Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Building Adaptive Systems
Search
Chris Keathley
May 28, 2020
Programming
43
2.7k
Building Adaptive Systems
Chris Keathley
May 28, 2020
Tweet
Share
More Decks by Chris Keathley
See All by Chris Keathley
Solid code isn't flexible
keathley
5
1.1k
Contracts for building reliable systems
keathley
6
900
Kafka, the hard parts
keathley
3
1.7k
Building Resilient Elixir Systems
keathley
7
2.3k
Consistent, Distributed Elixir
keathley
6
1.6k
Telling stories with data visualization
keathley
1
630
Easing into continuous deployment
keathley
2
390
Leveling up your git skills
keathley
0
770
Generative Testing in Elixir
keathley
0
520
Other Decks in Programming
See All in Programming
Flutterと Vibe Coding で個人開発!
hyshu
1
230
あまり知られていない MCP 仕様たち / MCP specifications that aren’t widely known
ktr_0731
0
220
MCPで実現できる、Webサービス利用体験について
syumai
7
2.4k
MySQL9でベクトルカラム登場!PHP×AWSでのAI/類似検索はこう変わる
suguruooki
1
280
AI Ramen Fight
yusukebe
0
130
リッチエディターを安全に開発・運用するために
unachang113
1
350
CLI ツールを Go ライブラリ として再実装する理由 / Why reimplement a CLI tool as a Go library
ktr_0731
3
960
「リーダーは意思決定する人」って本当?~ 学びを現場で活かす、リーダー4ヶ月目の試行錯誤 ~
marina1017
0
120
React 使いじゃなくても知っておきたい教養としての React
oukayuka
18
5.3k
令和最新版手のひらコンピュータ
koba789
4
1.3k
書き捨てではなく継続開発可能なコードをAIコーディングエージェントで書くために意識していること
shuyakinjo
0
200
Quality Gates in the Age of Agentic Coding
helmedeiros
PRO
1
120
Featured
See All Featured
Building Better People: How to give real-time feedback that sticks.
wjessup
367
19k
What’s in a name? Adding method to the madness
productmarketing
PRO
23
3.6k
Faster Mobile Websites
deanohume
308
31k
Making the Leap to Tech Lead
cromwellryan
134
9.5k
The Pragmatic Product Professional
lauravandoore
36
6.8k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
283
13k
Optimising Largest Contentful Paint
csswizardry
37
3.4k
Put a Button on it: Removing Barriers to Going Fast.
kastner
60
3.9k
Agile that works and the tools we love
rasmusluckow
329
21k
Principles of Awesome APIs and How to Build Them.
keavy
126
17k
Fashionably flexible responsive web design (full day workshop)
malarkey
407
66k
Scaling GitHub
holman
461
140k
Transcript
Chris Keathley / @ChrisKeathley /
[email protected]
Building Adaptive Systems
Server Server
Server Server I have a request
Server Server
Server Server
Server Server No Problem!
Server Server
Server Server Thanks!
Server Server
Server Server I have a request
Server Server
Server Server
Server Server I’m a little busy
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I’m a little busy I have more requests!
Server Server I don’t feel so good
Server
Server Welp
Server Welp
All services have objectives
A resilient service should be able to withstand a 10x
traffic spike and continue to meet those objectives
Lets Talk About… Queues Overload Mitigation Adaptive Concurrency
Lets Talk About… Queues Overload Mitigation Adaptive Concurrency
What causes overload?
What causes overload? Server Queue
What causes overload? Server Queue Processing Time Arrival Rate >
Little’s Law Elements in the queue = Arrival Rate *
Processing Time
Little’s Law Server 1 requests = 10 rps * 100
ms 100ms
Little’s Law Server 1 requests = 10 rps * 100
ms 100ms
Little’s Law Server 1 requests = 10 rps * 100
ms 100ms
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms BEAM Processes
Little’s Law Server 2 requests = 10 rps * 200
ms 200ms BEAM Processes CPU Pressure
Little’s Law Server 3 requests = 10 rps * 300
ms 300ms BEAM Processes CPU Pressure
Little’s Law Server 30 requests = 10 rps * 3000
ms 3000ms BEAM Processes CPU Pressure
Little’s Law Server 30 requests = 10 rps * ∞
ms ∞ BEAM Processes CPU Pressure
Little’s Law 30 requests = 10 rps * ∞ ms
Little’s Law ∞ requests = 10 rps * ∞ ms
Little’s Law ∞ requests = 10 rps * ∞ ms
This is bad
Lets Talk About… Queues Overload Mitigation Adaptive Concurrency
Lets Talk About… Queues Overload Mitigation Adaptive Concurrency
Overload Arrival Rate > Processing Time
Overload Arrival Rate > Processing Time We need to get
these under control
Load Shedding Server Queue Server
Load Shedding Server Queue Server Drop requests
Load Shedding Server Queue Server Drop requests Stop sending
Autoscaling
Autoscaling
Autoscaling Server DB Server
Autoscaling Server DB Server Requests start queueing
Autoscaling Server DB Server Server
Autoscaling Server DB Server Server Now its worse
Autoscaling needs to be in response to load shedding
Circuit Breakers
Circuit Breakers
Circuit Breakers Server Server
Circuit Breakers Server Server
Circuit Breakers Server Server Shut off traffic
Circuit Breakers Server Server
Circuit Breakers Server Server I’m not quite dead yet
Circuit Breakers are your last line of defense
Lets Talk About… Queues Overload Mitigation Adaptive Concurrency
Lets Talk About… Queues Overload Mitigation Adaptive Concurrency
We want to allow as many requests as we can
actually handle
None
Adaptive Limits Time Concurrency
Adaptive Limits Actual limit Time Concurrency
Adaptive Limits Actual limit Dynamic Discovery Time Concurrency
Load Shedding Server Server
Load Shedding Server Server Are we at the limit?
Load Shedding Server Server Am I still healthy?
Load Shedding Server Server
Load Shedding Server Server Update Limits
Adaptive Limits Time Concurrency Increased latency
Latency Successful vs. Failed requests Signals for Adjusting Limits
Additive Increase Multiplicative Decrease Success state: limit + 1 Backoff
state: limit * 0.95 Time Concurrency
Prior Art/Alternatives https://github.com/ferd/pobox/ https://github.com/fishcakez/sbroker/ https://github.com/heroku/canal_lock https://github.com/jlouis/safetyvalve https://github.com/jlouis/fuse
Regulator https://github.com/keathley/regulator
Regulator.install(:service, [ limit: {Regulator.Limit.AIMD, [timeout: 500]} ]) Regulator.ask(:service, fn ->
{:ok, Finch.request(:get, "https://keathley.io")} end) Regulator
Conclusion
Queues are everywhere
Those queues need to be bounded to avoid overload
If your system is dynamic, your solution will also need
to be dynamic
Go and build awesome stuff
Thanks Chris Keathley / @ChrisKeathley /
[email protected]