Collect cgroup stats one last time before exit

This commit adds a collection step in the Stop() task handler which will
retrieve the metrics available for this container at that time, and
store them until the next prometheus Collect() cycle.

This allows short-lived containers to be visible in prometheus, which
would otherwise be ignored (for example, running containerd-stress would
show something like 2 or 3 containers in the end, while now we can see
all of them). It also allows for more accurate collection when
long-running containers end (for example CPU usage could spike in the
last few seconds).

A simple case illustrating this with cpu usage would be:

  ctr run -t --rm docker.io/library/alpine:latest mycontainer sh -c 'yes > /dev/null & sleep 3 && pkill yes'

Signed-off-by: Mathieu Pasquet <mathieu.pasquet@alterway.fr>

metrics/cgroups/cgroups.go

@@ -70,6 +70,8 @@ func (m *cgroupsMonitor) Monitor(c runtime.Task) error {
 
 func (m *cgroupsMonitor) Stop(c runtime.Task) error {
 	info := c.Info()
+	t := c.(*linux.Task)
+	m.collector.collect(info.ID, info.Namespace, t.Cgroup(), m.collector.storedMetrics, nil)
 	m.collector.Remove(info.ID, info.Namespace)
 	return nil
 }

metrics/cgroups/metric.go

@@ -31,6 +31,10 @@ func (m *metric) desc(ns *metrics.Namespace) *prometheus.Desc {
 func (m *metric) collect(id, namespace string, stats *cgroups.Metrics, ns *metrics.Namespace, ch chan<- prometheus.Metric) {
 	values := m.getValues(stats)
 	for _, v := range values {
-		ch <- prometheus.MustNewConstMetric(m.desc(ns), m.vt, v.v, append([]string{id, namespace}, v.l...)...)
+		select {
+		case ch <- prometheus.MustNewConstMetric(m.desc(ns), m.vt, v.v, append([]string{id, namespace}, v.l...)...):
+		default:
+			break
+		}
 	}
 }

metrics/cgroups/metrics.go

@@ -40,6 +40,7 @@ func newCollector(ns *metrics.Namespace) *collector {
 	c.metrics = append(c.metrics, memoryMetrics...)
 	c.metrics = append(c.metrics, hugetlbMetrics...)
 	c.metrics = append(c.metrics, blkioMetrics...)
+	c.storedMetrics = make(chan prometheus.Metric, 100*len(c.metrics))
 	ns.Add(c)
 	return c
 }
@@ -62,6 +63,7 @@ type collector struct {
 	cgroups       map[string]*task
 	ns            *metrics.Namespace
 	metrics       []*metric
+	storedMetrics chan prometheus.Metric
 }
 
 func (c *collector) Describe(ch chan<- *prometheus.Desc) {
@@ -77,12 +79,23 @@ func (c *collector) Collect(ch chan<- prometheus.Metric) {
 		wg.Add(1)
 		go c.collect(t.id, t.namespace, t.cgroup, ch, wg)
 	}
+SelectLoop:
+	for {
+		select {
+		case value := <-c.storedMetrics:
+			ch <- value
+		default:
+			break SelectLoop
+		}
+	}
 	c.mu.RUnlock()
 	wg.Wait()
 }
 
 func (c *collector) collect(id, namespace string, cg cgroups.Cgroup, ch chan<- prometheus.Metric, wg *sync.WaitGroup) {
+	if wg != nil {
 		defer wg.Done()
+	}
 	stats, err := cg.Stat(cgroups.IgnoreNotExist)
 	if err != nil {
 		logrus.WithError(err).Errorf("stat cgroup %s", id)