Empowering AWS DevOps with Python and Machine Learning
summary
Uses Python data tools and a simple ML regression to map RPM to CPU (40k RPM → ~30% CPU) and to discover a latency cutoff at ~50% CPU where response times explode. Builds a pandas, Jupyter, Locust, boto3 workflow plus a scaling simulator and traffic shape generator to sweep autoscaling parameters. Simulation selects upper threshold 15% and scale-up increment 2; applied to production it yields average CPU ≈16%, peak ≈39% and stable P99 latency (~40 ms), so the change works.
AI-written summary · about
chapters (27)
- 0:00intro
- 0:20preamble
- 0:37intro
- 1:09agenda
- 1:53scaling is hard
- 3:24problem
- 7:04context
- 9:31approach
- 10:34load test one host
- 11:05retrieve data from aws to pandas
- 12:59analyse one host performance
- 13:14we can do better
- 14:28another thing - latency
- 17:03approaches
- 18:45create a scaling simulator
- 20:12create a traffic shape generator
- 20:52and run a first local experiment
- 22:40what does it mean?
- 22:59first attempt
- 24:13find the best parameters
- 25:08and the winner is
- 26:29which actually looks better
- 26:53testing new parameters in production
- 28:42simulation versus real life
- 31:19conclusion
- 31:41code
- 33:22about gustavo
related talks
14mBattle-tested tactics to accelerate data science and ML R&D with AWS
14mPerformance Optimization in High-Volume Transaction Systems
28mShift your ML model into overdrive with AWS DeepRacer analysis tools
17mR & Python playing nice, in Production
27mLeverage ML to automatically tune K8s apps
17mProductionizing Data Science Workloads
30mOn scaling Machine Learning microservices
32mAutomating Production Production Level Machine Learning on AWS