3
		numTries := 0
		for resPods == nil && numTries < maxTries {
			numTries++

			// Convert window (start, end) to (duration, offset) for querying Prometheus,
			// including handling Thanos offset
			durStr, offStr, err := batchWindow.DurationOffsetForPrometheus()
			if err != nil || durStr == "" {
				// Negative duration, so set empty results and don't query
				resPods = []*prom.QueryResult{}
				err = nil
				break
			}

			// Submit and profile query
			queryPods := fmt.Sprintf(queryFmtPods, env.GetPromClusterLabel(), durStr, resStr, offStr)
			queryProfile := time.Now()
			resPods, err = ctx.Query(queryPods).Await()
			if err != nil {
				log.Profile(queryProfile, fmt.Sprintf("CostModel.ComputeAllocation: pod query %d try %d failed: %s", numQuery, numTries, queryPods))
				resPods = nil
			}
		}

		if err != nil {
			return err
		}

		applyPodResults(window, resolution, podMap, clusterStart, clusterEnd, resPods)

		coverage = coverage.ExpandEnd(batchEnd)
		numQuery++
	}

	return nil
}

func applyPodResults(window kubecost.Window, resolution time.Duration, podMap map[podKey]*Pod, clusterStart, clusterEnd map[string]time.Time, resPods []*prom.QueryResult) {
	for _, res := range resPods {
		if len(res.Values) == 0 {
			log.Warningf("CostModel.ComputeAllocation: empty minutes result")
			continue
		}

		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		labels, err := res.GetStrings("namespace", "pod")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: minutes query result missing field: %s", err)
			continue
		}

		namespace := labels["namespace"]
		pod := labels["pod"]
		key := newPodKey(cluster, namespace, pod)

		// allocStart and allocEnd are the timestamps of the first and last
		// minutes the pod was running, respectively. We subtract one resolution
		// from allocStart because this point will actually represent the end
		// of the first minute. We don't subtract from allocEnd because it
		// already represents the end of the last minute.
		var allocStart, allocEnd time.Time
		startAdjustmentCoeff, endAdjustmentCoeff := 1.0, 1.0
		for _, datum := range res.Values {
			t := time.Unix(int64(datum.Timestamp), 0)

			if allocStart.IsZero() && datum.Value > 0 && window.Contains(t) {
				// Set the start timestamp to the earliest non-zero timestamp
				allocStart = t

				// Record adjustment coefficient, i.e. the portion of the start
				// timestamp to "ignore". That is, sometimes the value will be
				// 0.5, meaning that we should discount the time running by
				// half of the resolution the timestamp stands for.
				startAdjustmentCoeff = (1.0 - datum.Value)
			}

			if datum.Value > 0 && window.Contains(t) {
				// Set the end timestamp to the latest non-zero timestamp
				allocEnd = t

				// Record adjustment coefficient, i.e. the portion of the end
				// timestamp to "ignore". (See explanation above for start.)
				endAdjustmentCoeff = (1.0 - datum.Value)
			}
		}

		if allocStart.IsZero() || allocEnd.IsZero() {
			continue
		}

		// Adjust timestamps according to the resolution and the adjustment
		// coefficients, as described above. That is, count the start timestamp
		// from the beginning of the resolution, not the end. Then "reduce" the
		// start and end by the correct amount, in the case that the "running"
		// value of the first or last timestamp was not a full 1.0.
		allocStart = allocStart.Add(-resolution)
		// Note: the *100 and /100 are necessary because Duration is an int, so
		// 0.5, for instance, will be truncated, resulting in no adjustment.
		allocStart = allocStart.Add(time.Duration(startAdjustmentCoeff*100) * resolution / time.Duration(100))
		allocEnd = allocEnd.Add(-time.Duration(endAdjustmentCoeff*100) * resolution / time.Duration(100))

		// If there is only one point with a value <= 0.5 that the start and
		// end timestamps both share, then we will enter this case because at
		// least half of a resolution will be subtracted from both the start
		// and the end. If that is the case, then add back half of each side
		// so that the pod is said to run for half a resolution total.
		// e.g. For resolution 1m and a value of 0.5 at one timestamp, we'll
		//      end up with allocEnd == allocStart and each coeff == 0.5. In
		//      that case, add 0.25m to each side, resulting in 0.5m duration.
		if !allocEnd.After(allocStart) {
			allocStart = allocStart.Add(-time.Duration(50*startAdjustmentCoeff) * resolution / time.Duration(100))
			allocEnd = allocEnd.Add(time.Duration(50*endAdjustmentCoeff) * resolution / time.Duration(100))
		}

		// Set start if unset or this datum's start time is earlier than the
		// current earliest time.
		if _, ok := clusterStart[cluster]; !ok || allocStart.Before(clusterStart[cluster]) {
			clusterStart[cluster] = allocStart
		}

		// Set end if unset or this datum's end time is later than the
		// current latest time.
		if _, ok := clusterEnd[cluster]; !ok || allocEnd.After(clusterEnd[cluster]) {
			clusterEnd[cluster] = allocEnd
		}

		if pod, ok := podMap[key]; ok {
			// Pod has already been recorded, so update it accordingly
			if allocStart.Before(pod.Start) {
				pod.Start = allocStart
			}
			if allocEnd.After(pod.End) {
				pod.End = allocEnd
			}
		} else {
			// Pod has not been recorded yet, so insert it
			podMap[key] = &Pod{
				Window:      window.Clone(),
				Start:       allocStart,
				End:         allocEnd,
				Key:         key,
				Allocations: map[string]*kubecost.Allocation{},
			}
		}
	}
}

func applyCPUCoresAllocated(podMap map[podKey]*Pod, resCPUCoresAllocated []*prom.QueryResult) {
	for _, res := range resCPUCoresAllocated {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU allocation result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU allocation query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		cpuCores := res.Values[0].Value
		if cpuCores > MAX_CPU_CAP {
			klog.Infof("[WARNING] Very large cpu allocation, clamping to %f", res.Values[0].Value*(pod.Allocations[container].Minutes()/60.0))
			cpuCores = 0.0
		}
		hours := pod.Allocations[container].Minutes() / 60.0
		pod.Allocations[container].CPUCoreHours = cpuCores * hours

