Scaling The Software Development Process Lessons Learned from

Scaling The Software Development Process: Lessons Learned from The Sims Online Greg Kearney, Larry Mellon, Darrin West Spring 2003, GDC 1

Talk Overview • Covers: Software Engineering techniques to help when projects get big – Code structure – Work processes (for programmers) – Testing • Does Not Cover: – Game Design / Content Pipeline – Operations / Project Management 2

How to Apply it. • • We didn’t do all of this right away Improve what you can Don’t change too much at once Prove that it works, and others will take up the cause • Iterate 3

Match Process to Scale +tve Team Efficiency 0 Change to a new process “Everything’s Broken Hell” Team Size “Meeting Hell” Process for 5 to 15 programmers Process for 30 to 50 programmers 4

What You Should Leave With • TSO “Lessons Learned” – Where we were with our software process – What we did about it – How it helped • Some Rules of Thumb – General practices that tend to smooth software development @ scale – Not a blueprint for MMP development – Useful “frame of reference” 5

Classes of Lessons Learned & Rules • Architecture / Design: Keep it Simple – Minimizing dependencies, fatal couplings – Minimizing complexity, brittleness • Workspace Management: Keep it Clean – Code and directory structure – Check in and integration strategies • Dev. Support Structure: Make it Easy, Prove it – Testing – Automation -All of these had to change as we scaled up. -They eventually exceeded the team’s ability to deal with (using existing tools & processes). 6

Non-Geek Analogy –Sharpen your tools. –Clean up your mess. –Measure twice, cut once. –Stay with your buddy. Bad flashbacks found at: http: //www. easthamptonhigh. org/cernak/ http: //www. hancock. k 12. mi. us/high/art/wood/index. html 7

Key Factors Affecting Efficiency • High “Churn Rate”: large #coders times tightly coupled code equaled frequent breaks – Our code had a deep root system – And we had a forest of changes to make “Big root ball” found at: http: //www. on. ec. gc. ca/canwarn/norwich/norsummary-e. html 8

Make It Smaller Evolve 9

Key Factors Affecting Efficiency • “Key Logs”: some issues were preventing other issues from even being worked on 10

Key Factors Affecting Efficiency • A chain of single points of failure took out the entire team Login Create an avatar Enter a city Buy a house Enter a house Buy the chair Sit on a chair 11

So, What Did We Do That Worked • Switched to a logical architecture with less coupling • Switched to a code structure with fewer dependencies • Put in scaffolding to keep everyone working • Developed sophisticated configuration management • Instituted automated testing • Metrics, Metrics 12

So, What Did We Do That Didn’t? • • Long range milestone planning Network emulator(s) Over engineered a few things (too general) Some tasks failed due to: – Not replanning, reviewing long tasks – Not breaking up long tasks • Coding standard changed part way through • … 13

What we were faced with: • • • 750 K lines of legacy Windows code Port it to Linux Change from “multiplayer” to Client/Server 18 months Developers must remain alive after shipping Continuous releases starting at Beta 14

Go To Final Architecture ASAP 15

Go to final architecture ASAP Multiplayer: Client/Server: Client Sim Here be Sync Hell Client Sim Evolve Client Sim Nice Undemocratic Request/ Command Client 16

Final Architecture ASAP: “Refactoring” • Decomposed into Multiple dll’s – Found the Simulator • Interfaces • Reference Counting • Client/Server subclassing How it helped: –Reduced coupling. Even reduced compile times! –Developers in different modules broke each other less often. –We went everywhere and learned the code base. 17

Final Architecture ASAP: It Had to Always Run • But, clients would not behave predictably • We could not even play test • Game design was demoralized • We needed a bridge, now! ? ? 18

Final Architecture ASAP: Incremental Sync • A quick temporary solution… – Couldn’t wait for final system to be finished – High overhead, couldn’t ship it • We took partial state snapshots on the server and restored to them on the client How it helped: –Could finally see the game as it would be. –Allowed parallel game design and coding –Bought time to lay in the “right” stuff. 19

