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
Notes on Eventual Consistency
Search
UENISHI Kota
December 11, 2014
Technology
4.2k
5
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Notes on Eventual Consistency
結果整合性などの復習
UENISHI Kota
December 11, 2014
More Decks by UENISHI Kota
See All by UENISHI Kota
Storage Systems in Preferred Networks
kuenishi
0
97
Metadata Management in Distributed File Systems
kuenishi
2
570
Behind The Scenes: Cloud Native Storage System for AI
kuenishi
2
460
Apache Ozone behind Simulation and AI Industries
kuenishi
0
480
Distributed Deep Learning with Chainer and Hadoop
kuenishi
3
1.3k
A Few Ways to Accelerate Deep Learning
kuenishi
0
1.2k
Introducing Retz
kuenishi
5
1.3k
Introducing Retz and how to develop practical frameworks
kuenishi
3
830
Formalization and Proof of Distributed Systems (ja)
kuenishi
10
6.5k
Other Decks in Technology
See All in Technology
AIQAのナレッジ構築について
qatonchan
1
140
Claude Mythos、Fable...フロンティアAIの最新動向と企業のセキュリティ対策
flatt_security
0
180
MCPをつなげて作る組織横断のAIエージェント基盤
tsubakimoto_s
0
530
AIで楽になるはずが、なぜ疲れる?
kinopeee
0
150
CloudWatchから始めるAWS監視
butadora
0
310
【5分でわかる】セーフィー エンジニア向け会社紹介
safie_recruit
0
53k
Escolhendo LLMs na Prática: Lições Reais em Busca Agêntica no Mercado Livre —TDC 2026 Floripa
jpbonson
0
120
事業成長とAI活用を止めないデータ基盤アーキテクチャの設計思想
hiracky16
0
790
【CEDEC2026】ゲームシナリオライターを支援するAIツール開発の実践 ― 設計とプロンプトの工夫 ―
cygames
PRO
1
160
AI驚き屋発見器
yama3133
2
390
AIツールを導入しても生産性はあがらない? カオナビが直面した 3つの壁と乗り越え方。/ Overcoming 3 Barriers to AI-Driven Productivity at kaonavi
kaonavi
0
1.3k
BigQuery を検索ソースとした AI Agent の作り方って 〇〇 通りあんねん
satohjohn
0
150
Featured
See All Featured
How to audit for AI Accessibility on your Front & Back End
davetheseo
0
480
Lightning Talk: Beautiful Slides for Beginners
inesmontani
PRO
2
620
技術選定の審美眼(2025年版) / Understanding the Spiral of Technologies 2025 edition
twada
PRO
119
120k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
508
140k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
141
35k
ピンチをチャンスに:未来をつくるプロダクトロードマップ #pmconf2020
aki_iinuma
128
56k
