Operating online games is fun and challenging. Games are some of the spikiest workloads around, and real-time really means *real-time*. Randy shares many of the DevOps techniques he has been putting into practice at KIXEYE, including migrating to the cloud, organizing around services, and focusing on automation. He illustrates his points with war stories from operating large-scale services at Google and eBay.
Please see companion video at https://vimeo.com/95841677.
DevOps Chicago - The Game Of Operations and the Operation of Games
1. The Game of Operations
and
The Operation of Games
Randy Shoup
@randyshoup
linkedin.com/in/randyshoup
DevOps Chicago Meetup, May 19 2014
2. Background
CTO at KIXEYE
• Real-time strategy games for web and mobile
Director of Engineering for Google App
Engine
• World’s largest Platform-as-a-Service
Chief Engineer at eBay
• Multiple generations of eBay’s real-time
search infrastructure
3. Real-Time Strategy Games are
… • Real-time
• Spiky
• Computationally-
intensive
• Constantly evolving
• Constantly pushing
boundaries
Technically and
operationally demanding
4. Operating Games: Goals
Player Fun
• If players aren’t playing, we don’t have a business
• If players aren’t having fun, we don’t have a business for
long
• Fun includes game mechanics, feature set, quality,
performance
Studio Velocity
• 8 *highly independent* game studios
• Different tech stacks, tool chains, phases of development
Developer Productivity and Satisfaction
• We are a vendor; the studios are our customers
• Must be *strictly better* than the alternatives of build, buy,
borrow
Cost Efficiency
• More output for less
5. The Game of Operations
Cloud
• All studios and services moving to AWS
• Strong focus on automation
Services
• Small, focused teams
• Clean, well-defined interface to customers
DevOps
• Developers behave like Ops
• Ops behaves like Developers
7. Why Cloud? (The Obvious)
Provisioning Speed
• Minutes, not weeks
• Autoscaling in response to load
Near-Infinite Capacity
• No need to predict and plan for growth
• No need to defensively overprovision
Pay For What You Use
• No “utilization risk” from owning / renting
• If it’s not in use, spin it down
8. Why Cloud? (The Less
Obvious)
Instance Optimization Opportunities
• Instance shapes to fit most parts of the
solution space (compute-intensive, IO-
intensive, etc.)
• If the shape does not fit, try another
Service Quality
• Amazon and Google know how to run data
centers
• Battle-tested and highly automated
• World-class networking, both cluster fabric
and external peering
9. Why Cloud? (The
Fundamentals)
Right Side of History
• Almost impossible to beat Google / Amazon
buying power or operating efficiencies
• 2010s in computing are like 1910s in electric
power
• Soon it will be just as common to run your own
data center as it is to run your own electric power
generation (!)
Easy and Fun
• It Just Works ™
• Makes it easy to fall in love with infrastructure
10. Autoscaling
Games are very spiky
• Very unpredictable
• Huge variability between peak and trough
• Hits are self-reinforcing
Services and clients have to “flex”
• Clients back off in response to latency
• Services grow / shrink based on load
Service Cluster == AWS Auto-Scale Group
• Scale up or down based on predefined metrics,
thresholds
11. Automation Work at KIXEYE
Build / Deploy Pipeline
• One button
• Puppet -> Packer -> AMI -> Asgard
• No-downtime red-black deployment
• Futures: canarying, auto-rollback
Manageability
• Flume -> ElasticSearch / Kibana for logging
• Shinken -> PagerDuty for monitoring and
alerting
13. Service Teams
• Give teams autonomy
• Freedom to choose technology, methodology,
working environment
• Responsibility for the results of those choices
• Hold them accountable for *results*
• Give a team a goal, not a solution
• Let team own the best way to achieve the
goal
14. KIXEYE Service Chassis
• Goal: Produce a “chassis” for building scalable
game services
• Minimal resources, minimal direction
• 3 people x 1 month
• Consider building on open source projects
Team exceeded expectations
• Co-developed chassis, transport layer, service
template, build pipeline, red-black deployment, etc.
• Operability and manageability from the beginning
• Heavy use of Netflix open source projects
• 15 minutes from no code to running service in AWS
(!)
• Plan to open-source several parts of this work
16. Transition to Building Services
Common Chassis
• Make it trivially easy to build and maintain a service
Define Service Interface (Formally!)
• Propose, Discuss, Agree
Prototype Implementation
• Simplest thing that could possibly work
• Client can integrate with prototype
• Implementor can learn what works and what does not
Real Implementation
• Throw away the prototype (!)
Rinse and Repeat
17. Transition to Service
Relationships
Vendor – Customer Relationship
• Friendly and cooperative, but structured
• Clear ownership and division of responsibility
• Customer can choose to use service or not (!)
Service-Level Agreement (SLA)
• Promise of service levels by the service provider
• Customer needs to be able to rely on the service, like
a utility
Charging and Cost Allocation
• Charge customers for *usage* of the service
• Aligns economic incentives of customer and provider
• Motivates both sides to optimize
19. Instrumentation and
Measurement
Instrument Everything
• Machine / instance stats: CPU, memory, I/O
• Software infrastructure stats: database, message
queue
• Application stats: game client, game server, services
Make It Easy to Do the Right Thing ™
• Easy, reliable, low-latency
• Auto-tagged and searchable
Why?
• Measurement beats intuition every time; my own
intuition is usually wrong
• If you need to ssh into a box, instrumentation failed
you
20. One Team (!)
• Act as one team across development,
product, operations, etc.
• Solve problems instead of blaming and
pointing fingers
• Political games are not as fun as real-time
strategy games
21. Everyone Is Responsible for
Prod
Everyone’s incentives are aligned
Everyone is strongly motivated to have solid
instrumentation and monitoring
22. Organization: Learning Culture
Learn from mistakes and improve
• What did you do -> What did you learn
• Take emotion and personalization out of it
Encourage iteration and velocity
• “Failure is not falling down but refusing to get
back up” – Theodore Roosevelt
23. Google Blame-Free Post-
Mortems
Post-mortem After Every Incident
• Document exactly what happened
• What went right
• What went wrong
Open and Honest Discussion
• What contributed to the incident?
• What could we have done better?
Engineers compete to take personal
responsibility (!)
24. Transition to DevOps
Organization
• Studios make user-visible games
• Services provide common endpoints
Training / Retraining
• Common bootcamp
• Train devs as Ops, Ops as devs
You Build It, You Run It
• Transition on-call
• Use primary / secondary on-call as
apprenticeship