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: CPU allocation query result missing 'node': %s", key)
			continue
		}
		pod.Allocations[container].Properties.Node = node
	}
}

func applyCPUCoresRequested(podMap map[podKey]*Pod, resCPUCoresRequested []*prom.QueryResult) {
	for _, res := range resCPUCoresRequested {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU request result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU request query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		pod.Allocations[container].CPUCoreRequestAverage = res.Values[0].Value

		// If CPU allocation is less than requests, set CPUCoreHours to
		// request level.
		if pod.Allocations[container].CPUCores() < res.Values[0].Value {
			pod.Allocations[container].CPUCoreHours = res.Values[0].Value * (pod.Allocations[container].Minutes() / 60.0)
		}
		if pod.Allocations[container].CPUCores() > MAX_CPU_CAP {
			klog.Infof("[WARNING] Very large cpu allocation, clamping! to %f", res.Values[0].Value*(pod.Allocations[container].Minutes()/60.0))
			pod.Allocations[container].CPUCoreHours = res.Values[0].Value * (pod.Allocations[container].Minutes() / 60.0)
		}

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: CPU request query result missing 'node': %s", key)
			continue
		}
		pod.Allocations[container].Properties.Node = node
	}
}

func applyCPUCoresUsedAvg(podMap map[podKey]*Pod, resCPUCoresUsedAvg []*prom.QueryResult) {
	for _, res := range resCPUCoresUsedAvg {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU usage avg result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}
		container, err := res.GetString("container")
		if container == "" || err != nil {
			container, err = res.GetString("container_name")
			if err != nil {
				log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU usage avg query result missing 'container': %s", key)
				continue
			}
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		pod.Allocations[container].CPUCoreUsageAverage = res.Values[0].Value
		if res.Values[0].Value > MAX_CPU_CAP {
			klog.Infof("[WARNING] Very large cpu USAGE, dropping outlier")
			pod.Allocations[container].CPUCoreUsageAverage = 0.0
		}
	}
}

func applyCPUCoresUsedMax(podMap map[podKey]*Pod, resCPUCoresUsedMax []*prom.QueryResult) {
	for _, res := range resCPUCoresUsedMax {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU usage max result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container_name")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU usage max query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		if pod.Allocations[container].RawAllocationOnly == nil {
			pod.Allocations[container].RawAllocationOnly = &kubecost.RawAllocationOnlyData{
				CPUCoreUsageMax: res.Values[0].Value,
			}
		} else {
			pod.Allocations[container].RawAllocationOnly.CPUCoreUsageMax = res.Values[0].Value
		}
	}
}

func applyRAMBytesAllocated(podMap map[podKey]*Pod, resRAMBytesAllocated []*prom.QueryResult) {
	for _, res := range resRAMBytesAllocated {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM allocation result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM allocation query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		ramBytes := res.Values[0].Value
		hours := pod.Allocations[container].Minutes() / 60.0
		pod.Allocations[container].RAMByteHours = ramBytes * hours

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: RAM allocation query result missing 'node': %s", key)
			continue
		}
		pod.Allocations[container].Properties.Node = node
	}
}

func applyRAMBytesRequested(podMap map[podKey]*Pod, resRAMBytesRequested []*prom.QueryResult) {
	for _, res := range resRAMBytesRequested {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM request result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM request query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		pod.Allocations[container].RAMBytesRequestAverage = res.Values[0].Value

		// If RAM allocation is less than requests, set RAMByteHours to
		// request level.
		if pod.Allocations[container].RAMBytes() < res.Values[0].Value {
			pod.Allocations[container].RAMByteHours = res.Values[0].Value * (pod.Allocations[container].Minutes() / 60.0)
		}

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: RAM request query result missing 'node': %s", key)
			continue
		}
		pod.Allocations[container].Properties.Node = node
	}
}

func applyRAMBytesUsedAvg(podMap map[podKey]*Pod, resRAMBytesUsedAvg []*prom.QueryResult) {
	for _, res := range resRAMBytesUsedAvg {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM avg usage result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container")
		if container == "" || err != nil {
			container, err = res.GetString("container_name")
			if err != nil {
				log.DedupedWarningf(10, "CostModel.ComputeAllocation: CPU usage avg query result missing 'container': %s", key)
				continue
			}
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		pod.Allocations[container].RAMBytesUsageAverage = res.Values[0].Value
	}
}

func applyRAMBytesUsedMax(podMap map[podKey]*Pod, resRAMBytesUsedMax []*prom.QueryResult) {
	for _, res := range resRAMBytesUsedMax {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM usage max result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container_name")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: RAM usage max query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		if pod.Allocations[container].RawAllocationOnly == nil {
			pod.Allocations[container].RawAllocationOnly = &kubecost.RawAllocationOnlyData{
				RAMBytesUsageMax: res.Values[0].Value,
			}
		} else {
			pod.Allocations[container].RawAllocationOnly.RAMBytesUsageMax = res.Values[0].Value
		}
	}
}

func applyGPUsAllocated(podMap map[podKey]*Pod, resGPUsRequested []*prom.QueryResult, resGPUsAllocated []*prom.QueryResult) {
	if len(resGPUsAllocated) > 0 { // Use the new query, when it's become available in a window
		resGPUsRequested = resGPUsAllocated
	}
	for _, res := range resGPUsRequested {
		key, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: GPU request result missing field: %s", err)
			continue
		}

		pod, ok := podMap[key]
		if !ok {
			continue
		}

		container, err := res.GetString("container")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: GPU request query result missing 'container': %s", key)
			continue
		}

		if _, ok := pod.Allocations[container]; !ok {
			pod.AppendContainer(container)
		}

		hrs := pod.Allocations[container].Minutes() / 60.0
		pod.Allocations[container].GPUHours = res.Values[0].Value * hrs
	}
}

func applyNetworkTotals(podMap map[podKey]*Pod, resNetworkTransferBytes []*prom.QueryResult, resNetworkReceiveBytes []*prom.QueryResult) {
	for _, res := range resNetworkTransferBytes {
		podKey, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: Network Transfer Bytes query result missing field: %s", err)
			continue
		}

		pod, ok := podMap[podKey]
		if !ok {
			continue
		}

		for _, alloc := range pod.Allocations {
			alloc.NetworkTransferBytes = res.Values[0].Value / float64(len(pod.Allocations))
		}
	}
	for _, res := range resNetworkReceiveBytes {
		podKey, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: Network Receive Bytes query result missing field: %s", err)
			continue
		}

