James Peach created MESOS-8405:
----------------------------------
Summary: Update master task loss handling.
Key: MESOS-8405
URL: https://issues.apache.org/jira/browse/MESOS-8405
Project: Mesos
Issue Type: Bug
Reporter: James Peach
>From [~agentvindo.dev] in [r/64940|https://reviews.apache.org/r/64940/]:
{quote}
Ideally, we want terminal but unacknowledged tasks to still be marked
unreachable in some way, either via task state being TASK_UNREACHABLE or task
being present in unreachableTasks. This allows, for example, the WebUI to not
show sandbox links for unreachable tasks irrespective of whether they were
terminal or not before going unreachable.
But doing this is tricky for various reasons:
--> updateTask() doesn't allow a terminal state to be transitioned to
TASK_UNREACHABLE. Right now when we call updateTask for a terminal task, it
adds TASK_UNREACHABLE status to Task.statuses and also sends it to operator API
stream subscribers which looks incorrect. The fact that updateTask internally
deals with already terminal tasks is a bad design decision in retrospect. I
think the callers shouldn't call it for terminal tasks instead.
--> It's not clear to our users what a completed task means. The intention was
for this to hold a cache of terminal and acknowledged tasks for storing recent
history. The users of the WebUI probably equate "Completed Tasks" to terminal
tasks irrespective of their acknowledgement status, which is why it is
confusing for them to see terminal but unacknowledged tasks in the "Active
tasks" section in the WebUI.
--> When a framework reconciles the state of a task on an unreachable agent,
master replies with TASK_UNREACHABLE irrespective of whether the task was in a
non-terminal state or terminal but un-acknowledged state or terminal and
acknowledged state when the agent went unreachable.
I think the direction we want to go towards is
--> Completed tasks should consist of terminal unacknowledged and terminal
acknowled tasks, likely in two different data structures.
--> Unreachable tasks should consist of all non-complete tasks on an
unreachable agent. All the tasks in this map should be in TASK_UNREACHABLE
state.
{quote}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
