jumping/oculus

The metric correlation component of Etsy's Kale system

Oculus

Oculus is the anomaly correlation component of Etsy's Kale system.

It lets you search for metrics, using your choice of comparison algorithms:

and shows you other metrics which are similar:

You can even save interesting metrics into a collection, complete with your own notes:

And then if Oculus finds matches in your saved collections, it'll show you this alongside your other search results:

Installation Overview

Oculus consists of the following components:

• Sinatra Web App - the Oculus front end, used to search for metric correlations
• Skyline Import Script and Cronjob - Used to import data from Skyline into Elasticsearch
• ElasticSearch Cluster - The search backend for Oculus
• Worker Box(es) - Used to process metrics from Skyline and import them into Elasticsearch
• Working Skyline install - Oculus is fed using data from Skyline's data store. Without Skyline, it won't have any data to look at.

Where recommended server specs have been mentioned in this document, they're based on Etsy's usage of Oculus which is based around 250k metrics. Adjust as necessary for your own metric volume.

It's recommended that you work through the section in this README in the following order:

• ElasticSearch 1
• Workers
• Oculus Config File
• Skyline Import Script and Cronjob
• WebApp

Following the instructions in this order should result in a working Oculus setup with all the moving parts running and functioning correctly. Please note that these instructions are geared towards installing Oculus components on separate boxes. If you're installing everything on one box, then you can ignore replicated steps such as cloning the Oculus code - this only needs to be done once.

Also provided, but not necessary to get Oculus up and running are some utility scripts, in the section headed

• Misc Utility Scripts

##ElasticSearch

For Oculus to function properly, you will ideally need at least two ElasticSearch servers in separate clusters, which Oculus will rotate through. Oculus requires the addition of custom scoring plugins, but otherwise uses ElasticSearch out of the box.

###Recommended Server Spec

• At least 8GB RAM
• Quad Core Xeon 5620 CPU or comparable
• 1GB disk space

###Installation and Plugin Build (Applies to all cluster nodes)

• Install the Java JDK (on CentOS, this is yum install jdk)
• Download and extract elasticsearch from http://www.elasticsearch.org/download/ - here we'll assume /opt/elasticsearch
  • Oculus has been tested with version 0.20.5 - it will currently not build on version 0.90 and above
• Clone the Oculus repository to somewhere on your server - here we'll assume /opt/oculus
• Run the command mkdir elasticsearch-oculus-plugin
• Copy /opt/oculus/resources/elasticsearch-oculus-plugin to /opt/elasticsearch/elasticsearch-oculus-plugin
• cd to /opt/elasticsearch/elasticsearch-oculus-plugin
• Run the following command:

  rake build
  

• If successful, this will create a file called OculusPlugins.jar
• copy OculusPlugins.jar to /opt/elasticsearch/lib/OculusPlugins.jar

###Configuration

As noted above, the Elasticsearch servers which Oculus uses cannot be in the same cluster - Oculus needs at least two seperate servers (and clusters) to rotate between. To that end, the ElasticSearch configuration file needs to have the Cluster name and Node name manually specified. For simplicity, these instructions assume that you're going to name each cluster and node after the hostname of the server, but any name can be used.

The scoring plugins which Oculus uses for searching must also be added to the configuration file.

Change the following lines in /opt/elasticsearch/config/elasticsearch.yml, changing the cluster and nodename to match the servers you're using:

cluster.name: oculussearch01.mydomain.com
node.name: oculussearch01.mydomain.com

Add the following lines to /opt/elasticsearch/config/elasticsearch.yml, changing the cluster and nodename to match the servers you're using:

script.native:
  oculus_euclidian.type: com.etsy.oculus.tsscorers.EuclidianScriptFactory
  oculus_dtw.type: com.etsy.oculus.tsscorers.DTWScriptFactory

###Start Up ElasticSearch Once you've installed ElasticSearch, built the Oculus scoring plugins and edited the configuration file, you can start it up by running the following command from /opt/elasticsearch

bin/elasticsearch

You can verify that Elasticsearch is up and running by visiting http://locahost:9200 on the box you installed ElasticSearch on. If all is well, you should see JSON similar to the below:

{
  "ok" : true,
  "status" : 200,
  "name" : "oculussearch01",
  "version" : {
    "number" : "0.20.5",
    "snapshot_build" : false
  },
  "tagline" : "You Know, for Search"
}

##Workers

Oculus' import and backend functions are handled by a cluster of Worker boxes running Resque (https://github.com/resque/resque). You'll need a Worker Master, running Redis (the central data store used by Resque) and some resque workers, and optionally some more Worker Slave boxes running extra Resque workers.

###Recommended Server Spec

• At least 12GB RAM
• Quad Core Xeon 5620 CPU or comparable
• 1GB disk space

###Worker Master Box

The first worker box you'll need to get up and running is the Master box running Redis.