		pod, ok := podMap[podKey]
		if !ok {
			continue
		}

		for _, alloc := range pod.Allocations {
			alloc.NetworkReceiveBytes = res.Values[0].Value / float64(len(pod.Allocations))
		}
	}
}

func applyNetworkAllocation(podMap map[podKey]*Pod, resNetworkGiB []*prom.QueryResult, resNetworkCostPerGiB []*prom.QueryResult) {
	costPerGiBByCluster := map[string]float64{}

	for _, res := range resNetworkCostPerGiB {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		costPerGiBByCluster[cluster] = res.Values[0].Value
	}

	for _, res := range resNetworkGiB {
		podKey, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: Network allocation query result missing field: %s", err)
			continue
		}

		pod, ok := podMap[podKey]
		if !ok {
			continue
		}

		for _, alloc := range pod.Allocations {
			gib := res.Values[0].Value / float64(len(pod.Allocations))
			costPerGiB := costPerGiBByCluster[podKey.Cluster]
			alloc.NetworkCost = gib * costPerGiB
		}
	}
}

func resToNamespaceLabels(resNamespaceLabels []*prom.QueryResult) map[namespaceKey]map[string]string {
	namespaceLabels := map[namespaceKey]map[string]string{}

	for _, res := range resNamespaceLabels {
		nsKey, err := resultNamespaceKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			continue
		}

		if _, ok := namespaceLabels[nsKey]; !ok {
			namespaceLabels[nsKey] = map[string]string{}
		}

		for k, l := range res.GetLabels() {
			namespaceLabels[nsKey][k] = l
		}
	}

	return namespaceLabels
}

func resToPodLabels(resPodLabels []*prom.QueryResult) map[podKey]map[string]string {
	podLabels := map[podKey]map[string]string{}

	for _, res := range resPodLabels {
		podKey, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			continue
		}

		if _, ok := podLabels[podKey]; !ok {
			podLabels[podKey] = map[string]string{}
		}

		for k, l := range res.GetLabels() {
			podLabels[podKey][k] = l
		}
	}

	return podLabels
}

func resToNamespaceAnnotations(resNamespaceAnnotations []*prom.QueryResult) map[string]map[string]string {
	namespaceAnnotations := map[string]map[string]string{}

	for _, res := range resNamespaceAnnotations {
		namespace, err := res.GetString("namespace")
		if err != nil {
			continue
		}

		if _, ok := namespaceAnnotations[namespace]; !ok {
			namespaceAnnotations[namespace] = map[string]string{}
		}

		for k, l := range res.GetAnnotations() {
			namespaceAnnotations[namespace][k] = l
		}
	}

	return namespaceAnnotations
}

func resToPodAnnotations(resPodAnnotations []*prom.QueryResult) map[podKey]map[string]string {
	podAnnotations := map[podKey]map[string]string{}

	for _, res := range resPodAnnotations {
		podKey, err := resultPodKey(res, env.GetPromClusterLabel(), "namespace")
		if err != nil {
			continue
		}

		if _, ok := podAnnotations[podKey]; !ok {
			podAnnotations[podKey] = map[string]string{}
		}

		for k, l := range res.GetAnnotations() {
			podAnnotations[podKey][k] = l
		}
	}

	return podAnnotations
}

func applyLabels(podMap map[podKey]*Pod, namespaceLabels map[namespaceKey]map[string]string, podLabels map[podKey]map[string]string) {
	for podKey, pod := range podMap {
		for _, alloc := range pod.Allocations {
			allocLabels := alloc.Properties.Labels
			if allocLabels == nil {
				allocLabels = make(map[string]string)
			}
			// Apply namespace labels first, then pod labels so that pod labels
			// overwrite namespace labels.
			nsKey := newNamespaceKey(podKey.Cluster, podKey.Namespace)
			if labels, ok := namespaceLabels[nsKey]; ok {
				for k, v := range labels {
					allocLabels[k] = v
				}
			}
			if labels, ok := podLabels[podKey]; ok {
				for k, v := range labels {
					allocLabels[k] = v
				}
			}

			alloc.Properties.Labels = allocLabels
		}
	}
}

func applyAnnotations(podMap map[podKey]*Pod, namespaceAnnotations map[string]map[string]string, podAnnotations map[podKey]map[string]string) {
	for key, pod := range podMap {
		for _, alloc := range pod.Allocations {
			allocAnnotations := alloc.Properties.Annotations
			if allocAnnotations == nil {
				allocAnnotations = make(map[string]string)
			}
			// Apply namespace annotations first, then pod annotations so that
			// pod labels overwrite namespace labels.
			if labels, ok := namespaceAnnotations[key.Namespace]; ok {
				for k, v := range labels {
					allocAnnotations[k] = v
				}
			}
			if labels, ok := podAnnotations[key]; ok {
				for k, v := range labels {
					allocAnnotations[k] = v
				}
			}

			alloc.Properties.Annotations = allocAnnotations
		}
	}
}

func getServiceLabels(resServiceLabels []*prom.QueryResult) map[serviceKey]map[string]string {
	serviceLabels := map[serviceKey]map[string]string{}

	for _, res := range resServiceLabels {
		serviceKey, err := resultServiceKey(res, env.GetPromClusterLabel(), "namespace", "service")
		if err != nil {
			continue
		}

		if _, ok := serviceLabels[serviceKey]; !ok {
			serviceLabels[serviceKey] = map[string]string{}
		}

		for k, l := range res.GetLabels() {
			serviceLabels[serviceKey][k] = l
		}
	}

	// Prune duplicate services. That is, if the same service exists with
	// hyphens instead of underscores, keep the one that uses hyphens.
	for key := range serviceLabels {
		if strings.Contains(key.Service, "_") {
			duplicateService := strings.Replace(key.Service, "_", "-", -1)
			duplicateKey := newServiceKey(key.Cluster, key.Namespace, duplicateService)
			if _, ok := serviceLabels[duplicateKey]; ok {
				delete(serviceLabels, key)
			}
		}
	}

	return serviceLabels
}

func resToDeploymentLabels(resDeploymentLabels []*prom.QueryResult) map[controllerKey]map[string]string {
	deploymentLabels := map[controllerKey]map[string]string{}

	for _, res := range resDeploymentLabels {
		controllerKey, err := resultDeploymentKey(res, env.GetPromClusterLabel(), "namespace", "deployment")
		if err != nil {
			continue
		}

		if _, ok := deploymentLabels[controllerKey]; !ok {
			deploymentLabels[controllerKey] = map[string]string{}
		}

		for k, l := range res.GetLabels() {
			deploymentLabels[controllerKey][k] = l
		}
	}

