Approaching the Unacceptable Workload Boundary
summary
Systems degrade across a fuzzy unacceptable workload boundary rather than a sharp limit; load is usefully defined as the sum of residence times (average concurrency). Throughput vs concurrency scatter plots are fitted to the Universal Scalability Law to infer sigma and kappa (serialization and crosstalk), using hundreds of samples (one‑ to five‑minute averages) of per‑request latency. An Excel regression or existing APIs (example: VividCortex per‑query data) yields parameters, shows the knee and retrograde scalability, and predicts when the system enters the death spiral.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 5:14The Actual Boundaries Are Unknown
- 6:14The Buffer Zone Is Nonlinear
- 7:02Complex Systems Run In Degraded Mode
- 8:23What Is The Definition of Load?
- 9:35Load, Utilization, And Queueing
- 11:23What Is The Load Limit?
- 13:07What's the Definition of Scalability?
- 14:10Linear Scaling
- 15:41The Linear Scalability Equation
- 17:22Amdahl's Law Describes Serialization
- 17:59Amdahl's Law Has An Asymptote
- 18:40What If Workers Coordinate?
- 20:00How Bad Is Coordination?
- 20:35The Universal Scalability Law
- 22:21What is the USL Good For?
- 22:53What To Measure
- 24:23How To Measure Concurrency, Pt. 1
- 26:37Plot Your Data
- 28:03Plug The Data Into The Model
- 29:21Interpreting The Results What does the output mean?
- 30:00Paypal's NodeJS vs Java Benchmarks
related talks
25mAvoiding Overload in Distributed Systems
40mSystem Performance and Queuing Theory - Concepts and Application
54mThe `-ilities` of Software Engineering - How Qualities...
40mEverything Will Flow
34mDistributed Systems - What Can Go Wrong and How to Get It Right
01h 01mLittle things and big things: lessons for building large systems.
49mHigh Throughput and/or Low Latency
45mScaling under uncertainty