Algorithms in the real world (NUS Friday Hacks)

Have you ever had to learn an algorithm and thought, "No way is this ever going to come up in real life"? It turns out many of the topics you're learning in class come up in unexpected ways in the software industry. Come learn how Stripe uses algorithms to build the world's global payments network, setting the new standard in online payments

Transcript

1. Algorithms in the Real World "When am I ever going to need this?" Amy Nguyen an@stripe.com Software Engineer, Global Team (APAC)
2. Who am I? • Class of 2015 • Used to think I was bad at math • Failed algorithms in university and had to take it twice • Once had to write breadth-first search for a work project! • Still have a job as an engineer
3. What is Stripe? • Our mission: Increase the GDP of the internet. • How I like to think about it: Do whatever it takes to help people start businesses online.
5. • Algorithms appear in unexpected places • Unlike university, you're not alone! • You have to think beyond time and space complexity • Algorithms are easier to understand when you have a concrete problem and real-world constraints to think about Today's takeaways
6. Three Examples 1. Rate limiting the Stripe API 2. Two practical uses for compiler theory 3. Reducing human bias in machine learning algorithms
7. Rate limiting the Stripe API
8. What is rate limiting? • Limit the number of requests per second
9. What is rate limiting? • Limit the number of concurrent requests
10. Why rate limit?
11. 1. You want to limit the impact of one user's spikes Why rate limit?
12. 1. You want to limit the impact of one user's spikes 2. You're receiving heavy traffic and want to prioritize critical requests Why rate limit?
13. 1. You want to limit the impact of one user's spikes 2. You're receiving heavy traffic and want to prioritize critical requests 3. Your systems might be broken Why rate limit?
14. How do you rate limit? 1. Pretend you have buckets (one for every user)
15. How do you rate limit? 1. Pretend you have buckets (one for every user) 2. Every N seconds, add a drop to each non-full bucket
16. How do you rate limit? 1. Pretend you have buckets (one for every user) 2. Every N seconds, add a drop to each non-full bucket
17. How do you rate limit? 1. Pretend you have buckets (one for every user) 2. Every N seconds, add a drop to each non-full bucket
18. How do you rate limit? 1. Pretend you have buckets (one for every user) 2. Every N seconds, add a drop to each non-full bucket 3. For a request, only serve it if you have enough water
19. Practical uses for compiler theory
20. Sorbet: A static type checker for Ruby • Uses ASTs, CFGs, and DSLs to parse Ruby code • Abstract Syntax Tree (AST) and Context Free Grammar (CFG): Represents the structure of the code and rules for what is valid • Domain-specific Language (DSL): Specialized for Ruby type-checking, not a general-purpose language • Written in C++ for performance
21. Radar: Customized rules for fraud detection
22. Bias in machine learning decisions
23. Why did my card get rejected?
24. Decision trees Was the card issued in the US? yes no Were there more than 12 charges on the card in the last 30 minutes? Were there more than 4 charges on the card in the last 30 minutes? Fraudulent: 81 Legitimate: 19 P(fraud) = 81% Fraudulent: 12 Legitimate: 300 P(fraud) = 3.8% Fraudulent: 12 Legitimate: 8 P(fraud) = 60% Fraudulent: 4 Legitimate: 112 P(fraud) = 3.4% no yes no yes no
25. Random forests
29. Why did my card really get rejected?
30. "...when you ask a human you’re not really getting at the decision process. They make a decision first, and then you ask, and then they generate an explanation and that may not be the true explanation." Peter Norvig Director of Research, Google
31. “If I apply for a loan and I get turned down, whether it’s by a human or by a machine, and I say what’s the explanation, and it says well you didn’t have enough collateral. That might be the right explanation or it might be it didn’t like my skin colour. And I can’t tell from that explanation.” Peter Norvig Director of Research, Google
32. How could bias appear?
33. How to measure bias Fraud Not fraud Flagged Positive False positive Not flagged False negative Negative • Rate of true positives (good users are good) • Rate of false negatives (good users aren't rejected disproportionately)
34. Why can't we just fix it? • Complexity • User experience • Cost • Prioritization • Legality
36. "Haven't I taught you anything? What have I always told you? Never trust anything that can think for itself if you can't see where it keeps its brain?" Arthur Weasley
37. • Algorithms appear in unexpected places • Unlike university, you're not alone! • You have to think beyond time and space complexity • Algorithms are easier to understand when you have a concrete problem and real-world constraints to think about Today's takeaways
38. Thanks! • We're hiring interns and new graduates: stripe.com/jobs • Slides: speakerdeck.com/amyngyn