Sorry for the delayed response Imesh. On Thu, Nov 6, 2014 at 7:29 PM, Imesh Gunaratne <im...@apache.org> wrote:
> Thanks for your response Nirmal, please see my thoughts below: > > On Thu, Nov 6, 2014 at 7:38 PM, Nirmal Fernando <nirmal070...@gmail.com> > wrote: > >> AFAIU if it is statistics, it's all about random data, samples and >> normalization. You don't use all values to do estimations. And this is an >> estimation for gradient per say! >> > > True, however the random data needs to be accurate as much as possible. > Yes, Imesh but this was the first step. We were crawling and let's stand gradually. :-) > >>> Well, statistics we are calculating is for a cluster as a whole not >> member wise. Since, we autoscale a cluster. >> > > Yes for autoscaling a cluster the aggregated statistics should be > calculated against the cluster. However I do not think that we can mix each > statistic accorss members when calculating differences. Different members > of a cluster might be running at different resource usage levels at a given > point of time. > In my understanding, members of a cluster are homogeneous from the allocated resource point of view. It's true that their usage levels are different at a given point of time (which is obvious). I doubt whether your suggestion is scalable in a system where we have 10s of clusters and 100s of members in each cluster (since we need to run an execution plan for each member ). This would be very costly IMO. Therefore aggregation might needed to be done at the member level first and > then on the cluster level. WDYT? > > > -- > Imesh Gunaratne > > Technical Lead, WSO2 > Committer & PMC Member, Apache Stratos > -- Best Regards, Nirmal Nirmal Fernando. PPMC Member & Committer of Apache Stratos, Senior Software Engineer, WSO2 Inc. Blog: http://nirmalfdo.blogspot.com/
