Thank Matthieu for reading through this! On 02/10/2016 10:04 AM, Matthieu Huin wrote: > Comments inline > > ----- Original Message ----- >> From: "Tristan Cacqueray" <[email protected]> >> To: [email protected] >> Sent: Wednesday, February 10, 2016 2:23:17 AM >> Subject: [Softwarefactory-dev] ELK Stack discussion >> >> Hello folks, >> >> so I've started the process to integrate elasticsearch, logstash and >> kibana (ELK) in software-factory[0]. Mostly following this cool >> documentation[1]. >> >> Is there a reason why we sticked to java-1.7.0 ? Because it seems like >> gerrit and jenkins are working fine with 1.8.0 so that doesn't seems a >> blocker. > > None that I am aware of, is java pulled from a rpm ? >
Yes, CentOS provides both version of java, java-1.7.0-openjdk-*.rpm and java-1.8.0-openjdk-*.rpm >> Beside that, the recommended webserver is nginx... Would it be a problem >> to run both webserver so that the elk stack is isolated ? >> > > having to maintain 2 different kinds of webservers sounds a bit overkill > honestly ... > Do the performances drop significantly under Apache ? Although I guess it > could make sense in a distributed arch, see below. > I think it's less of a performance issue and more of a ease of management actually. Webserver configuration is very trivial compared to each virtualhost settings. And if nginx is better supported for elk, then it will be overkill for us to port it to apache actually. Well after testing, centos seems able to run both webserver side by side (providing they don't bind to the same port). >> >> Now the interesting part... >> >> Logs collection from jenkins seems doable using the logstash plugin, >> however openstack-infra is using a custom gearman client to plug into >> zuul server and do the log processing in python after job completion[2]. >> The plugin looks straightforward and already supported by jjb, but the >> gearman approach looks much better and may worth the time investment. >> >> Kibana looks pretty neat, should we look into that for ci monitoring or >> is grafana still trendy ? > > What does Kibana do that grafana doesn't ? According to https://discuss.elastic.co/t/elk-vs-grafana-influxdb/1686 It seems like Grafana could be used for time series metric and Kibana for logs. > >> And finally, how should this be deployed... >> * all services on managesf ? >> * elasticsearch side by side with mysql ? >> * completely distributed ? >> > > I am beginning to favor the following arch: > - SF core services on the main node: cauth, managesf, gerrit, jenkins, zuul, > mysql (for cauth/managesf/gerrit/jenkins) > - one node per optional service: redmine / ELK / extra jenkins master / ... > with the ELK node also hosting its own DB backend for performance. > After thinking some more, I guess SF should always be able to be deployed on a single node (to really benefit from a beefy system). And about distributed architecture, that should be left at the user discretion, the provided heat stack may supports an "inventory" input where an user can define how many nodes are associated with which roles. And we could validate that by testing a fully distributed setup with one node per service to demonstrate it's working fine. I guess we could have that many nodes: * mysql * gerrit * managesf/gateway * redmine * etherpad * lodgeit * jenkins0[1-9] * zuul * nodepool * smtp * statsd * gnocchi * grafana * elasticsearch * kibana * logstash * logstash-workers And service that scales horizontally like jenkins or logstash-workers could be supported using ansible inventory group such as [jenkins] node1 node2 node3 ... -Tristan
signature.asc
Description: OpenPGP digital signature
_______________________________________________ Softwarefactory-dev mailing list [email protected] https://www.redhat.com/mailman/listinfo/softwarefactory-dev
