fix: handling existing tokens in the cache

pkg/auth/cred.go

@@ -28,10 +28,9 @@ import (
 	"knative.dev/pkg/logging"
 )
 
-
-// expireEarlyTokenCredential is a wrapper around the azcore.TokenCredential that 
-// returns an earlier ExpiresOn timestamp to avoid conditions like clockSkew, or a race 
-// condition during polling. 
+// expireEarlyTokenCredential is a wrapper around the azcore.TokenCredential that
+// returns an earlier ExpiresOn timestamp to avoid conditions like clockSkew, or a race
+// condition during polling.
 // See: https://github.com/hashicorp/terraform-provider-azurerm/issues/20834 for more details
 type expireEarlyTokenCredential struct {
 	cred azcore.TokenCredential
@@ -48,8 +47,17 @@ func (w *expireEarlyTokenCredential) GetToken(ctx context.Context, options polic
 	if err != nil {
 		return azcore.AccessToken{}, err
 	}
+
 	logging.FromContext(ctx).Debug("adjusting token ExpiresOn")
-	token.ExpiresOn = time.Now().Add(2 * time.Hour)
+	twoHoursFromNow := time.Now().Add(2 * time.Hour)
+	// IMDS may have the MI token already, and have an expiration of less than 2h when we receive the token. We don't want to set that value beyond the ExpiresOn time and potentially miss a refresh
+	// So we just return earlier here. See discussion here: https://github.com/Azure/karpenter-provider-azure/pull/391/files#r1648633051
+	if token.ExpiresOn.Before(twoHoursFromNow) {
+		return token, nil
+	}
+	// If the token expires in more than 2 hours, this means we are taking in a new token with a fresh 24h expiration time or one already in the cache that hasn't been modified by us, so we want to set that to two hours so
+	// we can refresh it early to avoid the polling bugs mentioned in the above issue
+	token.ExpiresOn = twoHoursFromNow
 	return token, nil
 }