TubeMogul grew from few servers to over two thousands servers and handling over one trillion http requests a month, processed in less than 50ms each. To keep up with the fast growth, the SRE team had to implement an efficient Continuous Delivery infrastructure that allowed to do over 10,000 puppet deployment and 8,500 application deployment in 2014. In this presentation, we will cover the nuts and bolts of the TubeMogul operations engineering team and how they over come challenges.

1. How TubeMogul reached
10,000 Puppet deployment in
one year
May 26th, 2015
Nicolas Brousse | Sr. Director Of Operations Engineering | nicolas@tubemogul.com
Julien Fabre | Site Reliability Engineer | julien.fabre@tubemogul.com

2. Who are we?
TubeMogul
● Enterprise software company for digital branding
● Over 27 Billions Ads served in 2014
● Over 30 Billions Ad Auctions per day
● Bid processed in less than 50 ms
● Bid served in less than 80 ms (include network round trip)
● 5 PB of monthly video traffic served
● 1.3 EB of data stored

3. Who are we?
Operations Engineering
● Ensure the smooth day to day operation of the platform
infrastructure
● Provide a cost effective and cutting edge infrastructure
● Team composed of SREs, SEs and DBAs
● Managing over 2,500 servers (virtual and physical)

4. Our Infrastructure
Public Cloud On Premises
Multiple locations with a mix of Public Cloud and On Premises

5. ● Java (a lot!)
● MySQL
● Couchbase
● Vertica
● Kafka
● Storm
● Zookeeper, Exhibitor
● Hadoop, HBase, Hive
● Terracotta
● ElasticSearch, Kibana
● LogStash
● PHP, Python, Ruby, Go...
● Apache httpd
● Nagios
● Ganglia
Technology Hoarders
● Graphite
● Memcached
● Puppet
● HAproxy
● OpenStack
● Git and Gerrit
● Gor
● ActiveMQ
● OpenLDAP
● Redis
● Blackbox
● Jenkins, Sonar
● Tomcat
● Jetty (embedded)
● AWS DynamoDB, EC2, S3...

6. ● 2008 - 2010: Use SVN, Bash scripts and custom templates.
● 2010: Managing about 250 instances. Start looking at Puppet.
● 2011: Puppet 0.25 then 2.7 by EOY on 400 servers with 2 contributors.
● 2012: 800 servers managed by Puppet. 4 contributors.
● 2013: 1,000 servers managed by Puppet. 6 contributors.
● 2014: 1,500 servers managed by Puppet. Introduced Continuous Delivery
Workflow. 9 contributors. Start 3.7 migration.
● 2015: 2,000 servers managed by Puppet. 13 contributors.
Five Years Of Puppet!

7. ● 2000 nodes
● 225 unique nodes definition
● 1 puppetmaster
● 112 Puppet modules
Puppet Stats

8. ● Virtual and Physical Servers Configuration : Master mode
● Building AWS AMI with Packer : Master mode
● Local development environment with Vagrant : Master mode
● OpenStack deployment : Masterless mode
Where and how do we use Puppet ?

9. Code Review?

10. ● Gerrit, an industry standard : Eclipse, Google, Chromium, OpenStack,
WikiMedia, LibreOffice, Spotify, GlusterFS, etc...
● Fine Grained Permissions Rules
● Plugged to LDAP
● Code Review per commit
● Stream Events
● Use GitBlit
● Integrated with Jenkins and Jira
● Managing about 600 Git repositories
A Powerful Gerrit Integration

11. Gerrit in Action
verify -1 when no
ticket # or doesn’t
pass Jenkins code
validation

12. ● 1 job per module
● 1 job for the manifests and hiera data
● 1 job for the Puppet fileserver
● 1 job to deploy
Continuous Delivery with Jenkins
Global Jenkins stats for the past year
● ~10,000 Puppet deployment
● Over 8,500 Production App Deployment

13. Plugin : github.com/jenkinsci/job-dsl-plugin
● Automate the jobs creation
● Ensure a standard across all the jobs
● Versioned the configuration
● Apply changes to all your jobs without pain
● Test your configuration changes
Jenkins job DSL : code your Jenkins jobs

14. Team Awareness: HipChat Integration with Hubot

15. Infrastructure As Code
● Follow standard development lifecycle
● Repeatable and consistent server
provisioning
Continuous Delivery
● Iterate quickly
● Automated code review to improve code
quality
Reliability
● Improve Production Stability
● Enforce Better Security Practices
Puppet Continuous Delivery Workflow: The Vision

16. The Workflow

17. The Workflow : Puppet code logic
Puppet environments
● Dedicated node manifests (*.pp)
● Modules deployed by branch with Git submodules
All the data in Hiera
● Try to avoid params.pp class
● Store everything : modules parameters, classes, keys, passwords, ...

18. Puppet Code Hierarchy
/etc/puppet
├── puppet.conf, hiera.yaml, *.conf
├── hiera
└── environments
├── dev
│ ├── manifests
│ │ ├── nodes/*.pp
│ │ └── site.pp
│ └── modules
│ ├── activemq
│ ...
│ └── zookeeper
└── production
├── manifests
│ ├── nodes/*.pp
│ └── site.pp
└── modules
├── activemq
…
└── zookeeper
Git submodules, branch dev
Git submodules, branch production

19. Hiera Configuration
$ cat /etc/puppet/hiera.yaml
---
:backends:
- eyaml
- yaml
:yaml:
:datadir: /etc/puppet/hiera
:eyaml:
:pkcs7_private_key: /var/lib/puppet/hiera_keys/private_key.pkcs7.pem
:pkcs7_public_key: /var/lib/puppet/hiera_keys/public_key.pkcs7.pem
:hierarchy:
- fqdn/%{::fqdn}
- "%{::zone}/%{::vpc}/%{::hostgroup}"
- "%{::zone}/%{::vpc}/all"
- "%{::zone}/%{::hostgroup}"
- "%{::zone}/all"
- hostname/%{::hostname}
- hostgroup/%{::hostgroup}
- environment/%{::environment}
- common
:merge_behavior: deeper

