virtual-kubelet/providers/azure/metrics.go

package azure

import (
	"context"
	"strings"
	"time"

	"github.com/cpuguy83/strongerrors/status/ocstatus"
	"github.com/pkg/errors"
	"github.com/virtual-kubelet/virtual-kubelet/providers/azure/client/aci"
	"go.opencensus.io/trace"
	"golang.org/x/sync/errgroup"
	"k8s.io/api/core/v1"
	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
	stats "k8s.io/kubernetes/pkg/kubelet/apis/stats/v1alpha1"
)

// GetStatsSummary returns the stats summary for pods running on ACI
func (p *ACIProvider) GetStatsSummary(ctx context.Context) (summary *stats.Summary, err error) {
	ctx, span := trace.StartSpan(ctx, "GetSummaryStats")
	defer span.End()
	addAzureAttributes(span, p)

	p.metricsSync.Lock()
	defer p.metricsSync.Unlock()
	span.Annotate(nil, "acquired metrics mutex")

	if time.Now().Sub(p.metricsSyncTime) < time.Minute {
		span.AddAttributes(trace.BoolAttribute("preCachedResult", true), trace.StringAttribute("cachedResultSampleTime", p.metricsSyncTime.String()))
		return p.lastMetric, nil
	}
	span.AddAttributes(trace.BoolAttribute("preCachedResult", false), trace.StringAttribute("cachedResultSampleTime", p.metricsSyncTime.String()))

	select {
	case <-ctx.Done():
		return nil, ctx.Err()
	default:
	}

	defer func() {
		if err != nil {
			return
		}
		p.lastMetric = summary
		p.metricsSyncTime = time.Now()
	}()

	pods := p.resourceManager.GetPods()

	var errGroup errgroup.Group
	chResult := make(chan stats.PodStats, len(pods))

	end := time.Now()
	start := end.Add(-1 * time.Minute)

	sema := make(chan struct{}, 10)
	for _, pod := range pods {
		if pod.Status.Phase != v1.PodRunning {
			continue
		}
		pod := pod
		errGroup.Go(func() error {
			ctx, span := trace.StartSpan(ctx, "getPodMetrics")
			defer span.End()
			span.AddAttributes(
				trace.StringAttribute("UID", string(pod.UID)),
				trace.StringAttribute("Name", pod.Name),
				trace.StringAttribute("Namespace", pod.Namespace),
			)

			select {
			case <-ctx.Done():
				return ctx.Err()
			case sema <- struct{}{}:
			}
			defer func() {
				<-sema
			}()

			span.Annotate(nil, "Acquired semaphore")

			cgName := containerGroupName(pod)
			// cpu/mem and net stats are split because net stats do not support container level detail
			systemStats, err := p.aciClient.GetContainerGroupMetrics(ctx, p.resourceGroup, cgName, aci.MetricsRequest{
				Dimension:    "containerName eq '*'",
				Start:        start,
				End:          end,
				Aggregations: []aci.AggregationType{aci.AggregationTypeAverage},
				Types:        []aci.MetricType{aci.MetricTypeCPUUsage, aci.MetricTypeMemoryUsage},
			})
			if err != nil {
				span.SetStatus(ocstatus.FromError(err))
				return errors.Wrapf(err, "error fetching cpu/mem stats for container group %s", cgName)
			}
			span.Annotate(nil, "Got system stats")

			netStats, err := p.aciClient.GetContainerGroupMetrics(ctx, p.resourceGroup, cgName, aci.MetricsRequest{
				Start:        start,
				End:          end,
				Aggregations: []aci.AggregationType{aci.AggregationTypeAverage},
				Types:        []aci.MetricType{aci.MetricTyperNetworkBytesRecievedPerSecond, aci.MetricTyperNetworkBytesTransmittedPerSecond},
			})
			if err != nil {
				span.SetStatus(ocstatus.FromError(err))
				return errors.Wrapf(err, "error fetching network stats for container group %s", cgName)
			}
			span.Annotate(nil, "Got network stats")

			chResult <- collectMetrics(pod, systemStats, netStats)
			return nil
		})
	}

	if err := errGroup.Wait(); err != nil {
		return nil, errors.Wrap(err, "error in request to fetch container group metrics")
	}
	close(chResult)
	span.Annotate([]trace.Attribute{trace.Int64Attribute("nPods", int64(len(pods)))}, "Collected stats from Azure")

