Gerd Behrmann created CURATOR-328:
-------------------------------------
Summary: PathChildrenCache fails silently if server is unavailable
for sufficient time when client starts
Key: CURATOR-328
URL: https://issues.apache.org/jira/browse/CURATOR-328
Project: Apache Curator
Issue Type: Bug
Components: Recipes
Affects Versions: 2.10.0
Reporter: Gerd Behrmann
When initializing the PathChildrenCache, if the curator client is not yet
connected to the ZooKeeper server (e.g. the server is down or the network
connection is unavailable), then the internal initialization of the cache will
eventually fail silently and the cache stays empty even after the client
finally connects to the server and the path is populated with znodes.
The following unit test demonstrates the problem (the unit test is ugly as the
problem depends on timing, but it suffices to demonstrate the issue):
{code:java}
@Test
public void pathChildrenCacheTest() throws Exception
{
TestingServer server = new TestingServer(false);
Timing timing = new Timing();
CuratorFramework client = CuratorFrameworkFactory.newClient(
server.getConnectString(), timing.session(),
timing.connection(), new ExponentialBackoffRetry(1000, 3));
try {
new Thread() {
@Override
public void run()
{
try {
Thread.sleep(60000);
server.start();
} catch (Exception e) {
e.printStackTrace();
}
}
}.start();
client.start();
PathChildrenCache cache = new PathChildrenCache(client, "/", true);
cache.start();
client.blockUntilConnected();
client.create().creatingParentContainersIfNeeded().forPath("/baz",
new byte[] {1,2,3});
assertNotNull("/baz does not exist",
client.checkExists().forPath("/baz"));
/* Ugly hack for this test to ensure the cache got time to update
itself. */
Thread.sleep(1000);
assertNotNull("cache doesn't see /baz",
cache.getCurrentData("/baz"));
} finally {
client.close();
server.stop();
}
}
{code}
Here the server startup is delayed until some point after the curator client
was started and after the recipe has been created. Eventually the server starts
and the path is populated with data - some time is given for the cache to
update itself, yet no data is visible: The second assertion fails.
If the startup time is reduced to - say - 20 seconds, the test passes.
If the client is allowed to first connect to the server before creating the
recipe and then disconnect and reconnect after creating the recipe, then the
test passes too.
I tracked down the problem to the state change listener of the recipe: If the
connection to the server is down for long enough, the refresh call during the
background initialization will eventually fail (ensurePath throws an
exception). This isn't a problem as the recipe has a state change listener, so
it gets notified when the client eventually connects to the server. The
handleStateChange method however doesn't react to a CONNECTED event - only to a
RECONNECTED event. Thus if the client has been connected to the server in the
past, everything works, however if this is the first time it connects, the
recipe will not react to the event and thus not refresh itself.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
