[
https://issues.apache.org/jira/browse/MESOS-5181?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anand Mazumdar updated MESOS-5181:
----------------------------------
Description:
When a scheduler registers, the master will create a link from master to
scheduler. If this link breaks, the master will consider the scheduler
{{inactive}} and mark it as {{disconnected}}.
This causes a couple problems:
1) Master does not send offers to {{inactive}} schedulers. But these
schedulers might consider themselves "registered" in a one-way network
partition scenario.
2) Any calls from the {{inactive}} scheduler is still accepted, which leaves
the scheduler in a starved, but semi-functional state.
See the related issue for more context: MESOS-5180
There should be an additional guard for registered, but {{inactive}} schedulers
here:
https://github.com/apache/mesos/blob/94f4f4ebb7d491ec6da1473b619600332981dd8e/src/master/master.cpp#L1977
The HTTP API already does this:
https://github.com/apache/mesos/blob/94f4f4ebb7d491ec6da1473b619600332981dd8e/src/master/http.cpp#L459
Since the scheduler driver cannot return a 403, it may be necessary to return a
{{Event::ERROR}} and force the scheduler to abort.
was:
When a scheduler registers, the master will create a link from master to
scheduler. If this link breaks, the master will consider the scheduler
{{inactive}} and {{disconnected}}.
This causes a couple problems:
1) Master does not send offers to {{inactive}} schedulers. But these
schedulers are still considered "registered".
2) Any calls from the {{inactive}} scheduler is still accepted, which leaves
the scheduler in a starved, but semi-functional state.
See the related issue for more context: MESOS-5180
There should be an additional guard for registered, but {{inactive}} schedulers
here:
https://github.com/apache/mesos/blob/94f4f4ebb7d491ec6da1473b619600332981dd8e/src/master/master.cpp#L1977
The HTTP API already does this:
https://github.com/apache/mesos/blob/94f4f4ebb7d491ec6da1473b619600332981dd8e/src/master/http.cpp#L459
Since the scheduler driver cannot return a 403, it may be necessary to return a
{{Event::ERROR}} and force the scheduler to abort.
> Master should reject calls from the scheduler driver if the scheduler is not
> connected.
> ---------------------------------------------------------------------------------------
>
> Key: MESOS-5181
> URL: https://issues.apache.org/jira/browse/MESOS-5181
> Project: Mesos
> Issue Type: Bug
> Components: scheduler driver
> Affects Versions: 0.24.0
> Reporter: Joseph Wu
> Assignee: Anand Mazumdar
> Labels: mesosphere
>
> When a scheduler registers, the master will create a link from master to
> scheduler. If this link breaks, the master will consider the scheduler
> {{inactive}} and mark it as {{disconnected}}.
> This causes a couple problems:
> 1) Master does not send offers to {{inactive}} schedulers. But these
> schedulers might consider themselves "registered" in a one-way network
> partition scenario.
> 2) Any calls from the {{inactive}} scheduler is still accepted, which leaves
> the scheduler in a starved, but semi-functional state.
> See the related issue for more context: MESOS-5180
> There should be an additional guard for registered, but {{inactive}}
> schedulers here:
> https://github.com/apache/mesos/blob/94f4f4ebb7d491ec6da1473b619600332981dd8e/src/master/master.cpp#L1977
> The HTTP API already does this:
> https://github.com/apache/mesos/blob/94f4f4ebb7d491ec6da1473b619600332981dd8e/src/master/http.cpp#L459
> Since the scheduler driver cannot return a 403, it may be necessary to return
> a {{Event::ERROR}} and force the scheduler to abort.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