Why You Should Never Use an ORM
jnunemaker
PRO
61
9.9k
GraphQLの誤解/rethinking-graphql
sonatard
75
12k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
49
3.5k
Typedesign – Prime Four
hannesfritz
42
3.1k
We Analyzed 250 Million AI Search Results: Here's What I Found
joshbly
1
1.7k
Redefining SEO in the New Era of Traffic Generation
szymonslowik
1
370
Transcript
݁Ռ߹ੑͳͲͷ෮श 2014/12/11 Ϗοάσʔλج൫ษڧձ @NTTଂ௨ݚ Bashoδϟύϯɹ্
ࣗݾհ • @kuenishi • Github, Twitter, etc • ࢄγεςϜྺ6 •
Bashoδϟύϯͷํ͔Βདྷ·ͨ͠ • Riak CSͷ։ൃ • ͦͷଞຊͷ͜ͱ • msgpack-erlang ϝϯςφ
BashoͱRiak •ࢄσʔλϕʔεʁ •RiakΛ͍ͬͯΔʁ •BashoΛ͍ͬͯΔʁ
͋Β͢͡ •݁Ռ߹ੑڧ߹ੑͷྼԽ൛Ͱͳ͘ɺ ผछͷΛղͨ͘ΊͷҟͳΔఆٛ •ผछͷ is Մ༻ੑ •ηϚϯςΟΫε͕ҟͳΔͷͰΞϓϦͷઃܭ ͷํ͕ͪΐͬͱมΘΔ
݁Ռ߹ੑ͏ݹ͍ʁ •2006ͷٕज़Ͱ͠ΐ •Ϗοάσʔλؔͳ͘Ͷʁ •DynamoDBڧ߹ੑΛఏڙ͍ͯ͠Δ •ωοτϫʔΫΕͳ͍Ͱ͠ΐʁ •ͦΜͳͷͬͯΔਓ͍Δͷʁ •ΞϓϦ͕࡞Γʹ͍͘…
Ϗοάσʔλج൫ݚڀձͱ ݁Ռ߹ੑ •ϏοάσʔλΛѻ͏େنͳγεςϜʹͳ ΕͳΔ΄ͲյΕΔ෦ଟ͍ •͕͔͔͍ۚͬͯΔͷͰɺٻΊΒΕΔՄ༻ੑ ߴ͍ •ӡ༻ָ͕
None
݁Ռ߹ੑͷ࣮༻ྫʢۙʣ •σʔλϕʔεͷόοΫΞοϓ •rsyncͰͷϑΝΠϧͷόοΫΞοϓ •Google Wave aaaa bbbb y x
CAP Theorem • Consistent: ෳͷAtomic Objectʹର͢Δ ࿈ଓͨ͠ૢ࡞ (w1, w3, w4,
….) ͕શͯಉҰ Ͱ͋Δ͜ͱ (linearizable) • Available: Atomic Objectʹૢ࡞ w1, w2, …Λ࣮ߦͯ͠Ϩεϙϯε͕ಘΒΕΔ͜ͱ • Partition Tolerant: ૹͬͨϝοηʔδ͕ ࣦͯ͠ਖ਼͍͠ʢatomicʣͳϨεϙϯε͕ ಘΒΕΔ͜ͱ G1 G2 write read Gilbert and Lynch, Brewer’s Conjecture and the Feasibility of Consistent, Available, Partition-Tolerant Web Services
CAPఆཧ͕ఆٛ͢Δ߹ੑ •CAPఆཧ͕ఆٛ͢Δ߹ੑ㲈Linearizability •શͯͷෳͰɺશͯͷʢߋ৽ʣૢ࡞͕ॱ൪௨Γ ϦϓϨΠ͞ΕΔ͜ͱΛอো͢Δ •ACIDͷͦΕͱͪΐͬͱҧ͏
ͳͥڧ߹ੑͷ࣮ݱ͕͍͠ͷ͔ • ଟܾͱ͔ atomic broadcast Λ͏ͱͯ͠ύ ϑΥʔϚϯεͷϖφϧςΟ͕͋Δ • asynchrony +
partial failureͷ͠͞ • ࢮ׆ࢹ is hard => Downtime • ӡ༻ੑ • Γସ͑ɺΓ͠ɺ༳Ε·͢༳Ε·͢
Consistency͍͠ •ߋ৽ΛࢭΊΔʢAvailabilityΛԼ͛Δʣ͔ɺߋ৽ͷ্ॻ͖Λ ڐ͢ʢσʔλΛࣦ͏ʣ͔͔͠બࢶ͕ͳ͍ Server2 Server1 Server3 PUT V=42 PUT V=0
V=?
Atomic Broadcasting is Difficult • ϨϓϦέʔγϣϯॱ൪͕ೖΕସΘΔ • CPUͷΞτΦϒΦʔμʔ࣮ߦͱಉ͡ w1 w1
w1 w2 w2 w2 Actor 0 Actor 1 Actor 2 w2 w2 w1
Consensus Based Replication • ϨϓϦέʔγϣϯͷϦʔμʔΛଟܾͰબग़ • or ϨϓϦέʔγϣϯຖʹଟܾ w1 w1
w1 w2 w2 w2 Actor 0 Actor 1 Actor 2 w2 w2 w1
݁Ռ߹ੑ •Eventual Consistency •Ͳ͏͍͏ܦ࿏ΛḷΔʹͤΑɺෳ͕ ࠷ऴతʹಉ͡ঢ়ଶʹऩଋ͢Δ͜ͱ •Read Repair •AAE •CRDT v0
v1 •(Vector Clocks)
Siblings •ͱΓ͋͑ͣෳͷόʔδϣϯͷڞଘΛڐ͢ •Ͳͷόʔδϣϯ͕ਖ਼͍͔͠ɺ͘͠Ϛʔδ͢Δ͔ΛRead࣌ʹܾఆ Server2 Server1 Server3 PUT V=42 PUT V=0
