Benjamin Bannier created MESOS-8582:

             Summary: Add a way to make sure an agent always knows the full 
framework information of all frameworks executing operations on its resources
                 Key: MESOS-8582
             Project: Mesos
          Issue Type: Bug
          Components: agent, master, storage
    Affects Versions: 1.5.0
            Reporter: Benjamin Bannier

Currently an {{Operation}} only contains a {{FrameworkID}} of originating 
frameworks, but e.g., not the full {{FrameworkInfo}}. This is problematic in 
master failover scenarios where a master might learn about an operation 
triggered by a framework unknown to it. The way the master implementation is 
structured, we would like to create tracking structures for that framework 
(e.g., to sync with the allocator down the line), but cannot do so since we can 
only learn this information when either the framework reregisters, or an agent 
running tasks of that framework reconciles with the master. We also cannot use 
conjured uo dummy information until we  learn the true {{FrameworkInfo}} since 
some required fields in {{FrameworkInfo}} (namely {{FrameworkInfo.user}}) 
cannot be updated, see MESOS-703.

We should introduce a channel for agents to learn the full {{FrameworkInfo}} 
for all frameworks executing operations on its resources. For simplicity and 
symmetry with {{RunTaskMessage}} it seems that adding an explicit 
{{FrameworkInfo}} field to {{Operation}} would do the job (e.g., allow atomic 
information transfer when operations are sent to the agent or on reconciliation 
with newly elected masters.

This message was sent by Atlassian JIRA

Reply via email to