Everything Will Flow
summary
Claims that unbounded queues are fundamentally broken and that back pressure plus clear limits beat opaque buffering. Shows tools and code: Clojure libraries (a durable queue and S3 journal), an instrumented threadpool, and monitoring of high percentiles. Describes a production service that scales from about 40,000 requests to half a million requests per second, uses disk-backed buffering then background S3 uploads, and reports that moving to bounded/back-pressured design stabilizes the system.
AI-written summary · about
chapters (34)
- 0:00Queues
- 2:47Queue Theory
- 3:40Closed Systems
- 4:20Open Systems
- 5:13Exponential Distribution
- 5:59Pareto Distribution
- 6:40Arrival Times
- 7:35Latency Spectrogram
- 12:10Complex Tasks
- 12:48Un unbounded queues
- 13:18Too much data
- 13:35Drop the data
- 15:07Rejection
- 15:29The Only Thing We Can Do
- 16:50Queues Are Useful
- 17:23Unbuffered Queues
- 18:15Back Pressure
- 19:55MA
- 21:02Key Properties
- 21:56Background Process
- 23:18First Principles
- 23:54Closure Libraries
- 25:18S3 Journal
- 27:46JSON Endpoint
- 29:30Better Metrics
- 29:54Tools
- 30:55Chat Server
- 31:15Closure
- 32:15Active Publishing
- 32:50Buffer Write
- 34:17Spam
- 34:38Limit
- 35:23Problem Space
- 37:54Defaults
related talks
38mA Practical Look at Performance Theory
05mScale Fail
30mQueueing Theory, In Practice: Performance Modelling in Cloud-Native Territory [I]
46mQueueing Theory in Practice: Performance Modeling for the Working Engineer
45mReactive web services
44mActive Queue Management - What it is, and why you should care
19mFlash Sale Engineering
19mCome Hell or Still Water: Alleviating Tail Latency in Cloud Block Store