[jira] [Created] (MESOS-1696) Improve reconciliation between master and slave.

["Benjamin Mahler (JIRA)" <ji...@apache.org>]

Benjamin Mahler created MESOS-1696:
--------------------------------------

             Summary: Improve reconciliation between master and slave.
                 Key: MESOS-1696
                 URL: https://issues.apache.org/jira/browse/MESOS-1696
             Project: Mesos
          Issue Type: Bug
          Components: master, slave
            Reporter: Benjamin Mahler


As we update the Master to keep tasks in memory until they are both terminal and acknowledged (MESOS-1410), the lifetime of tasks in Mesos will look as follows:

{code}
Master           Slave
 {}               {}
{Tn}              {}  // Master receives Task T, non-terminal. Forwards to slave.
{Tn}             {Tn} // Slave receives Task T, non-terminal.
{Tn}             {Tt} // Task becomes terminal on slave. Update forwarded.
{Tt}             {Tt} // Master receives update, forwards to framework.
 {}              {Tt} // Master receives ack, forwards to slave.
 {}               {}  // Slave receives ack.
{code}

In the current form of reconciliation, the slave sends to the master all tasks that are not both terminal and acknowledged. At any point in the above lifecycle, the slave's re-registration message can reach the master.

Note the following properties:

*(1)* The master may have a non-terminal task, not present in the slave's re-registration message.
*(2)* The master may have a non-terminal task, present in the slave's re-registration message but in a different state.
*(3)* The slave's re-registration message may contain a terminal unacknowledged task unknown to the master.

In the current master / slave [reconciliation|https://github.com/apache/mesos/blob/0.19.1/src/master/master.cpp#L3146] code, the master assumes that case (1) is because a launch task message was dropped, and it sends TASK_LOST. We've seen above that (1) can happen even when the task reaches the slave correctly, so this can lead to inconsistency!

After chatting with [~vinodkone], we're considering updating the reconciliation to occur as follows:


→ Slave sends all tasks that are not both terminal and acknowledged, during re-registration. This is the same as before.

→ If the master sees tasks that are missing in the slave, the master sends a reconcile message to the slave for the tasks.

→ The slave will reply to reconcile messages with the latest state, or TASK_LOST if the task is not known to it. Preferably in a retried manner, unless we update socket closure on the slave to force a re-registration.



--
This message was sent by Atlassian JIRA
(v6.2#6252)