Hi David, Your feedback is encouraging!
Especially on a geographically distributed environment(e.g. cars as IoT clients, and its backends for analytics), such an algorithm based on a static threshold is not robust enough against various noises over data centers. I expect HTM is very robust against such noises. OK, then I will use HTM for this subject, staring with embedding HTM.java in Cassandra, and see how much it gets improved. Btw, I have some questions regarding the input format. 1. timestamp should be periodic? Should I give inputs periodically, at a fixed interval? Then, I need to take an average response time(or max if not available) for a period. e.g. timestamp(fixed interval), an average response time, node Or, can I simply feed as a health check response arrives? e.g. timestamp(variable), response time, node 2. one model for each node? To get anomaly score for each node, do I need to have one model for each node? Thanks, Takenori On Fri, Oct 23, 2015 at 9:03 PM, cogmission (David Ray) < [email protected]> wrote: > Hi Takenori, > > In my opinion (due to the constant time related pinging nature of the > health checks), is YES :-) HTM is **perfect** for this... Very interesting > application of NuPIC technology, by the way! > > Cheers, > David > > On Fri, Oct 23, 2015 at 2:13 AM, Takenori Sato <[email protected]> wrote: > >> Hello, >> >> In a distributed system, it is very important to know which node is more >> healthier than others to make a request. Or of course, when to determine >> one node should be treated as dead. >> >> For example, cassandra relies on phi accrual detector[1] to detect node >> down. A node does a gossip communication with 3 nodes every second, and >> exchanges information with each other. And its response time is used as an >> input for the failure detection. >> >> Also, a badness score is computed with such information, and which is >> used to choose a healthier node among replica nodes. >> >> But, I have seen many situations when it didn't work as expected, >> especially choosing a healthier node. >> >> On the other hand, I know any service provider makes some kind of health >> check request to detect if service is available or not. It may be just a >> simple ping, or HEAD request. >> >> Then, I just wondered if it is a good use case to use HTM for failure >> detection with such simple health check requests? >> >> For example, its input looks like this: >> >> time, node, avg response time(ms) >> 10:00:00, node1, 10 >> 10:00:00, node2, 9 >> ... >> 10:00:30, node1, 15 >> 10:00:30, node2, 10 >> ... >> >> >> [1] http://www.jaist.ac.jp/~defago/files/pdf/IS_RR_2004_010.pdf >> >> Thanks, >> Takenori >> > > > > -- > *With kind regards,* > > David Ray > Java Solutions Architect > > *Cortical.io <http://cortical.io/>* > Sponsor of: HTM.java <https://github.com/numenta/htm.java> > > [email protected] > http://cortical.io >