Final Architecture ASAP: Null View • Created Null View House. Sim on Windows – Same interface – Null (text output) implementation How it helped –No ifdef’s! –Done under Windows, we could test this first step. –We knew it was working during the port. –Allowed us to port to Linux only the “needed” parts. 20

Final Architecture ASAP: More “Bridges” • HSB’s: proxy on Linux, pass-through to a Windows Sim. How it helped –Could exercise Linux components before finishing House. Sim port. –Allowed us to debug server scale, performance and stability issues early. –Make best use of Windows developers. –Allowed single platform development. Faster compiles. • Disabled authentication, etc. How it helped –Could keep working even when some of the system wasn’t available. 21

Mainline *Must* Work! 22

If Mainline Doesn’t Work, Nobody Works • The Mainline source control branch *must* run • Never go dark: Demo/Play Test every day • If you hit a bug, do you sync to mainline, hoping someone else fixed it? Or did you just add it? –If mainline breaks for “only” an hour, the project loses a man-week. –If each developer breaks the mainline “only” once a month, it is broken every day. 23

Mainline must work: Sniff Test • Mainline was breaking for “simple” things. – Features you “didn’t touch” (and didn’t test). • Created an auto-test to exercise all core functions. • Quick to run. Fun to watch. Checked results. • Mandated that it pass before submitting code changes. • Break the build: “feed the pig”. How it helped –Very simple test. Amazing difference. –Sometimes we got lazy and trusted it too much. ! oh D 24

Mainline must work: Stages to “Sandboxing” 1. Got it to build reliably. 2. Instituted Auto-Builds: email all on failure. 3. Used a “Pumpkin” to avoid duplicate mergetest cycles, pulling partial submissions, . . . 4. Used a Pumpkin Queue when we really got rolling How it helped –Far fewer thumbs twiddled. –The extra process got on some people’s nerves. 25

Mainline must work: Sandboxing 5. Finally, went to per-developer branching. – – Develop on your own branch. Submit changes to an integration engineer. Full Smoke test run per submission/feature. If it worked, integrated to mainline in priority order, or else it is bounced. How it helped –Mainline *always* runs. Pull any time. –Releases are not delayed by partial features. –No more code freezes going to release. 26

Support Structure 27

Background: Support Structure • Team size placed design constraints on supporting tools – Automation: big win in big teams – Churn rate: tool accuracy / support cost • Types of tools – Data management: collection / corrolation – Testing: controlled, sync’ed, repeatable inputs – Baselines: my bug, your bug, or our bug? 28

Overview: Support Structure • Automated testing: designs to minimize impact of churn rate • Automated data collection / corrolation – Distributed sytem == distributed data – Dashboard / Esper / Monkey. Watcher • Use case: load testing – Controlled (tunable) inputs, observable results – “Scale&Break” 29

Problem: Testing Accuracy • Load & Regression: inputs must be – Accurate – Repeatable • Churn rate: logic/data in constant motion – How to keep testing client accurate? • Solution: game client becomes test client – Exact mimicry – Lower maintenance costs 30

Test Client == Game Client Test Client Game Client Test Control Game GUI State Commands Presentation Layer Client-Side Game Logic 31

Game Client: How Much To Keep? Game Client View Presentation Layer Logic 32

What Level To Test At? Mouse Clicks Game Client View Presentation Layer Logic Regression: Too Brittle (pixel shift) Load: Too Bulky 33

What Level To Test At? Game Client View Internal Events Presentation Layer Logic Regression: Too Brittle (Churn Rate vs Logic & Data) 34

Semantic Abstractions Basic gameplay changes less frequently than UI or protocol implementations. View Logic Null. View Client Presentation Layer Buy Lot Buy Object … Enter Lot ~¾ ~¼ Use Object 35

