> On June 23, 2017, 7:21 p.m., Adam B wrote:
> > src/slave/slave.cpp
> > Lines 1405-1407 (original), 1405-1408 (patched)
> > <https://reviews.apache.org/r/60405/diff/2/?file=1761692#file1761692line1405>
> >
> >     If this agent has refinements, and we send post format to an old 
> > master, will the old master safely reject the registration, crash and burn, 
> > or something in between?
> 
> Neil Conway wrote:
>     The master will basically consider the resources to be unreserved; 
> because the master and agent will have inconsistent views of the resource 
> state at the agent, this will cause problems.
>     
>     Since you need a new master to create reservation refinements in the 
> first place, you can only arrive in this situation by:
>     
>     Upgrading master
>     Upgrading agent
>     Creating res refinement
>     Downgrading master
>     
>     Which arguably falls under the "don't downgrade if you are using new 
> featues" bucket. But yes, this is certainly unfortunate. Hard to prevent 
> without introducing something similar to master capabilities, which we 
> definitely need (MESOS-5675). I'll drop this issue for now, since AFAIK 
> there's not much we can do to improve this in the short term.

Sounds good. Thanks for the explanation. At least we don't crash..


> On June 23, 2017, 7:21 p.m., Adam B wrote:
> > src/slave/slave.cpp
> > Line 1408 (original), 1412-1414 (patched)
> > <https://reviews.apache.org/r/60405/diff/2/?file=1761692#file1761692line1412>
> >
> >     We could at least log an INFO/WARN if we aren't able to downgrade, and 
> > still send it anyway.
> 
> Neil Conway wrote:
>     Hmm, not sure a warning/log is warranted here. In the common case 
> (refined reservation, >= 1.4 master, >= 1.4 agent), we will fail to downgrade 
> the resources, but that is fine and expected. Should we really be cluttering 
> the logs with this information?

Good point. Don't want to over-log in the common case. Would be nice if we had 
master capabilities so we only had to log failure-to-downgrade when we're 
(re)registering with an old master.


- Adam


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/60405/#review178834
-----------------------------------------------------------


On June 24, 2017, 10:30 a.m., Neil Conway wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/60405/
> -----------------------------------------------------------
> 
> (Updated June 24, 2017, 10:30 a.m.)
> 
> 
> Review request for mesos, Adam B and Michael Park.
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Documented resource format in agent <-> master protocol.
> 
> 
> Diffs
> -----
> 
>   src/messages/messages.proto 2c086263fdcee4d54a76a61379c2d4dba5271d23 
>   src/slave/slave.cpp f808458849bb9667a91abe18868751d377d36e0c 
> 
> 
> Diff: https://reviews.apache.org/r/60405/diff/3/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Neil Conway
> 
>

Reply via email to