Benjamin Mahler created MESOS-7569:
--------------------------------------
Summary: Allow "old" executors with half-open connections to be
preserved during agent upgrade / restart.
Key: MESOS-7569
URL: https://issues.apache.org/jira/browse/MESOS-7569
Project: Mesos
Issue Type: Bug
Components: agent
Reporter: Benjamin Mahler
Users who have executors in their cluster without the fix to MESOS-7057 will
experience these executors potentially being destroyed whenever the agent
restarts (or is upgraded).
This occurs when these old executors have connections idle for > 5 days
(default conntrack tcp timeout). At this point, the connection is timedout and
no longer tracked by conntrack. From what we've seen, if the agent stays up,
the packets still flow between the executor and agent. However, once the agent
restarts, in some cases (presence of a DROP rule, or some flavors of NATing),
the executor does not receive the RST/FIN from the kernel and will hold a
half-open TCP connection. At this point, when the executor responds to the
reconnect message from the restarted agent, it's half-open TCP connection
closes, and the executor will be destroyed by the agent.
In order to allow users to preserve the tasks running in these "old" executors
(i.e. without the MESOS-7057 fix), we can add *optional* retrying of the
reconnect message in the agent. This allows the old executor to correctly
establish a link to agent, when the second reconnect message is handled.
Longer term, heartbeating or TCP keepalives will prevent the connections from
reaching the conntrack timeout (see MESOS-7568).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