	// Prune duplicate deployments. That is, if the same deployment exists with
	// hyphens instead of underscores, keep the one that uses hyphens.
	for key := range deploymentLabels {
		if strings.Contains(key.Controller, "_") {
			duplicateController := strings.Replace(key.Controller, "_", "-", -1)
			duplicateKey := newControllerKey(key.Cluster, key.Namespace, key.ControllerKind, duplicateController)
			if _, ok := deploymentLabels[duplicateKey]; ok {
				delete(deploymentLabels, key)
			}
		}
	}

	return deploymentLabels
}

func resToStatefulSetLabels(resStatefulSetLabels []*prom.QueryResult) map[controllerKey]map[string]string {
	statefulSetLabels := map[controllerKey]map[string]string{}

	for _, res := range resStatefulSetLabels {
		controllerKey, err := resultStatefulSetKey(res, env.GetPromClusterLabel(), "namespace", "statefulSet")
		if err != nil {
			continue
		}

		if _, ok := statefulSetLabels[controllerKey]; !ok {
			statefulSetLabels[controllerKey] = map[string]string{}
		}

		for k, l := range res.GetLabels() {
			statefulSetLabels[controllerKey][k] = l
		}
	}

	// Prune duplicate stateful sets. That is, if the same stateful set exists
	// with hyphens instead of underscores, keep the one that uses hyphens.
	for key := range statefulSetLabels {
		if strings.Contains(key.Controller, "_") {
			duplicateController := strings.Replace(key.Controller, "_", "-", -1)
			duplicateKey := newControllerKey(key.Cluster, key.Namespace, key.ControllerKind, duplicateController)
			if _, ok := statefulSetLabels[duplicateKey]; ok {
				delete(statefulSetLabels, key)
			}
		}
	}

	return statefulSetLabels
}

func labelsToPodControllerMap(podLabels map[podKey]map[string]string, controllerLabels map[controllerKey]map[string]string) map[podKey]controllerKey {
	podControllerMap := map[podKey]controllerKey{}

	// For each controller, turn the labels into a selector and attempt to
	// match it with each set of pod labels. A match indicates that the pod
	// belongs to the controller.
	for cKey, cLabels := range controllerLabels {
		selector := labels.Set(cLabels).AsSelectorPreValidated()

		for pKey, pLabels := range podLabels {
			// If the pod is in a different cluster or namespace, there is
			// no need to compare the labels.
			if cKey.Cluster != pKey.Cluster || cKey.Namespace != pKey.Namespace {
				continue
			}

			podLabelSet := labels.Set(pLabels)
			if selector.Matches(podLabelSet) {
				if _, ok := podControllerMap[pKey]; ok {
					log.DedupedWarningf(5, "CostModel.ComputeAllocation: PodControllerMap match already exists: %s matches %s and %s", pKey, podControllerMap[pKey], cKey)
				}
				podControllerMap[pKey] = cKey
			}
		}
	}

	return podControllerMap
}

func resToPodDaemonSetMap(resDaemonSetLabels []*prom.QueryResult) map[podKey]controllerKey {
	daemonSetLabels := map[podKey]controllerKey{}

	for _, res := range resDaemonSetLabels {
		controllerKey, err := resultDaemonSetKey(res, env.GetPromClusterLabel(), "namespace", "owner_name")
		if err != nil {
			continue
		}

		pod, err := res.GetString("pod")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: DaemonSetLabel result without pod: %s", controllerKey)
		}

		podKey := newPodKey(controllerKey.Cluster, controllerKey.Namespace, pod)

		daemonSetLabels[podKey] = controllerKey
	}

	return daemonSetLabels
}

func resToPodJobMap(resJobLabels []*prom.QueryResult) map[podKey]controllerKey {
	jobLabels := map[podKey]controllerKey{}

	for _, res := range resJobLabels {
		controllerKey, err := resultJobKey(res, env.GetPromClusterLabel(), "namespace", "owner_name")
		if err != nil {
			continue
		}

		// Convert the name of Jobs generated by CronJobs to the name of the
		// CronJob by stripping the timestamp off the end.
		match := isCron.FindStringSubmatch(controllerKey.Controller)
		if match != nil {
			controllerKey.Controller = match[1]
		}

		pod, err := res.GetString("pod")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: JobLabel result without pod: %s", controllerKey)
		}

		podKey := newPodKey(controllerKey.Cluster, controllerKey.Namespace, pod)

		jobLabels[podKey] = controllerKey
	}

	return jobLabels
}

func applyServicesToPods(podMap map[podKey]*Pod, podLabels map[podKey]map[string]string, allocsByService map[serviceKey][]*kubecost.Allocation, serviceLabels map[serviceKey]map[string]string) {
	podServicesMap := map[podKey][]serviceKey{}

	// For each service, turn the labels into a selector and attempt to
	// match it with each set of pod labels. A match indicates that the pod
	// belongs to the service.
	for sKey, sLabels := range serviceLabels {
		selector := labels.Set(sLabels).AsSelectorPreValidated()

		for pKey, pLabels := range podLabels {
			// If the pod is in a different cluster or namespace, there is
			// no need to compare the labels.
			if sKey.Cluster != pKey.Cluster || sKey.Namespace != pKey.Namespace {
				continue
			}

			podLabelSet := labels.Set(pLabels)
			if selector.Matches(podLabelSet) {
				if _, ok := podServicesMap[pKey]; !ok {
					podServicesMap[pKey] = []serviceKey{}
				}
				podServicesMap[pKey] = append(podServicesMap[pKey], sKey)
			}
		}
	}

	// For each allocation in each pod, attempt to find and apply the list of
	// services associated with the allocation's pod.
	for key, pod := range podMap {
		for _, alloc := range pod.Allocations {
			if sKeys, ok := podServicesMap[key]; ok {
				services := []string{}
				for _, sKey := range sKeys {
					services = append(services, sKey.Service)
					allocsByService[sKey] = append(allocsByService[sKey], alloc)
				}
				alloc.Properties.Services = services

