JACKPOT: Online Experimentation of Cloud Microservices
summary
JACKPOT is a prototype system that runs online experiments across interacting microservices on a public container cloud, using Istio service mesh for traffic splits and distributed traces. It turns multiple KPIs into a single multivariate sigmoid utility (amplification factor makes constraints hard or soft), maintains Bayesian beliefs, and uses Monte Carlo + Thompson-sampling variants (2-STS top-2 sampling, 1-STS, UNIF). Evaluation with 10 paths, 50 requests/epoch, 100 epochs, 5 runs shows 2-STS finds the best path with ~49% fewer epochs than UNIF and ~63% fewer than 1-STS, while 1-STS maximizes mean reward to 0.7 fastest.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:22Cloud Microservices in Today's World
- 0:51Web & Mobile Online Experimentation
- 1:28Cloud Challenges
- 2:12Further Challenges Posed by Microservices
- 3:00Jackpot: Online Experimentation of Cloud Microservices
- 3:28Design Choices
- 5:15Jackpot Internals
- 6:05Jackpot's Workflow
- 6:32Multivariate Sigmoid
- 11:14Experiments
- 11:57Best Path Identification
- 13:12Utility Maximization
- 13:56Next Steps
related talks
18mChaos Experiments under the lens of AIOps
39mResilience & Chaos Engineering in the Cloud
56mHow to get a grip on your microservices system using a service-mesh
40mTo Build My Own Cloud with Blackjack…
51mHow to get a grip on your microservices system using a service mesh
36mWild Microservices and Imaginary DevOps
38mPatterns and Practices for Real-world event-driven microservices
11mFrom 0 to 10,000 experiments