Agile, Lean & Rugged - The Paper Edition
summary
Parnass argues that software ages unless it is designed for change, recommending modularity, information hiding, documentation and planning for replacement; Facebook’s Scuba team decouples memory from process to cut restarts from ~3 hours to 2–3 minutes per server and fleet downtime from ~12 hours to ~1 hour. McSherry shows systems that scale well can still be slower and defines a cost metric where many distributed frameworks never beat a single core. BlinkDB and sampling give 1% sampling ≈30× faster, and lineage-driven injector Molly finds faults by reasoning backward; focus on yield and harvest for availability.
AI-written summary · about
chapters (10)
- 0:00Start
- 1:41Yearly Performance Goals
- 4:04Looking at the Foundations
- 6:25David Palmer
- 9:03Software Aging
- 19:28Frontier Paper
- 19:58Analyses of the Size of Articles in Wikipedia
- 21:54How Does the Prox Hadoop Actually Work
- 23:15Aggregate Computations
- 27:32Probabilistic Availability
related talks
39mWe're Agile, we don't do documentation
36mGoing Modular: Teaching Old Docs New Tricks
58mNitpicking terminology - are we using the right terms
29mMove Fast and Document Things: Writing internal docs at fast-moving organizations
28mGood Enough at Scale: Documentation Infrastructure Under Capacity Constraints
22mUncharted waters: Documenting emerging technology
54mPanel: Mechanized Specifications for Real-World Programming Languages
11mInterview with Simon St. Laurent