Chun-Hung Hsiao created MESOS-9165:
--------------------------------------
Summary: Agent should commit suicide when got inconsistent ID in
DISCONNECTED state.
Key: MESOS-9165
URL: https://issues.apache.org/jira/browse/MESOS-9165
Project: Mesos
Issue Type: Bug
Affects Versions: 1.6.0, 1.5.0, 1.7.0
Reporter: Chun-Hung Hsiao
Test {{HookTest.MasterSlaveLostHookTest}} is flaky due to the following reason.
Normally, when a new agent register to the master,
{{Slave::doReliableRegistration}} keeps sending {{RegisterSlaveMessage}} until
the agent becomes {{RUNNING}}, which happens after the agent receives the first
{{SlaveRegisteredMessage}}. If another {{SlaveRegisteredMessage}} is received
with an inconsistent agent ID, the agent will commit suicide:
[https://github.com/apache/mesos/blob/master/src/slave/slave.cpp#L1525-L1529]
In the above unit test, however, the clock was advanced to simulate a network
partition. When the agent detected that the master was unreachable, it
transitioned into {{DISCONNECTED}} state. Meanwhile, the master marked the
agent as unreachable and removed it:
[https://github.com/apache/mesos/blob/master/src/master/master.cpp#L8788]
Later on, the master processed the second {{RegisterSlaveMessage}}, and
assigned a new agent ID. The slave received the {{SlaveRegisteredMessage}} and
accepted a different ID:
[https://github.com/apache/mesos/blob/master/src/slave/slave.cpp#L1478-L1522]
This behavior is unexpected. Instead, the agent should also commit suicide in
this case.
This becomes problematic since 1.5, because the agent holds a
{{LocalResourceProviderDaemon}} that assumes the agent ID never changes within
the agent actor's lifecycle:
[https://github.com/apache/mesos/blob/master/src/resource_provider/daemon.cpp#L154]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