			}
		}
	}
}

func applyControllersToPods(podMap map[podKey]*Pod, podControllerMap map[podKey]controllerKey) {
	for key, pod := range podMap {
		for _, alloc := range pod.Allocations {
			if controllerKey, ok := podControllerMap[key]; ok {
				alloc.Properties.ControllerKind = controllerKey.ControllerKind
				alloc.Properties.Controller = controllerKey.Controller
			}
		}
	}
}

func applyNodeCostPerCPUHr(nodeMap map[nodeKey]*NodePricing, resNodeCostPerCPUHr []*prom.QueryResult) {
	for _, res := range resNodeCostPerCPUHr {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node CPU cost query result missing field: %s", err)
			continue
		}

		instanceType, err := res.GetString("instance_type")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node CPU cost query result missing field: %s", err)
			continue
		}

		providerID, err := res.GetString("provider_id")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node CPU cost query result missing field: %s", err)
			continue
		}

		key := newNodeKey(cluster, node)
		if _, ok := nodeMap[key]; !ok {
			nodeMap[key] = &NodePricing{
				Name:       node,
				NodeType:   instanceType,
				ProviderID: cloud.ParseID(providerID),
			}
		}

		nodeMap[key].CostPerCPUHr = res.Values[0].Value
	}
}

func applyNodeCostPerRAMGiBHr(nodeMap map[nodeKey]*NodePricing, resNodeCostPerRAMGiBHr []*prom.QueryResult) {
	for _, res := range resNodeCostPerRAMGiBHr {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node RAM cost query result missing field: %s", err)
			continue
		}

		instanceType, err := res.GetString("instance_type")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node RAM cost query result missing field: %s", err)
			continue
		}

		providerID, err := res.GetString("provider_id")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node RAM cost query result missing field: %s", err)
			continue
		}

		key := newNodeKey(cluster, node)
		if _, ok := nodeMap[key]; !ok {
			nodeMap[key] = &NodePricing{
				Name:       node,
				NodeType:   instanceType,
				ProviderID: cloud.ParseID(providerID),
			}
		}

		nodeMap[key].CostPerRAMGiBHr = res.Values[0].Value
	}
}

func applyNodeCostPerGPUHr(nodeMap map[nodeKey]*NodePricing, resNodeCostPerGPUHr []*prom.QueryResult) {
	for _, res := range resNodeCostPerGPUHr {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node GPU cost query result missing field: %s", err)
			continue
		}

		instanceType, err := res.GetString("instance_type")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node GPU cost query result missing field: %s", err)
			continue
		}

		providerID, err := res.GetString("provider_id")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node GPU cost query result missing field: %s", err)
			continue
		}

		key := newNodeKey(cluster, node)
		if _, ok := nodeMap[key]; !ok {
			nodeMap[key] = &NodePricing{
				Name:       node,
				NodeType:   instanceType,
				ProviderID: cloud.ParseID(providerID),
			}
		}

		nodeMap[key].CostPerGPUHr = res.Values[0].Value
	}
}

func applyNodeSpot(nodeMap map[nodeKey]*NodePricing, resNodeIsSpot []*prom.QueryResult) {
	for _, res := range resNodeIsSpot {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		node, err := res.GetString("node")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: Node spot query result missing field: %s", err)
			continue
		}

		key := newNodeKey(cluster, node)
		if _, ok := nodeMap[key]; !ok {
			log.Warningf("CostModel.ComputeAllocation: Node spot  query result for missing node: %s", key)
			continue
		}

		nodeMap[key].Preemptible = res.Values[0].Value > 0
	}
}

func applyNodeDiscount(nodeMap map[nodeKey]*NodePricing, cm *CostModel) {
	if cm == nil {
		return
	}

	c, err := cm.Provider.GetConfig()
	if err != nil {
		log.Errorf("CostModel.ComputeAllocation: applyNodeDiscount: %s", err)
		return
	}

	discount, err := ParsePercentString(c.Discount)
	if err != nil {
		log.Errorf("CostModel.ComputeAllocation: applyNodeDiscount: %s", err)
		return
	}

	negotiatedDiscount, err := ParsePercentString(c.NegotiatedDiscount)
	if err != nil {
		log.Errorf("CostModel.ComputeAllocation: applyNodeDiscount: %s", err)
		return
	}

	for _, node := range nodeMap {
		// TODO GKE Reserved Instances into account
		node.Discount = cm.Provider.CombinedDiscountForNode(node.NodeType, node.Preemptible, discount, negotiatedDiscount)
		node.CostPerCPUHr *= (1.0 - node.Discount)
		node.CostPerRAMGiBHr *= (1.0 - node.Discount)
	}
}

func buildPVMap(pvMap map[pvKey]*PV, resPVCostPerGiBHour []*prom.QueryResult) {
	for _, res := range resPVCostPerGiBHour {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		name, err := res.GetString("volumename")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: PV cost without volumename")
			continue
		}

		key := newPVKey(cluster, name)

		pvMap[key] = &PV{
			Cluster:        cluster,
			Name:           name,
			CostPerGiBHour: res.Values[0].Value,
		}
	}
}

func applyPVBytes(pvMap map[pvKey]*PV, resPVBytes []*prom.QueryResult) {
	for _, res := range resPVBytes {
		key, err := resultPVKey(res, env.GetPromClusterLabel(), "persistentvolume")
		if err != nil {
			log.Warningf("CostModel.ComputeAllocation: PV bytes query result missing field: %s", err)
			continue
		}

		if _, ok := pvMap[key]; !ok {
			log.Warningf("CostModel.ComputeAllocation: PV bytes result for missing PV: %s", err)
			continue
		}

		pvMap[key].Bytes = res.Values[0].Value
	}
}

func buildPVCMap(window kubecost.Window, pvcMap map[pvcKey]*PVC, pvMap map[pvKey]*PV, resPVCInfo []*prom.QueryResult) {
	for _, res := range resPVCInfo {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		values, err := res.GetStrings("persistentvolumeclaim", "storageclass", "volumename", "namespace")
		if err != nil {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: PVC info query result missing field: %s", err)
			continue
		}

		namespace := values["namespace"]
		name := values["persistentvolumeclaim"]
		volume := values["volumename"]
		storageClass := values["storageclass"]

		pvKey := newPVKey(cluster, volume)
		pvcKey := newPVCKey(cluster, namespace, name)

