Merge pull request #138637 from aojea/automated-cherry-pick-of-#138571-upstream-release-1.34

Automated cherry pick of #138571: kube-proxy: don't do full periodic syncs on large cluster mode

pkg/proxy/iptables/proxier.go

@@ -745,7 +745,9 @@ func (proxier *Proxier) syncProxyRules() (retryError error) {
 	// Keep track of how long syncs take.
 	start := time.Now()
 
-	doFullSync := proxier.needFullSync || (time.Since(proxier.lastFullSync) > proxyutil.FullSyncPeriod)
+	doFullSync := proxier.needFullSync ||
+		// Avoid regular full syncs for large clusters.
+		((time.Since(proxier.lastFullSync) > proxyutil.FullSyncPeriod) && !proxier.largeClusterMode)
 
 	defer func() {
 		metrics.SyncProxyRulesLatency.WithLabelValues(string(proxier.ipFamily)).Observe(metrics.SinceInSeconds(start))

pkg/proxy/iptables/proxier_test.go

@@ -5644,6 +5644,21 @@ func TestSyncProxyRulesLargeClusterMode(t *testing.T) {
 		t.Errorf("numComments (%d) != 0 after partial resync when numEndpoints (%d) > threshold (%d)", numComments, expectedEndpoints+3, largeClusterEndpointsThreshold)
 	}
 
+	// Even if FullSyncPeriod has elapsed, large-cluster mode should keep this as
+	// a partial resync when there are no explicit changes requiring a full sync.
+	if !fp.largeClusterMode {
+		t.Fatalf("expected to be in large cluster mode")
+	}
+	expectedLastFullSync := time.Now().Add(-proxyutil.FullSyncPeriod).Add(-time.Second)
+	fp.lastFullSync = expectedLastFullSync
+	err := fp.syncProxyRules()
+	if err != nil {
+		t.Fatalf("syncProxyRules failed: %v", err)
+	}
+	if !fp.lastFullSync.Equal(expectedLastFullSync) {
+		t.Fatalf("expected periodic sync in large cluster mode to skip full sync: lastFullSync changed from %v to %v", expectedLastFullSync, fp.lastFullSync)
+	}
+
 	// Now force a full resync and confirm that it rewrites the older services with
 	// no comments as well.
 	fp.forceSyncProxyRules()