Great to know the summary. I will go through the JIRAs mentioned for more 
details.

Given the direction, I can spare some time to test affinity feature next week:-)

Thanks & Regards
Rohith Sharma K S

-----Original Message-----
From: Steve Loughran [mailto:[email protected]] 
Sent: 19 November 2015 19:14
To: [email protected]
Subject: Re: proposed: google+ hangout


> On 19 Nov 2015, at 05:39, Rohith Sharma K S <[email protected]> wrote:
> 
> 
> Hi floks,
> 
> I could not able to attend sync meeting:-( Apologies for this. 
> 
> @Attendees, Would anyone summarize ? And If any recorded audio/video 
> of meeting is there, I would love more to listen those discussions 
> directly:-)
> 
> Thanks & Regards
> Rohith Sharma K S
> 

I gave a walkthrough of some of the code; specifically the anti-affine 
placement, and showed the changes made to AppState and RoleHistory.

you can see the status of that work
https://issues.apache.org/jira/browse/SLIDER-82

I've done the mock & minicluster tests and am doing the functional tests 
(SLIDER-970).,

I couldn't demo the stuff working on a cluster as SLIDER-993 is stopping me 
talk to my kerberosed VM cluster; I did just demo that in the minicluster you 
dont' get as many workers as you want, as there's only one host to deploy on. 
The AA placement code will leave the number of containers below desired rather 
than compromise the anti-affinity requirements. I've also got the web UI 
showing this and adding some explanations, Look at 
https://issues.apache.org/jira/browse/SLIDER-979 if you want screen shots.

Thomas Weise came in with some of his experience, especially fair scheduler 
placement problems. YARN-1412, relating to when that scheduler relaxes its 
placement. Thomas? Remember how I said that it didn't matter as we never set 
relaxLocality==true? It turns out that I do that when escalating placements 
—after the non-relaxed placement has failed. I don't think YARN-1412 will 
matter then, as that request is only issued if the relax=false request wasn't 
satisfied after its time limit.

Anyway, if i do get some functional test working (one designed to scale), then 
I'll merge the stuff in at the weekend. At that point the code will be ready to 
play with —and I do need people to play with it.

That initial release will be : 
-check out and build yourself (no ASF packaged release &vote) -wont use the 
history of previous placements before starting the AA placement sequence. I've 
focused on the core AA code first

Reply via email to