		// pvcStart and pvcEnd are the timestamps of the first and last minutes
		// the PVC was running, respectively. We subtract 1m from pvcStart
		// because this point will actually represent the end of the first
		// minute. We don't subtract from pvcEnd because it already represents
		// the end of the last minute.
		var pvcStart, pvcEnd time.Time
		for _, datum := range res.Values {
			t := time.Unix(int64(datum.Timestamp), 0)
			if pvcStart.IsZero() && datum.Value > 0 && window.Contains(t) {
				pvcStart = t
			}
			if datum.Value > 0 && window.Contains(t) {
				pvcEnd = t
			}
		}
		if pvcStart.IsZero() || pvcEnd.IsZero() {
			log.Warningf("CostModel.ComputeAllocation: PVC %s has no running time", pvcKey)
		}
		pvcStart = pvcStart.Add(-time.Minute)

		if _, ok := pvMap[pvKey]; !ok {
			continue
		}

		pvMap[pvKey].StorageClass = storageClass

		if _, ok := pvcMap[pvcKey]; !ok {
			pvcMap[pvcKey] = &PVC{}
		}

		pvcMap[pvcKey].Name = name
		pvcMap[pvcKey].Namespace = namespace
		pvcMap[pvcKey].Volume = pvMap[pvKey]
		pvcMap[pvcKey].Start = pvcStart
		pvcMap[pvcKey].End = pvcEnd
	}
}

func applyPVCBytesRequested(pvcMap map[pvcKey]*PVC, resPVCBytesRequested []*prom.QueryResult) {
	for _, res := range resPVCBytesRequested {
		key, err := resultPVCKey(res, env.GetPromClusterLabel(), "namespace", "persistentvolumeclaim")
		if err != nil {
			continue
		}

		if _, ok := pvcMap[key]; !ok {
			continue
		}

		pvcMap[key].Bytes = res.Values[0].Value
	}
}

func buildPodPVCMap(podPVCMap map[podKey][]*PVC, pvMap map[pvKey]*PV, pvcMap map[pvcKey]*PVC, podMap map[podKey]*Pod, resPodPVCAllocation []*prom.QueryResult) {
	for _, res := range resPodPVCAllocation {
		cluster, err := res.GetString(env.GetPromClusterLabel())
		if err != nil {
			cluster = env.GetClusterID()
		}

		values, err := res.GetStrings("persistentvolume", "persistentvolumeclaim", "pod", "namespace")
		if err != nil {
			log.DedupedWarningf(5, "CostModel.ComputeAllocation: PVC allocation query result missing field: %s", err)
			continue
		}

		namespace := values["namespace"]
		pod := values["pod"]
		name := values["persistentvolumeclaim"]
		volume := values["persistentvolume"]

		podKey := newPodKey(cluster, namespace, pod)
		pvKey := newPVKey(cluster, volume)
		pvcKey := newPVCKey(cluster, namespace, name)

		if _, ok := pvMap[pvKey]; !ok {
			log.DedupedWarningf(5, "CostModel.ComputeAllocation: PV missing for PVC allocation query result: %s", pvKey)
			continue
		}

		if _, ok := podPVCMap[podKey]; !ok {
			podPVCMap[podKey] = []*PVC{}
		}

		pvc, ok := pvcMap[pvcKey]
		if !ok {
			log.DedupedWarningf(5, "CostModel.ComputeAllocation: PVC missing for PVC allocation query: %s", pvcKey)
			continue
		}

		count := 1
		if pod, ok := podMap[podKey]; ok && len(pod.Allocations) > 0 {
			count = len(pod.Allocations)
		} else {
			log.DedupedWarningf(10, "CostModel.ComputeAllocation: PVC %s for missing pod %s", pvcKey, podKey)
		}

		pvc.Count = count
		pvc.Mounted = true

		podPVCMap[podKey] = append(podPVCMap[podKey], pvc)
	}
}

func applyUnmountedPVs(window kubecost.Window, podMap map[podKey]*Pod, pvMap map[pvKey]*PV, pvcMap map[pvcKey]*PVC) {
	unmountedPVBytes := map[string]float64{}
	unmountedPVCost := map[string]float64{}

	for _, pv := range pvMap {
		mounted := false
		for _, pvc := range pvcMap {
			if pvc.Volume == nil {
				continue
			}
			if pvc.Volume == pv {
				mounted = true
				break
			}
		}

		if !mounted {
			gib := pv.Bytes / 1024 / 1024 / 1024
			hrs := window.Minutes() / 60.0 // TODO improve with PV hours, not window hours
			cost := pv.CostPerGiBHour * gib * hrs
			unmountedPVCost[pv.Cluster] += cost
			unmountedPVBytes[pv.Cluster] += pv.Bytes
		}
	}

	for cluster, amount := range unmountedPVCost {
		container := kubecost.UnmountedSuffix
		pod := kubecost.UnmountedSuffix
		namespace := kubecost.UnmountedSuffix
		node := ""

		key := newPodKey(cluster, namespace, pod)
		podMap[key] = &Pod{
			Window:      window.Clone(),
			Start:       *window.Start(),
			End:         *window.End(),
			Key:         key,
			Allocations: map[string]*kubecost.Allocation{},
		}

		podMap[key].AppendContainer(container)
		podMap[key].Allocations[container].Properties.Cluster = cluster
		podMap[key].Allocations[container].Properties.Node = node
		podMap[key].Allocations[container].Properties.Namespace = namespace
		podMap[key].Allocations[container].Properties.Pod = pod
		podMap[key].Allocations[container].Properties.Container = container
		pvKey := kubecost.PVKey{
			Cluster: cluster,
			Name:    kubecost.UnmountedSuffix,
		}
		unmountedPVs := kubecost.PVAllocations{
			pvKey: {
				ByteHours: unmountedPVBytes[cluster] * window.Minutes() / 60.0,
				Cost:      amount,
			},
		}
		podMap[key].Allocations[container].PVs = podMap[key].Allocations[container].PVs.Add(unmountedPVs)
	}
}

func applyUnmountedPVCs(window kubecost.Window, podMap map[podKey]*Pod, pvcMap map[pvcKey]*PVC) {
	unmountedPVCBytes := map[namespaceKey]float64{}
	unmountedPVCCost := map[namespaceKey]float64{}

	for _, pvc := range pvcMap {
		if !pvc.Mounted && pvc.Volume != nil {
			key := newNamespaceKey(pvc.Cluster, pvc.Namespace)

			gib := pvc.Volume.Bytes / 1024 / 1024 / 1024
			hrs := pvc.Minutes() / 60.0
			cost := pvc.Volume.CostPerGiBHour * gib * hrs
			unmountedPVCCost[key] += cost
			unmountedPVCBytes[key] += pvc.Volume.Bytes
		}
	}

