[
https://issues.apache.org/jira/browse/MESOS-1453?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Benjamin Mahler updated MESOS-1453:
-----------------------------------
Description:
Currently, we only reply to a reconciliation request when the task state
differs. This makes it difficult for a framework to implement reconciliation:
if the framework does not receive an update for the task, it cannot determine
if this was because the state is valid, or the slave was transitioning, or the
master has not recovered yet.
As a consequence of this, frameworks looking to reconcile tasks need to
continually retry at small intervals to compensate for the lack of ability to
differentiate the cases listed above.
A solution to this problem is to send an update for all of the tasks, if
possible. This means we no longer compare task states to determine when to send
an update.
was:
Currently, we only reply to a reconciliation request when the task state
differs. This makes it difficult for a framework to implement reconciliation:
if the framework does not receive an update for the task, it cannot determine
if this was because the state is valid, or the slave was transitioning, or the
master has not recovered yet.
As a consequence of this, frameworks looking to reconcile tasks need to
continually retry at small intervals to compensate for the lack of ability to
differentiate the cases listed above.
A solution to this problem is update the reconciliation API to instead just
send the task ids to be reconciled, and the master will send an update for all
of the tasks, if possible. This means we no longer compare task states to
determine when to send an update.
> Update reconciliation semantics send statuses for each task.
> ------------------------------------------------------------
>
> Key: MESOS-1453
> URL: https://issues.apache.org/jira/browse/MESOS-1453
> Project: Mesos
> Issue Type: Improvement
> Affects Versions: 0.19.0
> Reporter: Benjamin Mahler
> Fix For: 0.20.0
>
>
> Currently, we only reply to a reconciliation request when the task state
> differs. This makes it difficult for a framework to implement reconciliation:
> if the framework does not receive an update for the task, it cannot determine
> if this was because the state is valid, or the slave was transitioning, or
> the master has not recovered yet.
> As a consequence of this, frameworks looking to reconcile tasks need to
> continually retry at small intervals to compensate for the lack of ability to
> differentiate the cases listed above.
> A solution to this problem is to send an update for all of the tasks, if
> possible. This means we no longer compare task states to determine when to
> send an update.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
