mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Megha Sharma <mshar...@apple.com>
Subject Re: Review Request 61473: Do not kill non partition aware tasks.
Date Mon, 07 Aug 2017 19:16:48 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/61473/#review182310
-----------------------------------------------------------



We talked about 2 approaches and approach 2 seemed like a cleaner way to address the issue.

Approach 1: ShutdownFrameworkMessage

1. Upon agent re-registration, master will add tasks even for non-PA frameworks on this agent.
This is needed by the master to do correct resource accounting and not offer resources already
in use on this agent. We need to mutate the TaskState on the Task before adding them to the
master's data structures since the TaskState might be non-terminal when the agent sends these
tasks with ReregisterSlaveMessage. And the master has already sent TASK_LOST for these tasks
to the frameworks so we need to set the TaskState to TASK_LOST so that any future reconciliations
with the framework doesn't have this task transitioning from TASK_LOST to TASK_RUNNNG/TASK_STARTING.
This is to avoid unnecessary confusion about task state as observed by the framework but indeed
this could have happened with non-strict registry as well where the framework can actually
receive a non terminal task state update after receiving a TASK_LOST for the same task in
the past.

2. When the agent re-registers, the master will continue to send a ShutdownFrameworkMessage
to the agent to kill the tasks pertaining to non-PA frameworks on the agent as it does today.
An additional optional field will be added to the ShutdownFrameworkMessage to indicate whether
or not the shutdown was initiated internally.

3. During framework shutdown the state of the framework is set to Framework::TERMINATING which
prevents it from launching new tasks. Here, since the framework is not really terminating
so in order to allow it to launch new tasks, the agent will not set the state to terminating
if the ShutdownFrameworkMessage is generated internally.

4. The framework shutdown today doesn't generate any status updates which needs to change.
The status updates will be sent if the framework shutdown is triggered internally, this is
needed to remove the tasks of non-PA frameworks that got added when the agent re-registered
(1).

Approach 2: Do not shutdown non-PA framework when agent re-registers and let the frameworks
make the decision on what needs to be done when they receive non-terminal status updates for
tasks for which they have already received a TASK_LOST. This hopefully won't break any frameworks
since it could have happened in the past with non-strict registry as well and frameworks should
be resilient enough to handle this scenario.

Let me know if I have missed anything

- Megha Sharma


On Aug. 7, 2017, 6:23 p.m., Megha Sharma wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/61473/
> -----------------------------------------------------------
> 
> (Updated Aug. 7, 2017, 6:23 p.m.)
> 
> 
> Review request for mesos and Jiang Yan Xu.
> 
> 
> Bugs: MESOS-7215
>     https://issues.apache.org/jira/browse/MESOS-7215
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Master will not kill the tasks for non-Partition aware frameworks
> when an unreachable agent re-registers with the master.
> Master used to send a ShutdownFrameworkMessages to the agent
> to kill the tasks from non partition aware frameworks including the
> ones that are still registered which was problematic because the offer
> from this agent could still go to the same framework which could then
> launch new tasks. The agent would then receive tasks of the same
> framework and ignore them because it thinks the framework is shutting
> down. The framework is not shutting down of course, so from the master
> and the scheduler’s perspective the task is pending in STAGING forever
> until the next agent reregistration, which could happen much later.
> This commit fixes the problem by not shutting down the non-partition
> aware frameworks on such an agent.
> 
> 
> Diffs
> -----
> 
>   src/master/http.cpp 959091c8ec03b6ac7bcb5d21b04d2f7d5aff7d54 
>   src/master/master.hpp b802fd153a10f6012cea381f153c28cc78cae995 
>   src/master/master.cpp 7f38a5e21884546d4b4c866ca5918db779af8f99 
>   src/tests/partition_tests.cpp 62a84f797201ccd18b71490949e3130d2b9c3668 
> 
> 
> Diff: https://reviews.apache.org/r/61473/diff/1/
> 
> 
> Testing
> -------
> 
> make check
> 
> 
> Thanks,
> 
> Megha Sharma
> 
>


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message