Merge pull request #4222 from fisherxu/automated-cherry-pick-of-#4160-upstream-release-1.9v1.9.5origin/release-1.9

Automated cherry pick of #4160: fix: Ignore the cache timestamp of the MachineInfo Metrics

1 files changed, 11 insertions, 0 deletions

diff --git a/edge/pkg/edged/edged.go b/edge/pkg/edged/edged.go
index 32e847ced..e402f4c82 100644
--- a/edge/pkg/edged/edged.go
+++ b/edge/pkg/edged/edged.go
@@ -534,6 +534,17 @@ func newEdged(enable bool) (*edged, error) {
 	if ed.machineInfo, err = ed.newMachineInfo(); err != nil {
 		return nil, err
 	}
+	// Avoid adding timestamp to machine metrics when they're cached
+	// Please refer to the following K8s PRs for more details:
+	//    link: https://github.com/kubernetes/kubernetes/pull/95210#issuecomment-726143798
+	//    link: https://github.com/kubernetes/kubernetes/pull/97006
+	// Analysis of causes :
+	//   1. The cadvisor collects machine metrics when it launches and then caches them for later scrape.
+	//      It is assumed that as long as the machine is not restarted, the hardware will not change and so do the machine metrics.
+	//   2. This timestamp of the machine metrics should be set when they're scraped, not when they're cached.
+	//   3. If using the timestamp when the machine metrics are cached, the timestamp when these metrics are scraped will be outdated
+	//      which may cause issues like 'out of order sample' when sending these metrics to prometheus.
+	ed.machineInfo.Timestamp = time.Time{}
 
 	// create a log manager
 	logManager, err := logs.NewContainerLogManager(runtimeService, ed.os, "10Mi", 5)