	for key, amount := range unmountedPVCCost {
		container := kubecost.UnmountedSuffix
		pod := kubecost.UnmountedSuffix
		namespace := key.Namespace
		node := ""
		cluster := key.Cluster

		podKey := newPodKey(cluster, namespace, pod)
		podMap[podKey] = &Pod{
			Window:      window.Clone(),
			Start:       *window.Start(),
			End:         *window.End(),
			Key:         podKey,
			Allocations: map[string]*kubecost.Allocation{},
		}

		podMap[podKey].AppendContainer(container)
		podMap[podKey].Allocations[container].Properties.Cluster = cluster
		podMap[podKey].Allocations[container].Properties.Node = node
		podMap[podKey].Allocations[container].Properties.Namespace = namespace
		podMap[podKey].Allocations[container].Properties.Pod = pod
		podMap[podKey].Allocations[container].Properties.Container = container
		pvKey := kubecost.PVKey{
			Cluster: cluster,
			Name:    kubecost.UnmountedSuffix,
		}
		unmountedPVs := kubecost.PVAllocations{
			pvKey: {
				ByteHours: unmountedPVCBytes[key] * window.Minutes() / 60.0,
				Cost:      amount,
			},
		}
		podMap[podKey].Allocations[container].PVs = podMap[podKey].Allocations[container].PVs.Add(unmountedPVs)

	}
}

// LB describes the start and end time of a Load Balancer along with cost
type LB struct {
	TotalCost float64
	Start     time.Time
	End       time.Time
}

func getLoadBalancerCosts(resLBCost, resLBActiveMins []*prom.QueryResult, resolution time.Duration) map[serviceKey]*LB {
	lbMap := make(map[serviceKey]*LB)
	lbHourlyCosts := make(map[serviceKey]float64)
	for _, res := range resLBCost {
		serviceKey, err := resultServiceKey(res, env.GetPromClusterLabel(), "namespace", "service_name")
		if err != nil {
			continue
		}
		lbHourlyCosts[serviceKey] = res.Values[0].Value
	}
	for _, res := range resLBActiveMins {
		serviceKey, err := resultServiceKey(res, env.GetPromClusterLabel(), "namespace", "service_name")
		if err != nil || len(res.Values) == 0 {
			continue
		}
		if _, ok := lbHourlyCosts[serviceKey]; !ok {
			log.Warningf("CostModel: failed to find hourly cost for Load Balancer: %v", serviceKey)
			continue
		}

		s := time.Unix(int64(res.Values[0].Timestamp), 0)
		// subtract resolution from start time to cover full time period
		s = s.Add(-resolution)
		e := time.Unix(int64(res.Values[len(res.Values)-1].Timestamp), 0)
		hours := e.Sub(s).Hours()

		lbMap[serviceKey] = &LB{
			TotalCost: lbHourlyCosts[serviceKey] * hours,
			Start:     s,
			End:       e,
		}
	}
	return lbMap
}

func applyLoadBalancersToPods(lbMap map[serviceKey]*LB, allocsByService map[serviceKey][]*kubecost.Allocation) {
	for sKey, lb := range lbMap {
		totalHours := 0.0
		allocHours := make(map[*kubecost.Allocation]float64)
		// Add portion of load balancing cost to each allocation
		// proportional to the total number of hours allocations used the load balancer
		for _, alloc := range allocsByService[sKey] {
			// Determine the (start, end) of the relationship between the
			// given LB and the associated Allocation so that a precise
			// number of hours can be used to compute cumulative cost.
			s, e := alloc.Start, alloc.End
			if lb.Start.After(alloc.Start) {
				s = lb.Start
			}
			if lb.End.Before(alloc.End) {
				e = lb.End
			}
			hours := e.Sub(s).Hours()
			// A negative number of hours signifies no overlap between the windows
			if hours > 0 {
				totalHours += hours
				allocHours[alloc] = hours
			}
		}

		// Distribute cost of service once total hours is calculated
		for alloc, hours := range allocHours {
			alloc.LoadBalancerCost += lb.TotalCost * hours / totalHours
		}
	}
}

// getNodePricing determines node pricing, given a key and a mapping from keys
// to their NodePricing instances, as well as the custom pricing configuration
// inherent to the CostModel instance. If custom pricing is set, use that. If
// not, use the pricing defined by the given key. If that doesn't exist, fall
// back on custom pricing as a default.
func (cm *CostModel) getNodePricing(nodeMap map[nodeKey]*NodePricing, nodeKey nodeKey) *NodePricing {
	// Find the relevant NodePricing, if it exists. If not, substitute the
	// custom NodePricing as a default.
	node, ok := nodeMap[nodeKey]
	if !ok || node == nil {
		if nodeKey.Node != "" {
			log.DedupedWarningf(5, "CostModel: failed to find node for %s", nodeKey)
		}
		return cm.getCustomNodePricing(false)
	}

	// If custom pricing is enabled and can be retrieved, override detected
	// node pricing with the custom values.
	customPricingConfig, err := cm.Provider.GetConfig()
	if err != nil {
		log.Warningf("CostModel: failed to load custom pricing: %s", err)
	}
	if cloud.CustomPricesEnabled(cm.Provider) && customPricingConfig != nil {
		return cm.getCustomNodePricing(node.Preemptible)
	}

	node.Source = "prometheus"

	// If any of the values are NaN or zero, replace them with the custom
	// values as default.
	// TODO:CLEANUP can't we parse these custom prices once? why do we store
	// them as strings like this?

