A truly HA cluster is often not required depending on what you use it for. Can 
you share what your application is?

The EventServer in my experience (I wrote the pages references below) has never 
crashed because of input. I think the only crash modes I’ve seen involved disk 
full on some service from log accumulation, or too much data accumulated. The 
EventServer is designed to ignore bad input, not exception out.

If you really do need HA, all services (HBase, HDFS, Elasticsearch, PIO) are 
designed with their own HA (multi-master, redundant data) configs. You would 
put the PIO eventserver and predictionserver on multiple machines behind a load 
balancer for instance.

If you see stoppage due to disk full, you might want to plug those issues 
because HA won’t help there. Monitoring core/CPU, memory, and disk usage is IMO 
more important than full HA unless you do indeed have a business critical use 
for PIO.


On Aug 7, 2017, at 1:39 AM, ROSSI Anthony <anthony.ro...@sword-group.com> wrote:

Hi, 

I think you could use one of the ActionML install docs for this purpose.
For example, you could use this one, and modify it to suit your purpose : 
http://actionml.com/docs/small_ha_cluster 
<http://actionml.com/docs/small_ha_cluster>
This one could also be pertinent : 
http://actionml.com/docs/single_driver_machine 
<http://actionml.com/docs/single_driver_machine>

Regards, 
Anthony ROSSI
De : 李文良 <william...@is-land.com.tw>
Envoyé : lundi 7 août 2017 10:30:08
À : user@predictionio.incubator.apache.org
Objet : How to config a high availability eventserver for PredictionIO ?
 
Hi, 

EventServer plays a critical role in PredictionIO platform and any abnormal 
events on it might cause the system error or halt. We'd like to build up a high 
availability (HA) architecture for stability and loading balance in our 
production environment. However, I can find out the wanted document from PIO 
web site. Could you please share your related experience or any suggestions for 
us how to config a HA eventserver ? Thanks.
-- 
Regards

William Lee

Reply via email to