[
https://issues.apache.org/jira/browse/MESOS-7735?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16074575#comment-16074575
]
Michael Park commented on MESOS-7735:
-
{noformat}
commit e097f21124d16c80544af7aabc83bab5fb712453
Author: Michael Park
Date: Thu Jun 29 23:49:10 2017 -0700
Performed validation/normalization of `Resource`s before authorization.
Review: https://reviews.apache.org/r/60564
{noformat}
{noformat}
commit 63c8b1d63be363e96d7c2672d9d65b9d67ca48ed
Author: Michael Park
Date: Thu Jun 29 23:30:16 2017 -0700
Updated `validateAndNormalizeResources` to operate on `Operation`s.
Review: https://reviews.apache.org/r/60563
{noformat}
{noformat}
commit 710b72179938cac2100c90277ce7ced5c8ca3401
Author: Michael Park
Date: Thu Jun 29 20:53:59 2017 -0700
Updated `accept` to perform operation adjustment in one place.
It used to be that the minor adjustments that were made to operations
were done in various places across `accept` and `_accept`.
The "executor-injection" for LAUNCH_GROUP was at the beginning of
`accept`, "allocation-info-injection" for MULTI_ROLE was after offer
validation, and "health-check-injection" for LAUNCH was in `_accept`.
The `Master::accept` function is now broken down into distinct
"metrics accounting", "offer validation", "operation-adjustments", and
"authorization" stages.
Review: https://reviews.apache.org/r/60562
{noformat}
> The master crashes when state endpoint is hit during a task authorization.
> --
>
> Key: MESOS-7735
> URL: https://issues.apache.org/jira/browse/MESOS-7735
> Project: Mesos
> Issue Type: Bug
> Components: master
>Reporter: Michael Park
>Assignee: Michael Park
>Priority: Blocker
>
> With reservation refinement, the construction of {{Resources}} requires
> {{Resource}} objects to have been validated and converted to the
> "post-reservation-refinement" format. Generally speaking, validation and
> conversion are the first steps we take with given {{Resource}} objects prior
> to proceeding. In the master currently, we perform authorization first with
> not-yet-validated, not-yet-converted {{Resource}} objects. During the
> authorization phase, we add tasks with not-yet-validated, not-yet-converted
> resources into {{framework->pendingTasks}} as well as
> {{slave->pendingTasks}}.
> (https://github.com/apache/mesos/blob/master/src/master/master.cpp#L3974-L3999).
> If one hits the state endpoint on the master during this time, we get to
> https://github.com/apache/mesos/blob/master/src/master/http.cpp#L278 which
> tries to construct a {{Resources}} with {{taskInfo.resources()}} which is
> not-yet-validated nor converted.
> I think the correct fix here is to perform validation / conversion prior to
> authorization. The authorization code currently is written to carefully
> inspect fields in both "pre-reservation-refinement" and
> "post-reservation-refinement" formats. By performing validation / conversion
> first, the authorization code would be simplified, and we're also much less
> likely to make mistakes such as this.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
