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
ABEJA Platform での MLOps LINE×ABEJA MLOps Study ...
Search
Yusuke Ueno
April 24, 2019
Technology
0
730
ABEJA Platform での MLOps LINE×ABEJA MLOps Study @FUKUOKA
Yusuke Ueno
April 24, 2019
Tweet
Share
Other Decks in Technology
See All in Technology
Pure Goで体験するWasmの未来
askua
1
150
Oracle Base Database Service 技術詳細
oracle4engineer
PRO
11
77k
VCC 2025 Write-up
bata_24
0
130
API提供者のためのMCPサーバー設計ガイド / MCP Server Design Guide for API Providers
yokawasa
0
230
Oracle Cloud Infrastructure:2025年9月度サービス・アップデート
oracle4engineer
PRO
0
300
analysis パッケージの仕組みの上でMulti linter with configを実現する / Go Conference 2025
k1low
1
240
pprof vs runtime/trace (FlightRecorder)
task4233
0
130
Goに育てられ開発者向けセキュリティ事業を立ち上げた僕が今向き合う、AI × セキュリティの最前線 / Go Conference 2025
flatt_security
0
250
あなたのWebサービスはAIに自動テストしてもらえる?アクセシビリティツリーで読み解く、AIの『視点』
yusukeiwaki
1
3.3k
PLaMo2シリーズのvLLM実装 / PFN LLM セミナー
pfn
PRO
2
550
PyCon JP 2025 DAY1 「Hello, satellite data! ~Pythonではじめる衛星データ解析~」
ra0kley
0
710
#普通の文系サラリーマンチャレンジ 自分でアプリ開発と電子工作を続けたら人生が変わった
tatsuya1970
0
730
Featured
See All Featured
Designing Experiences People Love
moore
142
24k
What's in a price? How to price your products and services
michaelherold
246
12k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
49
3.1k
Embracing the Ebb and Flow
colly
88
4.8k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
285
14k
Why You Should Never Use an ORM
jnunemaker
PRO
59
9.5k
GraphQLとの向き合い方2022年版
quramy
49
14k
The Cult of Friendly URLs
andyhume
79
6.6k
Site-Speed That Sticks
csswizardry
10
860
ReactJS: Keep Simple. Everything can be a component!
pedronauck
667
120k
KATA
mclloyd
32
14k
The Language of Interfaces
destraynor
162
25k
Transcript
Software Engineer at ABEJA Yusuke Ueno ABEJA Platform Ͱͷ ML
Ops
ࠓ͢͜ͱ • ABEJA Platform ͱʁ • ػցֶशͷ࣮ݧཧʹ͍ͭͯ • ABEJA Platform
Ͱͷ࣮ݧཧͱͦͷ࣮
ABEJA Platform ͱʁ
Copyright © 2019 ABEJA, Inc. All rights reserved.
None
Copyright © 2019 ABEJA, Inc. All rights reserved. نײ
Copyright © 2019 ABEJA, Inc. All rights reserved. ML Ops
ͱ? DevOps ͜ͷΑ͏ͳҹ • Development ͱ Operation ؒͷϓϩηεվળ • ΞϓϦέʔγϣϯͷσϦόϦೳྗΛ͋͛ΔจԽతֶɺ ϓϥΫςΟεɺπʔϧ
Copyright © 2019 ABEJA, Inc. All rights reserved. ML Ops
͜͏ఆٛͯ͠Έ·͢ • ML Engineer ͱ Development ؒͷϓϩηεվળ • Ϗδωεʹద༻Ͱ͖Δਫ਼ΛͭϞσϧΛఏڙ͢Δೳྗ Λ্͛ΔจԽతֶɺϓϥΫςΟεɺπʔϧ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࠓֶश෦ʹ͍ͭͯ
Copyright © 2019 ABEJA, Inc. All rights reserved. ֶश ΠςϨʔςΟϒͳ࡞ۀ
• ֶशίʔυͷ࡞ɾमਖ਼ • ҟͳΔΦϓςΟϚΠβͰͷࢼߦ • ϋΠύʔύϥϝʔλͷௐ • αϯϓϦϯάํ๏ͷमਖ਼ • ҟͳΔόʔδϣϯͷϥΠϒϥϦͷ༻ • ϥϯμϜγʔυͷมߋ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧͷཧ͕ॏཁ ҰճҰճͷ࣮ݧͷ݅ͱ݁ՌΛه͍ͯ͠ͳ͍ͱɺޙͰਫ਼
͕ྑ͔ͬͨ࣌ͷ࣮ݧΛ࠶ݱͰ͖ͳ͍ هͯ͠ɺӾཡͰ͖ΔΑ͏ʹ͓ͯ͘͠ඞཁ͕͋Δ
• σʔληοτ • ίʔυ • ύϥϝʔλ • ࣮ߦڥ • ࣮ݧ݁ՌʢධՁࢦඪʣ
• ॏΈύϥϝʔλ • ϩά • ࣮ߦ࣌ؒ ه • ࣮ݧ݁Ռͷൺֱ • ৄࡉใͷදࣔ • ࣮ݧ݅ • ࣮ݧ݁Ռ • ՄࢹԽʢը૾ͳͲʣ • ϝϯόʔؒͰͷڞ༗ • աڈͷ࣮ݧͷݕࡧ • Ӿཡ Ӿཡ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ ϝϯόʔؒͰͷڞ༗ όʔδϣϯཧ ՄࢹԽ ֶशδϣϒؒͰͷൺֱ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. σʔληοτͷόʔδϣϯཧ ̎ͭͷίϯϙʔωϯτΛ༻ҙ
• Datalake • ