On Wed, Dec 7, 2016 at 3:26 PM, Vinod Kone <[email protected]> wrote:

> We are trying to move away from un-versioned REST endpoints like "/reload"
> in favor of versioned endpoint(s) that take calls. We don't have a good
> story yet for endpoints that are not hosted by Master and Agent actors. So,
> at first blush, I'm wary of this solution.
>
> Ok, this definitely throws a spanner to the idea of having a reload
endpoint.

> - Why not scan the config directory every time? Are you trying to avoid the
> > speed hit from disk reads?
> >
>
> Why not do this? If a network config doesn't exist in memory, refresh
> configs by reading from disk and check again; basically acting like a
> cache. Seems straight forward enough?
>
This does seem feasible. We were trying to avoid unnecessary reads . For
example a framework continuously launches containers on a non-existent
network (due to operator error or malicious intent). This would result in
quite a few unnecessary reads. However, if the reload endpoint is not a
possibility than this would be the route we have to take.



-- 
Avinash Sridharan, Mesosphere
+1 (323) 702 5245

Reply via email to