Merge pull request #2829 from kragniz/prometheus-oom-count

Expose OOM event count to prometheus

Author: iwankgb

cmd/cadvisor.go

@@ -112,6 +112,7 @@ var (
 		container.CPUTopologyMetrics:             struct{}{},
 		container.ResctrlMetrics:                 struct{}{},
 		container.CPUSetMetrics:                  struct{}{},
+		container.OOMMetrics:                     struct{}{},
 	}
 )
 

cmd/cadvisor_test.go

@@ -109,6 +109,7 @@ func TestToIncludedMetrics(t *testing.T) {
 			container.CPUTopologyMetrics:             struct{}{},
 			container.ResctrlMetrics:                 struct{}{},
 			container.CPUSetMetrics:                  struct{}{},
+			container.OOMMetrics:                     struct{}{},
 		},
 		container.AllMetrics,
 		{},

container/factory.go

@@ -64,6 +64,7 @@ const (
 	CPUTopologyMetrics             MetricKind = "cpu_topology"
 	ResctrlMetrics                 MetricKind = "resctrl"
 	CPUSetMetrics                  MetricKind = "cpuset"
+	OOMMetrics                     MetricKind = "oom_event"
 )
 
 // AllMetrics represents all kinds of metrics that cAdvisor supported.
@@ -89,6 +90,7 @@ var AllMetrics = MetricSet{
 	CPUTopologyMetrics:             struct{}{},
 	ResctrlMetrics:                 struct{}{},
 	CPUSetMetrics:                  struct{}{},
+	OOMMetrics:                     struct{}{},
 }
 
 func (mk MetricKind) String() string {

info/v1/container.go

@@ -963,6 +963,8 @@ type ContainerStats struct {
 	Resctrl ResctrlStats `json:"resctrl,omitempty"`
 
 	CpuSet CPUSetStats `json:"cpuset,omitempty"`
+
+	OOMEvents uint64 `json:"oom_events,omitempty"`
 }
 
 func timeEq(t1, t2 time.Time, tolerance time.Duration) bool {

manager/container.go

@@ -27,6 +27,7 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/google/cadvisor/cache/memory"
@@ -102,6 +103,8 @@ type containerData struct {
 
 	// resctrlCollector updates stats for resctrl controller.
 	resctrlCollector stats.Collector
+
+	oomEvents uint64
 }
 
 // jitter returns a time.Duration between duration and duration + maxFactor * duration,
@@ -668,6 +671,9 @@ func (cd *containerData) updateStats() error {
 			klog.V(2).Infof("Failed to add summary stats for %q: %v", cd.info.Name, err)
 		}
 	}
+
+	stats.OOMEvents = atomic.LoadUint64(&cd.oomEvents)
+
 	var customStatsErr error
 	cm := cd.collectorManager.(*collector.GenericCollectorManager)
 	if len(cm.Collectors) > 0 {

manager/manager.go

@@ -24,6 +24,7 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/google/cadvisor/accelerators"
@@ -35,7 +36,7 @@ import (
 	"github.com/google/cadvisor/events"
 	"github.com/google/cadvisor/fs"
 	info "github.com/google/cadvisor/info/v1"
-	"github.com/google/cadvisor/info/v2"
+	v2 "github.com/google/cadvisor/info/v2"
 	"github.com/google/cadvisor/machine"
 	"github.com/google/cadvisor/nvm"
 	"github.com/google/cadvisor/perf"
@@ -1237,6 +1238,24 @@ func (m *manager) watchForNewOoms() error {
 			if err != nil {
 				klog.Errorf("failed to add OOM kill event for %q: %v", oomInstance.ContainerName, err)
 			}
+
+			// Count OOM events for later collection by prometheus
+			request := v2.RequestOptions{
+				IdType: v2.TypeName,
+				Count:  1,
+			}
+			conts, err := m.getRequestedContainers(oomInstance.ContainerName, request)
+			if err != nil {
+				klog.V(2).Infof("failed getting container info for %q: %v", oomInstance.ContainerName, err)
+				continue
+			}
+			if len(conts) != 1 {
+				klog.V(2).Info("Expected the request to match only one container")
+				continue
+			}
+			for _, cont := range conts {
+				atomic.AddUint64(&cont.oomEvents, 1)
+			}
 		}
 	}()
 	return nil

metrics/prometheus.go

@@ -1757,6 +1757,17 @@ func NewPrometheusCollector(i infoProvider, f ContainerLabelsFunc, includedMetri
 			},
 		}...)
 	}
+	if includedMetrics.Has(container.OOMMetrics) {
+		c.containerMetrics = append(c.containerMetrics, containerMetric{
+			name:      "container_oom_events_total",
+			help:      "Count of out of memory events observed for the container",
+			valueType: prometheus.CounterValue,
+			getValues: func(s *info.ContainerStats) metricValues {
+				return metricValues{{value: float64(s.OOMEvents), timestamp: s.Timestamp}}
+			},
+		})
+	}
+
 	return c
 }
 

metrics/testdata/prometheus_metrics

@@ -358,6 +358,9 @@ container_network_udp_usage_total{container_env_foo_env="prod",container_label_f
 container_network_udp_usage_total{container_env_foo_env="prod",container_label_foo_label="bar",id="testcontainer",image="test",name="testcontaineralias",udp_state="listen",zone_name="hello"} 0 1395066363000
 container_network_udp_usage_total{container_env_foo_env="prod",container_label_foo_label="bar",id="testcontainer",image="test",name="testcontaineralias",udp_state="rxqueued",zone_name="hello"} 0 1395066363000
 container_network_udp_usage_total{container_env_foo_env="prod",container_label_foo_label="bar",id="testcontainer",image="test",name="testcontaineralias",udp_state="txqueued",zone_name="hello"} 0 1395066363000
+# HELP container_oom_events_total Count of out of memory events observed for the container
+# TYPE container_oom_events_total counter
+container_oom_events_total{container_env_foo_env="prod",container_label_foo_label="bar",id="testcontainer",image="test",name="testcontaineralias",zone_name="hello"} 0 1395066363000
 # HELP container_perf_events_total Perf event metric.
 # TYPE container_perf_events_total counter
 container_perf_events_total{container_env_foo_env="prod",container_label_foo_label="bar",cpu="0",event="instructions",id="testcontainer",image="test",name="testcontaineralias",zone_name="hello"} 123 1395066363000