Andrei Budnik created MESOS-9054:
------------------------------------

             Summary: Scheduler driver hangs on syncing its state with 
ZooKeeper during Master detection
                 Key: MESOS-9054
                 URL: https://issues.apache.org/jira/browse/MESOS-9054
             Project: Mesos
          Issue Type: Bug
          Components: scheduler driver
            Reporter: Andrei Budnik


A framework (namely, Marathon) uses scheduler driver (V0 API) to connect to the 
Mesos master, but never receives `registered()`. The hanging framework prints 
the following messages:
{code:java}
2018-06-26 05:30:23: I0626 05:30:23.899340 14465 sched.cpp:232] Version: 1.4.0
2018-06-26 05:30:24: I0626 05:30:24.022102 14523 group.cpp:341] Group process 
(zookeeper-group(1)@10.136.5.234:15101) connected to ZooKeeper
2018-06-26 05:30:24: I0626 05:30:24.022148 14523 group.cpp:831] Syncing group 
operations: queue size (joins, cancels, datas) = (0, 0, 0)
2018-06-26 05:30:24: I0626 05:30:24.022166 14523 group.cpp:419] Trying to 
create path '/mesos' in ZooKeeper{code}
When the framework calls `scheduler_driver->start()`, it creates and spawns 
`ZooKeeperMasterDetectorProcess`, which creates a detector of type 
`zookeeper::Group`. After the detector connects to ZK, it 
[calls|https://github.com/apache/mesos/blob/77921200a69564f966917bfc8e07a3d1e3ada196/include/mesos/zookeeper/watcher.hpp#L49-L52]
 `zookeeper::Group::connected()`. Then, `Group` tries to 
`[sync()|https://github.com/apache/mesos/blob/77921200a69564f966917bfc8e07a3d1e3ada196/src/zookeeper/group.cpp#L374]`,
 which calls 
`[create()|https://github.com/apache/mesos/blob/77921200a69564f966917bfc8e07a3d1e3ada196/src/zookeeper/group.cpp#L851]`.
 At this point we call 
`[zk->create()|https://github.com/apache/mesos/blob/77921200a69564f966917bfc8e07a3d1e3ada196/src/zookeeper/group.cpp#L421]`,
 which is a *synchronous* call, see 
[`dispatch(...).get()`|https://github.com/apache/mesos/blob/77921200a69564f966917bfc8e07a3d1e3ada196/src/zookeeper/zookeeper.cpp#L602-L610].

Since ZK library or ZK itself might hang, a scheduler driver can stuck in this 
state, so a framework will never receive any callbacks.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to