[fix][broker] Check replication cluster before starting the replicator

Fixes apache#20010

### Motivation

`PersistentTopicTest.testCreateTopicWithZombieReplicatorCursor` is flaky
because the cursor could still be created again in `startReplicator`,
which could be called by:

```
onPoliciesUpdate
  checkReplicationAndRetryOnFailure
    checkReplication
```

### Modifications

- Call `checkReplicationCluster` before calling `startReplicator`.
- Sleep for a while in the test to reduce the flakiness caused by the
  asynchronous update of the policies

pulsar-broker/src/main/java/org/apache/pulsar/broker/service/persistent/PersistentTopic.java

@@ -1542,7 +1542,13 @@ public CompletableFuture<Void> checkReplication() {
                 continue;
             }
             if (!replicators.containsKey(cluster)) {
-                futures.add(startReplicator(cluster));
+                futures.add(checkReplicationCluster(cluster).thenCompose(clusterExists -> {
+                    if (clusterExists) {
+                        return startReplicator(cluster);
+                    } else {
+                        return CompletableFuture.completedFuture(null);
+                    }
+                }));
             }
         }
 

pulsar-broker/src/test/java/org/apache/pulsar/broker/service/persistent/PersistentTopicTest.java

@@ -38,6 +38,7 @@
 import java.io.ByteArrayOutputStream;
 import java.lang.reflect.Field;
 import java.nio.charset.StandardCharsets;
+import java.time.Duration;
 import java.util.ArrayList;
 import java.util.Collection;
 import java.util.Collections;
@@ -52,6 +53,7 @@
 import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.function.Supplier;
 import lombok.Cleanup;
+import lombok.extern.slf4j.Slf4j;
 import org.apache.bookkeeper.client.LedgerHandle;
 import org.apache.bookkeeper.mledger.ManagedCursor;
 import org.apache.bookkeeper.mledger.ManagedLedger;
@@ -83,6 +85,7 @@
 import org.testng.annotations.DataProvider;
 import org.testng.annotations.Test;
 
+@Slf4j
 @Test(groups = "broker")
 public class PersistentTopicTest extends BrokerTestBase {
 
@@ -576,6 +579,10 @@ public void testCreateTopicWithZombieReplicatorCursor(boolean topicLevelPolicy)
         };
         assertEquals(getCursors.get(), Collections.singleton(conf.getReplicatorPrefix() + "." + remoteCluster));
 
+        // PersistentTopics#onPoliciesUpdate might happen in different threads, so there might be a race between two
+        // updates of the replication clusters. So here we sleep for a while to reduce the flakiness.
+        Thread.sleep(100);
+
         if (topicLevelPolicy) {
             admin.topics().setReplicationClusters(topicName, Collections.emptyList());
         } else {
@@ -584,8 +591,14 @@ public void testCreateTopicWithZombieReplicatorCursor(boolean topicLevelPolicy)
         admin.clusters().deleteCluster(remoteCluster);
         // Now the cluster and its related policy has been removed but the replicator cursor still exists
 
-        topic.initialize().get(3, TimeUnit.SECONDS);
-        Awaitility.await().atMost(3, TimeUnit.SECONDS)
-                .until(() -> !topic.getManagedLedger().getCursors().iterator().hasNext());
+        Awaitility.await().atMost(Duration.ofSeconds(10)).until(() -> {
+            log.info("Before initialize...");
+            try {
+                topic.initialize().get(3, TimeUnit.SECONDS);
+            } catch (ExecutionException e) {
+                log.warn("Failed to initialize: {}", e.getCause().getMessage());
+            }
+            return !topic.getManagedLedger().getCursors().iterator().hasNext();
+        });
     }
 }