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
Road to masterless multi-node distributed syste...
Search
udit
December 12, 2018
Technology
0
51
Road to masterless multi-node distributed system in Elixir
udit
December 12, 2018
Tweet
Share
More Decks by udit
See All by udit
Scalable dist-sys from grounds up
yudistrange
0
170
Other Decks in Technology
See All in Technology
JAWS UG AI/ML #32 Amazon BedrockモデルのライフサイクルとEOL対応/How Amazon Bedrock Model Lifecycle Works
quiver
1
750
パフォーマンスチューニングのために普段からできること/Performance Tuning: Daily Practices
fujiwara3
2
200
Databricks Free Editionで始めるMLflow
taka_aki
0
770
How Fast Is Fast Enough? [PerfNow 2025]
tammyeverts
2
250
GPUをつかってベクトル検索を扱う手法のお話し~NVIDIA cuVSとCAGRA~
fshuhe
0
370
AWS re:Invent 2025事前勉強会資料 / AWS re:Invent 2025 pre study meetup
kinunori
0
1.1k
組織全員で向き合うAI Readyなデータ利活用
gappy50
5
2.1k
Mackerelにおけるインシデント対応とポストモーテム - 現場での工夫と学び
taxin
0
110
Boxを“使われる場”にする統制と自動化の仕組み
demaecan
0
180
OPENLOGI Company Profile for engineer
hr01
1
46k
DMARCは導入したんだけど・・・現場のつぶやき 〜 BIMI?何それ美味しいの?
hirachan
1
140
Kotlinで型安全にバイテンポラルデータを扱いたい! ReladomoラッパーをAIと実装してみた話
itohiro73
3
220
Featured
See All Featured
Chrome DevTools: State of the Union 2024 - Debugging React & Beyond
addyosmani
9
950
How Fast Is Fast Enough? [PerfNow 2025]
tammyeverts
2
250
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
32
1.7k
Connecting the Dots Between Site Speed, User Experience & Your Business [WebExpo 2025]
tammyeverts
10
640
Building an army of robots
kneath
306
46k
Designing Experiences People Love
moore
142
24k
Scaling GitHub
holman
463
140k
Navigating Team Friction
lara
190
15k
The World Runs on Bad Software
bkeepers
PRO
72
11k
jQuery: Nuts, Bolts and Bling
dougneiner
65
7.9k
How STYLIGHT went responsive
nonsquared
100
5.9k
Context Engineering - Making Every Token Count
addyosmani
8
330
Transcript
The Road to a in Elixir Masterless Multi-node Distributed System
BUILDING AND SERVICES SYSTEMS
•WHY THIS JOURNEY? •THE START •WHERE WE ARE •THE ROAD
AHEAD AGENDA
WHY THIS JOURNEY?
PLATFORM TO DISCOVER AND BUY THE BEST EXPERIENCES IN EVENTS,
TRAVEL AND FOOD IN CITIES ACROSS INDIA
FOUNDED IN 2014 10K+ EVENTS, 3M+ TICKETS IN 2018 ACQUIRED
BY PAYTM IN 2017
MAKE INSIDER A PLATFORM FOR HOSTING AND PARTICIPATING IN DIGITAL
EVENTS
LIVE QUIZ
A PLATFORM THAT ENABLES INTERACTIONS ON A LIVE VIDEO STREAM
interaction by Aman from the Noun Project
FLEXIBLE AND SCALABLE FUEL INTERACTIONS flexible by Anup from the
Noun Project PLATFORM TO
Elixir Functional Programming Immutable Data Structures Concurrency Fault Tolerance Battle
Tested VM Erlang Interoperability
THE START
M W W W W NODE
MASTER CANNOT HANDLE Problem PLAYERS
M W W W W NODE SM SM
UNNECESSARY CHANGE CALLS CASTS
PLAYER STATE TRANSFER GENSERVER ETS
M W W W W NODE C NODE B NODE
A
SENDING MESSAGE TO A PID OVER THE NETWORK COPIES THE
MESSAGE OVER THE NETWORK Problem
Manifold github.com/discordapp/manifold Batching of messages passed between nodes
MANIFOLD WORKS FOR FAN-OUTs WHAT ABOUT FAN-INs? Problem
SO MUCH CHATTER MESSAGE PASSING Problem BECAME EXPENSIVE
MQTT Publisher/Subscriber Messaging Protocol Built on top of TCP/IP Lightweight
and Bandwidth Efficient Quality of Service Data Agnostic Message Queueing Telemetry Transport
Distributed, Scalable and Highly Extensible MQTT message broker Written in
Erlang/OTP http://emqtt.io
E M Q W W W W M /connect /connect
/connect /connect
Problem GLOBAL LOCK ACROSS WORKERS TO HANDLE PLAYER CONNECTIONS
Workers subscribe to /connect/<slot id> Players connect to /connect/<slot id>
Slot = Hash(player_id) T P I O S C SLOTTED
E M Q W W W W M /connect/1 /connect/2
/connect/2 /connect/1
Problem CRASHES APPLICATION
SUPERVISION TREES
Master Worker Worker Worker App Supervisor Master Supervisor Worker Supervisor
CURRENT ARCHITECTURE IS TIGHTLY COUPLED TO QUIZ Problem
HOW TO SUPPORT STATES AND TRANSITIONS FOR DIFFERENT ENGAGEMENTS? Problem
FSM Event State Engagement State gen_statem in Erlang/OTP
WHERE WE ARE
M W W W W Director
WE HAVE IN PLACE BUILDING BLOCKS
HOW TO SCALE IN THE LONG RUN?
WE NEED TO THINK LONG TERM
INTERMISSION
THE ROAD AHEAD
Multi Node
M W W W W M W W W W
GM
libcluster https://github.com/bitwalker/libcluster Form clusters of Erlang nodes Static or Dynamic
Node Membership Custom Strategy to Deal with Nodes Joining/Leaving Notification when Nodes Join/Leave
Resilience
ETS Immortal https://github.com/danielberkompas/immortal Keep ETS alive using a heir process
Keep ETS alive when owner dies Give ownership back to owner after it reboots
NODE A NODE B MASTER MA WA WB MB WA
WA WA WB WB WB MB WB MA WB WB WB WA WA WA WA AND WORKER REPLICA
Masterless
M W W W W GM M W W W
W GM M W W W W GM NODE A NODE C NODE B
CRDT Conflict-free Coordination-free Commutative or Convergent
CRDT Flags Maps Counters Sets
https://kubernetes.io/ SETUP/TEARDOWN ON-DEMAND FOR EVENTS INFRASTRUCTURE
RECAP
• https://elixirschool.com/en/ • https://ferd.ca/the-zen-of-erlang.html • http://blog.plataformatec.com.br/ • http://basho.com/tag/riak/ • https://www.theerlangelist.com/
Resources
Books Litte Elixir and OTP https://www.manning.com/books/the-little-elixir-and-otp-guidebook Learn you some Erlang
http://learnyousomeerlang.com Designing for Scalability with Erlang/OTP http://shop.oreilly.com/product/0636920024149.do
DISTRIBUTED SYSTEMS ARE HARD ELIXIR MAKES THINGS EASIER
Thanks! @kirang89 @yudistrange https://bit.ly/2QVT7CB