	var s stats.Summary
	s.Node = stats.NodeStats{
		NodeName: p.nodeName,
	}
	s.Pods = make([]stats.PodStats, 0, len(chResult))

	for stat := range chResult {
		s.Pods = append(s.Pods, stat)
	}

	return &s, nil
}

func collectMetrics(pod *v1.Pod, system, net *aci.ContainerGroupMetricsResult) stats.PodStats {
	var stat stats.PodStats
	containerStats := make(map[string]*stats.ContainerStats, len(pod.Status.ContainerStatuses))
	stat.StartTime = pod.CreationTimestamp

	for _, m := range system.Value {
		// cpu/mem stats are per container, so each entry in the time series is for a container, not the container group.
		for _, entry := range m.Timeseries {
			if len(entry.Data) == 0 {
				continue
			}

			var cs *stats.ContainerStats
			for _, v := range entry.MetadataValues {
				if strings.ToLower(v.Name.Value) != "containername" {
					continue
				}
				if cs = containerStats[v.Value]; cs == nil {
					cs = &stats.ContainerStats{Name: v.Value, StartTime: stat.StartTime}
					containerStats[v.Value] = cs
				}
			}
			if cs == nil {
				continue
			}

			if stat.Containers == nil {
				stat.Containers = make([]stats.ContainerStats, 0, len(containerStats))
			}

			data := entry.Data[len(entry.Data)-1] // get only the last entry
			switch m.Desc.Value {
			case aci.MetricTypeCPUUsage:
				if cs.CPU == nil {
					cs.CPU = &stats.CPUStats{}
				}

				// average is the average number of millicores over a 1 minute interval (which is the interval we are pulling the stats for)
				nanoCores := uint64(data.Average * 1000000)
				usageNanoSeconds := nanoCores * 60
				cs.CPU.Time = metav1.NewTime(data.Timestamp)
				cs.CPU.UsageCoreNanoSeconds = &usageNanoSeconds
				cs.CPU.UsageNanoCores = &nanoCores

				if stat.CPU == nil {
					var zero uint64
					stat.CPU = &stats.CPUStats{UsageNanoCores: &zero, UsageCoreNanoSeconds: &zero, Time: metav1.NewTime(data.Timestamp)}
				}
				podCPUSec := *stat.CPU.UsageCoreNanoSeconds
				podCPUSec += usageNanoSeconds
				stat.CPU.UsageCoreNanoSeconds = &podCPUSec

				podCPUCore := *stat.CPU.UsageNanoCores
				podCPUCore += nanoCores
				stat.CPU.UsageNanoCores = &podCPUCore
			case aci.MetricTypeMemoryUsage:
				if cs.Memory == nil {
					cs.Memory = &stats.MemoryStats{}
				}
				cs.Memory.Time = metav1.NewTime(data.Timestamp)
				bytes := uint64(data.Average)
				cs.Memory.UsageBytes = &bytes
				cs.Memory.WorkingSetBytes = &bytes

				if stat.Memory == nil {
					var zero uint64
					stat.Memory = &stats.MemoryStats{UsageBytes: &zero, WorkingSetBytes: &zero, Time: metav1.NewTime(data.Timestamp)}
				}
				podMem := *stat.Memory.UsageBytes
				podMem += bytes
				stat.Memory.UsageBytes = &podMem
				stat.Memory.WorkingSetBytes = &podMem
			}
		}
	}

	for _, m := range net.Value {
		if stat.Network == nil {
			stat.Network = &stats.NetworkStats{}
		}
		// network stats are for the whole container group, so there should only be one entry here.
		if len(m.Timeseries) == 0 {
			continue
		}
		entry := m.Timeseries[0]
		if len(entry.Data) == 0 {
			continue
		}
		data := entry.Data[len(entry.Data)-1] // get only the last entry

		bytes := uint64(data.Average)
		switch m.Desc.Value {
		case aci.MetricTyperNetworkBytesRecievedPerSecond:
			stat.Network.RxBytes = &bytes
		case aci.MetricTyperNetworkBytesTransmittedPerSecond:
			stat.Network.TxBytes = &bytes
		}
		stat.Network.Time = metav1.NewTime(data.Timestamp)
		stat.Network.InterfaceStats.Name = "eth0"
	}

	for _, cs := range containerStats {
		stat.Containers = append(stat.Containers, *cs)
	}

	stat.PodRef = stats.PodReference{
		Name:      pod.Name,
		Namespace: pod.Namespace,
		UID:       string(pod.UID),
	}

	return stat
}