20. Hiera eyaml : github.com/TomPoulton/hiera-eyaml
● Hiera backend
● Easy to use
● Powerful CLI : eyaml edit /etc/puppet/hiera/secrets.yaml
Encrypt Your Secrets
$ cat secret.yaml
---
ec2::access_key_id: ENC[PKCS7,MIIBiQYJKoZIhvcNAQcDoIIBejCCAXYCAQAxggEhMII
IBHQIBADAFMAACAQEwDQYJKoZIhvcNAQEBBQAEggEAVIa28OwyaqI5N1TDCvVkBZz3YG+s+Hfzr0lqgcvRCIuJ
Gpq28sQmmuBaQjWY38i86ZSFu0gM6saOHfG64OzVlurO7k/l0CKeL0JfXNaVM4TUqMaN9dSkL5e2vsmpLKrMAS
awmarqbLYwllTrTe32H4NWxU1e+qWLeUMr9ciBnA3W1Azm4RIo+3bsvgvMfdks....=]

21. Encrypt Files
Blackbox : github.com/StackExchange/blackbox
● Use GPG to encrypt secret files
● Easy to add/delete team members
● No need to change your Puppet code !
# modules/${modules_name}/files/credentials.yaml.gpg
file { ‘/etc/app/credentials.yaml’:
ensure => ‘file’,
owner => ‘root’,
group => ‘root’,
mode => ‘0644’,
source => ‘puppet:///modules/${module_name}/credentials.yaml’
}

22. The Workflow

23. The Workflow : bottlenecks
● Only Ops team members can commit (SRE, SE)
● Review and validation is done only by a SRE
● Jenkins will verify the code but will not validate the commit
● Static Puppet environments
● Rely a lot on server hostnames

24. Flexibility : R10K github.com/adrienthebo/r10k !
● Dynamic environments
● No Git submodules anymore ! : - )
● Easy to reproduce any environment
● Can use private and forge Puppet modules
● Can use branches and tags
● Based on Puppetfile
Puppet Workflow Reloaded!

25. R10K
$ cat Puppetfile
forge "https://forgeapi.puppetlabs.com"
# Forge modules
mod 'pdxcat/collectd'
mod 'puppetlabs/rabbitmq'
mod 'arioch/redis'
mod 'maestrodev/wget'
mod 'puppetlabs/apt'
mod 'puppetlabs/stdlib'
# Tubemogul modules
mod "hosts",
:git => 'ssh://<gerrit_host>/puppet/modules/hosts',
:branch => 'dev'
mod "timezone",
:git => 'ssh://<gerrit_host>/puppet/modules/timezone',
:branch => 'dev'
...

26. Puppet Workflow Reloaded!
Better code organization : Roles and Profiles
● Represent the business logic : Roles
o Highest abstraction layer
o Use Profiles for implementation
● Implement the applications : Profiles
o Remove potential code duplication
o Use modules and other Puppet resources

27. Roles/Profiles Pattern
class role::logs {
include profile::base
include profile::logstash::server
include profile::elasticsearch
}
class profile::logstash {
$version = hiera('profile::logstash::server::version', '1.4.2')
$es_host = hiera('profile::logstash::server::es_host', 'es01')
$redis_host = hiera('profile::logstash::server::redis_host', 'redis01')
class { 'logstash':
package_url => "https://download.elasticsearch.org/logstash/.../logstash_${version}.deb",
java_install => true,
}
logstash::configfile { 'input_redis':
content => template('logstash/configfile/logstash.input_redis.conf.erb'),
order => 10,
}
logstash::configfile { 'output_es':
content => template('logstash/configfile/logstash.output_es.conf.erb'),
order => 30,
}
}

28. Do not rely on hostname : nodeless approach
● Facts to guide Puppet
● No node myawesomeserver { } anymore
● Enforce a cluster vision
● site.pp gives the configuration logic
Puppet Workflow Reloaded!
# /etc/puppet/manifests/site.pp
node default {
if $::ec2_tag_tm_role {
notify { "Using role : ${ec2_tag_tm_role}": }
include "role::${::ec2_tag_tm_role}"
} else {
fail(‘No role found. Nothing to configure.’)
}
}

29. ● Specify tags during the provisioning
● Retrieve tags with AWS Ruby SDK and create facts
● New hierarchy
AWS EC2 tags
$ facter -p | grep ec2_tag
ec2_tag_cluster => rtb-bidder
ec2_tag_nagios_host => mgmt01
ec2_tag_name => bidder
ec2_tag_pupenv => production
ec2_tag_tm_role => rtb::bidder
:hierarchy:
- "%{::zone}/%{::ec2_tag_vpc}/%{::ec2_tag_cluster}"
- "%{::zone}/%{::ec2_tag_vpc}/all"
- "%{::zone}/all"
- vpc/%{::ec2_tag_vpc}/%{::ec2_tag_cluster}
- vpc/%{::ec2_tag_vpc}/all
- environment/%{::environment}
- common

30. New merging and reviewing rules
● Everyone can commit a Puppet code
● Allow everyone to review a Puppet change (+1)
● Allow SE and SRE to validate a Puppet change (+2)
● Auto validation/merging in dev if at least 80% of test (+2)

31. Next improvements
● Acceptance testing with Beaker and Docker
● Full test provisioning with ServerSpec
● PuppetDB to improve the reporting
● Dedicated Puppet Masters

32. Nicolas Brousse
Julien Fabre
@orieg
@julien_fabre