V=0 or 42 V=0 V=0 or 42 V=42
APΛ࣮ݱ •ωοτϫʔΫஅ͕ى͖͍ͯͯͱΓ͋͑ͣॻ͖ࠐΈΛڐ͢ Server2 Server1 Server3 PUT V=42 PUT V=0 Server4
෮چͨ͠Βॻ͖͢ ྆ํ͓࣋ͬͯ͘
γϣοϐϯάΧʔτͷྫ •UnionΛͱΕΑ͍ Server2 Server1 Server3 PUT cart=[a,b,d] PUT cart=[a,b,c] union([a,b,c],
[a,b,d]) => [a,b,c,d] [a,b,c] [a,b,c] or [a,b,d] [a,b,d]
Read Repair v2 v2 get(“conferences/thoughtworks”) Get Handler (FSM) client Riak
Coordinating node Cluster 6 7 8 9 10 11 12 13 14 15 16 R=2 v1 v2 v2 v1 v2 v1 v1 v2 v2
Active Anti Entropy • APࢦͷDBͷσʔλྼԽΛ͙ ͨΊͷόοΫάϥϯυॲཧ • Merkle-TreeΛͬͯύʔςΟγϣ ϯຖͷʮνΣοΫαϜʯΛܭࢉ •
ࠩΛݟ͚ͭͨΒͦ͜ΛRead Repair͢Δ hash(vnode=0, pid=0) hash(vnode=1, pid=0) hash(vnode=2, pid=0)
CRDT • ॱ൪͕ೖΕସΘͬͯ݁Ռ͕มΘΒͳ͍ܕ • update(w1, update(w2, Data0) = update(w2, update(w1,
Data0) = Data w1 w1 w1 w2 w2 w2 Actor 0 Actor 1 Actor 2 w1(w2(Data0)) => Data w1(w2(Data0)) => Data w2(w1(Data0)) => Data
CRDT: PN-Counter • merge • {a: {1,-1}, b: {1,0}, c:
{2,0}} • {a: {0,0}, b: {2, 0}, c: {0, -2}} • => {a: {1,-1}, b:{2,0}, c:{2,-2}} => 2 • update • a͕ {increment, 3} Λड͚͚Δͱ • {a: {4,-1}, b: {1,0}, c: {2,0}}
CRDT: OR-Sets • merge • {a:{“foo”:true}, b:{“bar”:false}} • + {a:{“foo”:true},
b:{“foo”:false, “bar”:false}} • => {a:{“foo”:true}, b:{“foo”:false, “bar”:true}} • => [“bar”] • update • add: {a:{}} => +”foo” => {a:{“foo”:false}} • remove: {a: {“foo”:false}} => {a: {“foo”:true}}
ӡ༻ָ͕ • ΧδϡΞϧʹϊʔυωοτϫʔΫΛ্͛Լ͛Ͱ͖Δ • ߹ੑΛอͭͨΊͷϚελʔ͕୭͔Λؾʹ͢Δඞཁ͕ͳ͍ • ڧ߹ੑΛอͭͨΊͷΦϖϨʔγϣϯ͕μϯλΠϜʹͳ Βͳ͍ • ߹ੑνΣοΫɺϦΧόϦɺόοΫΞοϓ
• ނো࣌ͷΦϖϨʔγϣϯ͕͔ͳΓ୯७
݁Ռ߹ੑΛ࠾༻ͨ͠ ߹ͷ՝ •;ͭ͏ͷϓϩάϥϛϯάͱҟͳΔηϚϯςΟ ΫεʹͳΔ •ΞϓϦέʔγϣϯʹ͜Ε·ͰͱҟͳΔલఏΛཁ ٻ͢Δ͜ͱʹͳΔ •CRDTͰҰ෦ղܾɺ͚ͩͲ…
Ԡ༻ྫ
League of Legends •MMORPGͷνϟοτՄ༻ੑͱϨε ϙϯελΠϜ໋͕ •10ms ͕ੜࢮΛ͚Δ (C) Riot Games
•Riak্Ͱಈ͘ “ߴՄ༻” Ϋϥυ ετϨʔδ •ΦϒδΣΫτͷϝλσʔλ݁Ռ ߹తσʔλߏ •: 5GBͷσʔλ͕Concurrentʹ Ξοϓϩʔυ͞Ε͖ͯͨΒʁ •:
͔ͦ͠Ε͕ผͷେͷ ΞοϓϩʔυͩͬͨΒʁ /foo.bar
σʔληϯλʔؒϨϓϦέʔγϣϯ •DCؒωοτϫʔΫଓੑ ଳҬ·ͰؚΊͯৗʹਖ਼͘͠ӡ ༻͢Δͷ͕͍͠ •CAPఆཧͷཁ͔Βɺಉظత ϨϓϦέʔγϣϯ͍͠ •Մ༻ੑΛอͭͨΊʹɺ݁Ռ ߹͢ΔσʔλϞσϧΛ࠾༻
·ͱΊ •݁Ռ߹ੑڧ߹ੑͷྼԽ൛Ͱͳ͘ɺผछͷ Λղͨ͘ΊͷҟͳΔఆٛ •ผछͷ is Մ༻ੑ •ηϚϯςΟΫε͕ҟͳΔͷͰΞϓϦͷઃܭͷํ͕ ͪΐͬͱมΘΔ •݁Ռ߹ੑΛอͭͨΊͷ͍͔ͭ͘ͷٕज़Λհ
We are hiring. •࣮ੈքͷࢄγεςϜͷ ʹڵຯ͋Δਓʂ •@BashoJapan •
[email protected]
Questions?