Everything Will Flow
summary
Argues that unbounded queues cause sudden collapse and recommends back pressure instead of buffering, showing a simulation where slight fattening of task tails triggers an orders-of-magnitude latency jump. Describes production choices: core.async semantics, Netty thread pools, instrumented thread-pool metrics with quantiles (median vs 99.9% can differ ~3x), and building durable-queue and s3-journal to handle 40k→300k qps and about 3–4 trillion events without OOM by persisting to disk before S3. Gives concrete fixes: throttling with credits and drop/reject/pause strategies.
AI-written summary · about
chapters (21)
- 0:00Start
- 3:44QUEUEING THEORY (noun)
- 5:21OPEN
- 6:04PRODUCERS ARE EXPONENTIALLY
- 6:28CONSUMERS ARE PARETO
- 7:00A SIMULATION
- 11:41WHAT WE'VE LEARNED
- 13:25DROP
- 14:25REJECT
- 15:42BACKPRESSURE SWIMS UPSTREAM
- 16:52UNBUFFERED QUEUES ARE CLOSED SYSTEMS
- 17:11BACKPRESSURE AINT FREE
- 20:18A CONCRETE EXAMPLE
- 21:07KEY PROPERTIES
- 22:10PREVIOUSLY...
- 24:35AND SO...
- 26:31NETTY'S THREADING MODEL
- 27:22METRICS?
- 31:27AN ABSTRACT EXAMPLE
- 32:51PUBLISHING
- 39:21WE TALKED ABOUT QUEUES
related talks
02h 42m!!! Do not publish or share !!!!
42mHow Flow Works and other curiosities
40mResilient Design 101: Queue Theory
25mAvoiding Overload in Distributed Systems
50mHow Flow Works and other Curiosities
54mHow I Learned to Stop Worrying & Love the Flow
44mDeep Dive talk | Resilience in Microservices with Congestion Control
47mEverything Loosely Coupled - Advanced Integration Patterns and Their Trade-Offs