kaisun2000 commented on a change in pull request #1035:
URL: https://github.com/apache/helix/pull/1035#discussion_r435711656
##########
File path:
helix-core/src/test/java/org/apache/helix/integration/TestZkCallbackHandlerLeak.java
##########
@@ -317,6 +324,150 @@ public boolean verify() throws Exception {
System.out.println("END " + clusterName + " at " + new
Date(System.currentTimeMillis()));
}
+ @Test
+ public void testDanglingCallbackHanlderFix() throws Exception {
+ String className = TestHelper.getTestClassName();
+ String methodName = TestHelper.getTestMethodName();
+ String clusterName = className + "_" + methodName;
+ final int n = 3;
+ final String zkAddr = ZK_ADDR;
+ System.out.println("START " + clusterName + " at " + new
Date(System.currentTimeMillis()));
+
+ TestHelper.setupCluster(clusterName, zkAddr, 12918, "localhost", "TestDB",
1, // resource
+ 32, // partitions
+ n, // nodes
+ 2, // replicas
+ "MasterSlave", true);
+
+ final ClusterControllerManager controller =
+ new ClusterControllerManager(zkAddr, clusterName, "controller_0");
+ controller.syncStart();
+
+ MockParticipantManager[] participants = new MockParticipantManager[n];
+ for (int i = 0; i < n; i++) {
+ String instanceName = "localhost_" + (12918 + i);
+ participants[i] = new MockParticipantManager(zkAddr, clusterName,
instanceName);
+ participants[i].syncStart();
+ }
+
+ Boolean result =
+ ClusterStateVerifier
+ .verifyByZkCallback(new
ClusterStateVerifier.BestPossAndExtViewZkVerifier(zkAddr,
+ clusterName));
+ Assert.assertTrue(result);
+
+ //HelixManager rpManager = HelixManagerFactory
+ // .getZKHelixManager(clusterName, "", InstanceType.SPECTATOR, ZK_ADDR);
+ //rpManager.connect();
+ ClusterSpectatorManager rpManager = new ClusterSpectatorManager(ZK_ADDR,
clusterName, "router");
+ rpManager.syncStart();
+ RoutingTableProvider rp = new RoutingTableProvider(rpManager,
PropertyType.CURRENTSTATES);
+
+ Thread.sleep(5000);
Review comment:
@jiajunwang. This is a good one for discussion as how best can we make
the test more stable.
In this sleep(), we wait for RP to get to a stable state of installing the
watches, callbackhandlers. However, previous verifyWithBestPossible whatever is
waiting for cluster to be stable. That is for controller. As far I know there
is not equivalent for RP. And is there a way we can do similar thing for RP to
reach a stable state?
line 374 sleep(). We need to wait till RP's zkHelixManager finish
handleAllEvent(). By that time, it would have duplicated CB.
line 389, we need to wait till participant session expiration finishing
handleNewSession(), and RP's leaked version of CB's handleChildrenChange()
finishing processing. This reproduce the dangling CB's leak on server.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
