300 likes | 426 Views
Rapidly Developing FarmVille. How we built and scaled a # 1 Facebook game in 5 weeks. Amitt Mahajan Lead Developer – FarmVille March 9, 2010. 5 week development cycle. FarmVille. 6 developers (PHP/Flash) 2 artists 3 producers/designers. At Launch. Today. FarmVille.
E N D
Rapidly Developing FarmVille How we built and scaled a #1 Facebook game in 5 weeks • AmittMahajan • Lead Developer – FarmVille • March 9, 2010
5 week development cycle FarmVille • 6 developers (PHP/Flash) • 2 artists • 3 producers/designers
At Launch • Today FarmVille • 18K users/day after 24 hrs • 1M users/day after 4 days • 110M+ Installs • 31M Players/Day
My Goals • Reduce the amount of time it takes you to develop your games • Give you a blueprint for how to reliably scale your games to millions without your servers falling over
What slows down developers? • Other developers • Design / Art • Production (Copy) • Lack of knowledge
Flash (Client) • PHP (Server) Increasing Developer Efficiency • All developers know both • Design DOESN’T rule all • Developers co-own features with designers
Data Driven Design • Shoot for a content pipeline that doesn’t need a developer • Data driven can be as easy as a designer-editable XML file No devneeded!
String Tables • A string table is an external file that holds strings for the app • Best practice to prepare for localization • Developers are not blocked by production • Allows quick response to Facebook TOS changes
Abstracted Network Layer PHP Flash • Feature A • Feature B • Feature A Service • Feature B Service Networked Actions • Feature developers get client/server communication and serialization for free Goal Action Dispatcher Network Layer Network Layer Validation Checks Validation Checks AMF Protocol (RPC) Result Batching Action Batching
Network Layer Batching • No need to do the work twice • Reduced server load in FarmVille by 50%! Server Client Only One Request onPlow() Plow Plow onPlow() Plant Plant onPlant() Plow …
Network Layer Validation • Solves problem of unfiltered input • Ensures data is received by server in-order it was sent
Social Network Wrapper • Single place to perform Facebook API calls • Working with a highly dynamic API can be difficult • Abstracting FB calls makes them easier to integrate • Allows for cross-platform games
Continuous Deployment • Build latest version of source-repo • Deploy to test “auto app” on Facebook to surface production issues early
Continuous Deployment FarmVille QA Process Auto app Staging Production Quick Smoke Full Test Pass
FarmVille Traffic Growth 1 Million DAUsevery week for 20 weeks source: developeranalytics.com
Scalable Server Architecture • We had no choice but to scale on the cloud • Every part of the server architecture scales horizontally • No single points of failure • Take the DB out of the equation
Round Robin DNS … • Load Balancer #1 • Load Balancer #N • Auto-scaling web array (PHP) … • Web Server #M • Web Server #1 User Data & Updates • Memcache Pool Lazy Writes DB Layer Data Reads … • DB-#1 • Master • DB-#K • Master … • DB-#1 • Slave • DB-#K • Slave
All you need to implement is… … • Web Server #M • Web Server #1 …everything else is off-the-shelf components! …
Reducing Load-times • Users want responsive pages • Show something immediately • Player’s won’t wait for your app to load • Stream non-critical content
Avoid Remote Calls • Remote server calls are slow and unreliable • Aim for no remote calls during load • Embed data into Javascript Facebook Iframe Inline-JS here Flash ClientEmbed
Caching Slow Calls • Build in FacebookAPI caching within the social network wrapper • Write a DB wrapper to generate and cache SQL
FarmVille Page Profiler • Catch and eliminate all un-cached Facebook and DB calls Goal
Fault Tolerance • Servers do go down randomly • Build redundancy on all levels of server architecture • Facebook is a dependency also • Aim for DB-less & Facebook-less modes • “Defcon”-style error management
Runtime Config • How do we progressively keep the game running? • Create all features with kill-switches • Create a web dashboard to allow non-technical folks to help out
Is it still running? • Notify your team when things break using server monitoring (Nagios/Munin) • What to watch • Server load/traffic graphs • Memcacheevictions
Stats Tracking • How do we know what users are seeing? • Have the client send back statistics • Number of errors • Load-times • Take a metrics driven approach to error handling
Before Launching • Get confidence that your stuff actually works • Perform load testing • Social games are a marathon not a sprint • Sleep before launching!