Scriptable User Play Sessions • Test Scripts: Specific / ordered inputs – Single user play session – Multiple user play session • Sim. Script – Collection: Presentation Layer “primitives” – Synchronization: wait_until, remote_command – State probes: arbitrary game state • Avatar’s body skill, lamp on/off, … 36

Scriptable User Play Sessions • Scriptable play sessions: big win – Load: tunable based on actual play – Regression: walk a set of avatars thru various play sessions, validating correctness per step • Gameplay semantics: very stable – UI / protocols shifted constantly – Game play remained (about) the same 37

Automated Test: Team Baselines • Hourly “critical path” stability tests – Sync / clean / build / test – Validate Mainline / Servers • Snifftest weather report – Hourly testing – Constant reporting 38

How Automated Testing Helped • Current, accurate baseline for developers • Scale&break found many bugs • Greatly increased stability – Code base was “safe” – Server health was known (and better) 39

Tools & Large Teams • High tool ROI – team_size * automation_savings • Faster triage – Quickly narrow down problem – across any system component • Monitoring tools became a focal point • Wiki: central doc repository 40

Monitoring / Diagnostics When you can measure what you are speaking about and can express it in numbers, you know something about it. But when you cannot measure it, when you cannot express it in numbers, your knowledge is of a meager and unsatisfactory kind. " - Lord Kelvin • De. Marco: You cannot control what you cannot measure. • Maxwell: To measure is to know. • Pasteur: A science is as mature as its measurement tools. 41

Dashboard • System resource & health tool – CPU / Memory / Disk / … • Central point to access – Status – Test Results – Errors – Logs – Cores –… 42

Test Central / Monkey Watcher • Test Central UI – Control rig for developers & testers • Monkey Watcher – – – Collects & stores (distributed) test results Produces summarized reports across tests Filters known defects Provides baseline of correctness Web frontend, unique IDs per test 43

Esper • In-game profiler for a distributed system • Internal probes may be viewed – Per process / machine / cluster – Time view or summary view • Automated data management – Coders: add one line probe – Esper: data shows up on web site 44

Use Case: Scale & Break • Never too early to begin scaling – Idle: keep doubling server processes – Busy: double #users, dataset size • Fix what broke, start again • Tune input scripts using Beta data 45

Load Testing: Data Flow Resource Metrics Load Testing Team Debugging Data Client Metrics Load Control Rig Test Test Client Client Test Client Test Driver CPU Game Traffic System Monitors Server Cluster Internal Probes

Outline: Wrapup • • Wins / Losses Rules: Analysis & Discussion Recommended reading Questions 47

Process: Wins / Losses • Wins – Module decomposition • Logical: client / server architecture • Physical: code structure – Scaffolding for parallel development – Tools to improve workflow – Automated Regression / Load 48

Process: Wins / Losses • Losses – Early lack of tools – #ifdef as a cross-platform port – Single points of failure blocked entire development team 49

Not Done Yet: More Challenges • How to ship, and ship… • How to balance infrastructure cleanup against new feature development • … 50

Rules of Thumb (1) • KISS: software and processes • Incremental changes – – : “Baby-Steps” • Continual tool/process improvement 51

Rules of Thumb (2) • Mainline has got to work • Get something on the ground. Quickly. 52

Rules of Thumb (3) • • • Key Logs: break up quickly, ruthlessly Scaffolding: keep others working Do important things, not urgent things Module separation (logically, physically) If you can’t measure it, you don’t understand it 53

Final Rule: “Sharpen The Saw” • Efficiency impacted by – Component coupling / team size – Compile / load / test / analyze cycle • Tool Justification in large teams – Large ROI @ large scale – 5% gain across 30 programmers – “Fred Brooks”: 31 st programmer… 54

Recommended Reading • Influences – Extreme Programming – Scott Meyers: large-scale software engineering – Gamma et al: Design Patterns • Caveat Emptor: slavish following not encouraged – Consider “ground conditions” for your project 55

Questions & Answers 56