Fix: Network Manager race condition

integration_test/ops_agent_test.go

@@ -3953,8 +3953,9 @@ func checkExpectedHealthCheckResult(t *testing.T, output string, name string, ex
 func getRecentServiceOutputForPlatform(platform string) string {
 	if gce.IsWindows(platform) {
 		cmd := strings.Join([]string{
-			"$Past = (Get-Date) - (New-TimeSpan -Minute 1)",
-			"Get-WinEvent -MaxEvents 10 -FilterHashtable @{ Logname='Application'; ProviderName='google-cloud-ops-agent'; StartTime=$Past } | select -ExpandProperty Message",
+			"$ServiceStart = (Get-EventLog -LogName 'System' -Source 'Service Control Manager' -EntryType 'Information' -Message '*Google Cloud Ops Agent service entered the running state*' -Newest 1).TimeGenerated",
+			"$QueryStart = $ServiceStart - (New-TimeSpan -Seconds 30)",
+			"Get-WinEvent -MaxEvents 10 -FilterHashtable @{ Logname='Application'; ProviderName='google-cloud-ops-agent'; StartTime=$QueryStart } | select -ExpandProperty Message",
 		}, ";")
 		return cmd
 	}
@@ -4233,6 +4234,41 @@ func TestNoNvmlOtelReceiverWithoutGpu(t *testing.T) {
 	})
 }
 
+func TestRestartVM(t *testing.T) {
+	t.Parallel()
+	gce.RunForEachPlatform(t, func(t *testing.T, platform string) {
+		t.Parallel()
+
+		ctx, logger, vm := agents.CommonSetup(t, platform)
+		if err := agents.SetupOpsAgent(ctx, logger.ToMainLog(), vm, ""); err != nil {
+			t.Fatal(err)
+		}
+
+		cmdOut, err := gce.RunRemotely(ctx, logger.ToMainLog(), vm, "", getRecentServiceOutputForPlatform(vm.Platform))
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		// Ensure sure all healthchecks pass before the restart
+		checkExpectedHealthCheckResult(t, cmdOut.Stdout, "Network", "PASS", "")
+		checkExpectedHealthCheckResult(t, cmdOut.Stdout, "Ports", "PASS", "")
+		checkExpectedHealthCheckResult(t, cmdOut.Stdout, "API", "PASS", "")
+
+		if err := gce.RestartInstance(ctx, logger, vm); err != nil {
+			t.Fatal(err)
+		}
+
+		cmdOut, err = gce.RunRemotely(ctx, logger.ToMainLog(), vm, "", getRecentServiceOutputForPlatform(vm.Platform))
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		checkExpectedHealthCheckResult(t, cmdOut.Stdout, "Network", "PASS", "")
+		checkExpectedHealthCheckResult(t, cmdOut.Stdout, "Ports", "PASS", "")
+		checkExpectedHealthCheckResult(t, cmdOut.Stdout, "API", "PASS", "")
+	})
+}
+
 func TestMain(m *testing.M) {
 	code := m.Run()
 	gce.CleanupKeysOrDie()

internal/healthchecks/network_check.go

@@ -17,6 +17,7 @@ package healthchecks
 import (
 	"errors"
 	"net/http"
+	"time"
 
 	"github.com/GoogleCloudPlatform/ops-agent/internal/logs"
 )
@@ -29,7 +30,15 @@ type networkRequest struct {
 }
 
 func (r networkRequest) SendRequest(logger logs.StructuredLogger) error {
-	response, err := http.Get(r.url)
+	var response *http.Response
+	var err error
+	for i := 0; i < 5; i++ {
+		response, err = http.Get(r.url)
+		if err == nil && response.StatusCode == http.StatusOK {
+			break
+		}
+		time.Sleep(1 * time.Second)
+	}
 	if err != nil {
 		if isTimeoutError(err) || isConnectionRefusedError(err) {
 			return r.healthCheckError

systemd/google-cloud-ops-agent.service

@@ -14,7 +14,8 @@
 
 [Unit]
 Description=Google Cloud Ops Agent
-Wants=google-cloud-ops-agent-fluent-bit.service google-cloud-ops-agent-opentelemetry-collector.service google-cloud-ops-agent-diagnostics.service
+Wants=google-cloud-ops-agent-fluent-bit.service google-cloud-ops-agent-opentelemetry-collector.service google-cloud-ops-agent-diagnostics.service network-online.target
+After=network-online.target
 
 [Service]
 Type=oneshot