	if node.CostPerCPUHr == 0 || math.IsNaN(node.CostPerCPUHr) {
		log.Warningf("CostModel: node pricing has illegal CostPerCPUHr; replacing with custom pricing: %s", nodeKey)
		cpuCostStr := customPricingConfig.CPU
		if node.Preemptible {
			cpuCostStr = customPricingConfig.SpotCPU
		}
		costPerCPUHr, err := strconv.ParseFloat(cpuCostStr, 64)
		if err != nil {
			log.Warningf("CostModel: custom pricing has illegal CPU cost: %s", cpuCostStr)
		}
		node.CostPerCPUHr = costPerCPUHr
		node.Source += "/customCPU"
	}

	if math.IsNaN(node.CostPerGPUHr) {
		log.Warningf("CostModel: node pricing has illegal CostPerGPUHr; replacing with custom pricing: %s", nodeKey)
		gpuCostStr := customPricingConfig.GPU
		if node.Preemptible {
			gpuCostStr = customPricingConfig.SpotGPU
		}
		costPerGPUHr, err := strconv.ParseFloat(gpuCostStr, 64)
		if err != nil {
			log.Warningf("CostModel: custom pricing has illegal GPU cost: %s", gpuCostStr)
		}
		node.CostPerGPUHr = costPerGPUHr
		node.Source += "/customGPU"
	}

	if node.CostPerRAMGiBHr == 0 || math.IsNaN(node.CostPerRAMGiBHr) {
		log.Warningf("CostModel: node pricing has illegal CostPerRAMHr; replacing with custom pricing: %s", nodeKey)
		ramCostStr := customPricingConfig.RAM
		if node.Preemptible {
			ramCostStr = customPricingConfig.SpotRAM
		}
		costPerRAMHr, err := strconv.ParseFloat(ramCostStr, 64)
		if err != nil {
			log.Warningf("CostModel: custom pricing has illegal RAM cost: %s", ramCostStr)
		}
		node.CostPerRAMGiBHr = costPerRAMHr
		node.Source += "/customRAM"
	}

	return node
}

// getCustomNodePricing converts the CostModel's configured custom pricing
// values into a NodePricing instance.
func (cm *CostModel) getCustomNodePricing(spot bool) *NodePricing {
	customPricingConfig, err := cm.Provider.GetConfig()
	if err != nil {
		return nil
	}

	cpuCostStr := customPricingConfig.CPU
	gpuCostStr := customPricingConfig.GPU
	ramCostStr := customPricingConfig.RAM
	if spot {
		cpuCostStr = customPricingConfig.SpotCPU
		gpuCostStr = customPricingConfig.SpotGPU
		ramCostStr = customPricingConfig.SpotRAM
	}

	node := &NodePricing{Source: "custom"}

	costPerCPUHr, err := strconv.ParseFloat(cpuCostStr, 64)
	if err != nil {
		log.Warningf("CostModel: custom pricing has illegal CPU cost: %s", cpuCostStr)
	}
	node.CostPerCPUHr = costPerCPUHr

	costPerGPUHr, err := strconv.ParseFloat(gpuCostStr, 64)
	if err != nil {
		log.Warningf("CostModel: custom pricing has illegal GPU cost: %s", gpuCostStr)
	}
	node.CostPerGPUHr = costPerGPUHr

	costPerRAMHr, err := strconv.ParseFloat(ramCostStr, 64)
	if err != nil {
		log.Warningf("CostModel: custom pricing has illegal RAM cost: %s", ramCostStr)
	}
	node.CostPerRAMGiBHr = costPerRAMHr

	return node
}

// NodePricing describes the resource costs associated with a given node, as
// well as the source of the information (e.g. prometheus, custom)
type NodePricing struct {
	Name            string
	NodeType        string
	ProviderID      string
	Preemptible     bool
	CostPerCPUHr    float64
	CostPerRAMGiBHr float64
	CostPerGPUHr    float64
	Discount        float64
	Source          string
}

// Pod describes a running pod's start and end time within a Window and
// all the Allocations (i.e. containers) contained within it.
type Pod struct {
	Window      kubecost.Window
	Start       time.Time
	End         time.Time
	Key         podKey
	Allocations map[string]*kubecost.Allocation
}

// AppendContainer adds an entry for the given container name to the Pod.
func (p Pod) AppendContainer(container string) {
	name := fmt.Sprintf("%s/%s/%s/%s", p.Key.Cluster, p.Key.Namespace, p.Key.Pod, container)

	alloc := &kubecost.Allocation{
		Name:       name,
		Properties: &kubecost.AllocationProperties{},
		Window:     p.Window.Clone(),
		Start:      p.Start,
		End:        p.End,
	}
	alloc.Properties.Container = container
	alloc.Properties.Pod = p.Key.Pod
	alloc.Properties.Namespace = p.Key.Namespace
	alloc.Properties.Cluster = p.Key.Cluster

	p.Allocations[container] = alloc
}

// PVC describes a PersistentVolumeClaim
// TODO:CLEANUP move to pkg/kubecost?
// TODO:CLEANUP add PersistentVolumeClaims field to type Allocation?
type PVC struct {
	Bytes     float64   `json:"bytes"`
	Count     int       `json:"count"`
	Name      string    `json:"name"`
	Cluster   string    `json:"cluster"`
	Namespace string    `json:"namespace"`
	Volume    *PV       `json:"persistentVolume"`
	Mounted   bool      `json:"mounted"`
	Start     time.Time `json:"start"`
	End       time.Time `json:"end"`
}

// Cost computes the cumulative cost of the PVC
func (pvc *PVC) Cost() float64 {
	if pvc == nil || pvc.Volume == nil {
		return 0.0
	}

	gib := pvc.Bytes / 1024 / 1024 / 1024
	hrs := pvc.Minutes() / 60.0

	return pvc.Volume.CostPerGiBHour * gib * hrs
}

// Minutes computes the number of minutes over which the PVC is defined
func (pvc *PVC) Minutes() float64 {
	if pvc == nil {
		return 0.0
	}

	return pvc.End.Sub(pvc.Start).Minutes()
}

// String returns a string representation of the PVC
func (pvc *PVC) String() string {
	if pvc == nil {
		return "<nil>"
	}
	return fmt.Sprintf("%s/%s/%s{Bytes:%.2f, Cost:%.6f, Start,End:%s}", pvc.Cluster, pvc.Namespace, pvc.Name, pvc.Bytes, pvc.Cost(), kubecost.NewWindow(&pvc.Start, &pvc.End))
}

// PV describes a PersistentVolume
// TODO:CLEANUP move to pkg/kubecost?
type PV struct {
	Bytes          float64 `json:"bytes"`
	CostPerGiBHour float64 `json:"costPerGiBHour"`
	Cluster        string  `json:"cluster"`
	Name           string  `json:"name"`
	StorageClass   string  `json:"storageClass"`
}

// String returns a string representation of the PV
func (pv *PV) String() string {
	if pv == nil {
		return "<nil>"
	}
	return fmt.Sprintf("%s/%s{Bytes:%.2f, Cost/GiB*Hr:%.6f, StorageClass:%s}", pv.Cluster, pv.Name, pv.Bytes, pv.CostPerGiBHour, pv.StorageClass)
}