[
https://issues.apache.org/jira/browse/MESOS-5359?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15298436#comment-15298436
]
Anand Mazumdar commented on MESOS-5359:
---------------------------------------
Ideally, yes. The delay should be configurable by a flag.
Have a look at how we are already doing so for the old driver based interface
{{src/sched/flags.hpp}}. The only difference here is that we just need a single
maximum connection backoff variable e.g.,
{{MESOS_CONNECTION_BACKOFF_MAX=~500ms}}. The scheduler library can then do a
linear backoff after picking a random delay between 0 and maxBackoff for
initiating the connection with the master.
> The scheduler library should have a delay before initiating a connection with
> master.
> -------------------------------------------------------------------------------------
>
> Key: MESOS-5359
> URL: https://issues.apache.org/jira/browse/MESOS-5359
> Project: Mesos
> Issue Type: Bug
> Affects Versions: 0.29.0
> Reporter: Anand Mazumdar
> Assignee: José Guilherme Vanz
> Labels: mesosphere
>
> Currently, the scheduler library {{src/scheduler/scheduler.cpp}} does have an
> artificially induced delay when trying to initially establish a connection
> with the master. In the event of a master failover or ZK disconnect, a large
> number of frameworks can get disconnected and then thereby overwhelm the
> master with TCP SYN requests.
> On a large cluster with many agents, the master is already overwhelmed with
> handling connection requests from the agents. This compounds the issue
> further on the master.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
