"Puppet at Scale – Case Study of PayPal's Learnings" by Stan Hsu, Senior Dev Manager, PayPal.
Presentation Overview: Large scale and app level management pose challenges to any implementation of puppet. Come and learn some of the challenges PayPal Deployment Systems team faced and the how these were overcome.
Speaker Bio: Stan Hsu is the Senior Dev Manager for PayPal's deployment systems team. His team is currently responsible to build out a new deployment system based on puppet. In his tenure at eBay/PayPal, he's had the unique experience of having had access to all data centers in both eBay and PayPal to help build out of new deployment systems for production and QA environments. His interests include application at scale, scalability, performance tuning, and usability. In his previous roles he has managed teams at Tibco, Crossworlds, and HP.
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
Puppet at Scale – Case Study of PayPal's Learnings - PuppetConf 2013
1. PUPPET AT SCALE:
CASE STUDY OF LEARNINGS AT PAYPAL
Stan Hsu, Harendra Narayan, Chris Huang
August 23, 2013
2. PAYPAL SCALE
2 Confidential and Proprietary
• PayPal is part of eBay Inc
• 132 million active registered
accounts
• 25 currencies in 193 markets
• Net Total Payment Volume: $43 Billion in Q2
• 7.6 million payments per day
• $5,277 in Total Payment Volume every second
3. THE CHALLENGE
3 Confidential and Proprietary
• 100GB+ and counting…
• 4000+ packages
• 20-50 new packages introduced every release
• complex dependency graph across domains and services
• Build a new system for deploying
application and system software
• Massive scale in production across
multiple data centers
• Thousands of stages in QA
• 3000+ dev & QE in 10+ offices across
time zones and geographic regions
4.
5. SCALABILITY CHALLENGE
Challenges:
• Traditional 1 app 1 module does not scale
• High velocity environment with ever increasing speed of
change
• New pkgs, sunset pkgs, dependency changes, dev staff
operating 24x5
• Lack of puppet expertise to complement 3000+ technology
staff across geographic regions
Solution:
• Ninja engine generate resources dynamically
• Dependency discovery
• Puppet code change not required
Confidential and Proprietary
6. ROLES & LABELS
Role
• One role per pool
• Define a set of packages to install
Label
• A set of versioned packages
• Backed by a yum repository
6 Confidential and Proprietary
9. 9
SYSTEM HIERARCHY
Confidential and Proprietary
HomeWeb
1001
HomeWeb
1002
HomeSvc
1001
HomeSvc
1002
HomeBE
2020
HomeBE
2021
Host
HomeWeb HomeSvc HomeBEPool
SF
NA
ProdEnv
DC
Geo
LAX
10. ENC / HIERA
10 Confidential and Proprietary
• Mongo DB for hierarchical datastore
• Reduced multiple Hiera calls to one for classes, role,
parameters look up
• Efficiency & easier debug
• Created a web based tool to visualize Hiera data
• REST API for CRUD operations on Hiera data
13. MCOLLECTIVE AT SCALE
• Query systems through facts, agents, or regular expressions
[peadmin@puppet ~]$ mco find -F processorcount=24
• Verify package versions in all systems for simple auditing purposes
[peadmin@puppet ~]$ mco package status python-qpid
---- package agent summary ----
Nodes: 3366 / 3366
Versions: 3356 * 0.14-6.el5, 10 * absent
Elapsed Time: 34.68 s
• Kick off puppet runs
• ssh script replacement
• REST API enables Mcollective to web and other tools
13 Confidential and Proprietary
14. 14 Confidential and Proprietary
Large number of
Applications and Services
Real Time Status Updates
of Puppet Runs
17. PROGRESS PUPPET MODULE
• Open source code available at
• https://github.com/hunner/progress_mq
• Modified since then
• To enable/disable messages for different puppet runs
• To format JSON messages when writing to a file
Confidential and Proprietary