Here's how to get it up and running:

• Install redis (on CentOS, yum install redis)
• Start up Redis (on CentOS, service redis start)
• Verify you can connect to Redis by running redis-cli
  • If everything is working, you should see this prompt: redis 127.0.0.1:6379>
• Type exit to return back to the Shell.
• Clone the Oculus repository to somewhere on your server - here we'll assume /opt/oculus
• cd into the directory where you cloned Oculus
• run the command bundle install
• Edit the file Rakefile, changing the URL on line 66 to match the URL to your redis instance
• Run the command mkdir /var/run/oculus and make sure the user you will run your workers as can read and write to it.
• Run the command mkdir /var/log/oculus and make sure the user you will run your workers as can read and write to it.
• To start the workers, from the directory where you installed Oculus run rake resque:start_workers
• Optional Steps
  • You can start up a Resque web console to monitor what's going on by running resque-web from the binary directory of the resque Gem (for example /usr/lib64/ruby/gems/1.9.1/gems/resque-1.23.0/bin/resque-web)
  • If you want to edit the number of Resque workers run on the master box, edit lines 21 and 22 of the Rakefile (changing the number '22' to whatever you like, then run rake resque:restart_workers

###Worker Slave Box(es) - Optional

Once you've got your Worker Master set up and running Redis, you can also optionally set up any number of Worker Slave boxes (which just run Resque Workers) you want - we run 3 at Etsy.

• Clone the Oculus repository to somewhere on your server - here we'll assume /opt/oculus
• cd into the directory where you cloned Oculus
• run the command bundle install
• Edit the file Rakefile, changing the URL on line 66 to match the URL to your redis instance on the Worker Master
• Run the command mkdir /var/run/oculus and make sure the user you will run your workers as can read and write to it.
• Run the command mkdir /var/log/oculus and make sure the user you will run your workers as can read and write to it.
• To start the workers, from the directory where you installed Oculus, run rake resque:start_workers
• Optional Steps
  • If you want to edit the number of Resque workers run on this box, edit lines 21 and 22 of the Rakefile (changing the number '22' to whatever you like, then run rake resque:restart_workers

##Oculus Config File

Now that you've got your Elasticsearch servers all prepped and your Workers happily waiting for work, before we can start chucking data Oculus the next step is to set up the Oculus config file - config/config.yml. Here's a sample config file:

results_explain: 0
elasticsearch:
  servers:
     - "http://oculussearch01.mycompany.com:9200"
     - "http://oculussearch02.mycompany.com:9200"
  index: "metrics"
  timeout: 30
  phrase_slop: 20
  scorers:
    dtw:
      radius: 5
      scale_points: 25
    euclidian:
      scale_points: 25
skyline:
  host: "skyline.mycompany.com"
  port: 6379
  listener_port: 2015
redis:
  host: "oculusredis.mycompany.com"
  port: 6379

At this stage, all you need to do is the following:

• Add your elasticsearch servers (one per line) to the elasticsearch->servers section
• Add the hostname and port of your Worker Master's redis instance to the redis section
• Add your Skyline host, port and listener port to the skyline section.

You can ignore the other settings under elasticsearch for now - Have a look at the "Help" page in Oculus once you've gotten it up and running if you want to know more about these. For now, the defaults will suffice!

You should make sure that your config file is pushed out to all of your Oculus servers.

##Skyline Import Script and Cronjob

Now that you've got your search servers all ready to populate, you'll need to set up the script that imports data from Skyline into your search indexes. This can run on any of the servers you're using to run Oculus.

• Double check that you've configured your Oculus config file correctly as specified in the above section.
• cd into the scripts directory under the directory where you cloned Oculus.
• Run ./import.rb
• If all is working correctly, you should see output similar to this:

 Active ES Server: http://oculussearch01.mycompany.com:9200
 Next ES Server: http://oculussearch02.mycompany.com:9200
 Recreating indexes
 Creating redis jobs...
 Getting unique metric names
 Found 250578 metric names
 187 workers working
 311 process_redis_metrics jobs left to run
 187 workers working
 309 process_redis_metrics jobs left to run
 187 workers working
 148 process_redis_metrics jobs left to run
 187 workers working
 113 process_redis_metrics jobs left to run
 187 workers working
 110 process_redis_metrics jobs left to run
 153 workers working
 0 process_redis_metrics jobs left to run
 Setting active search server to http://oculussearch02.mycompany.com:9200
 Oculus import finished in 36.634672763 seconds

• Next, create a directory for the cronjob to log into (here we've used /var/log/oculus)
• Create a cronjob as follows, changing the full path to where you cloned the Oculus code, and the cron frequency as applicable):

*/2 * * * * /opt/oculus/scripts/import.rb > /var/log/oculus/import.log 2>&1

• Your Oculus search index will now be updated from Skyline every 2 minutes.

##Web App Install

Now that we've got all of the moving parts set up and started updating our search indexes with metric data, the final step is to get the Oculus front end web app set up.

• Clone the Oculus repository to somewhere on your server - here we'll assume /opt/oculus
• cd into the directory where you cloned Oculus
• run the command bundle install
• Make sure you've got your correctly configured configuration file in the config directory under your Oculus checkout
• Run the command thin start
• If everything's working correctly, you should see output similar to the following:

>> Using rack adapter
>> Thin web server (v1.5.0 codename Knife)
>> Maximum connections set to 1024
>> Listening on 0.0.0.0:3000, CTRL+C to stop

• The final step is to log onto the Oculus admin console (http://yourserver:3000/admin, username admin and password admin), then click "Reinitialize Collections." Click OK to confirm.
• You're all done! Oculus should now be running at http://yourserver:3000 - you can now start searching for metrics!!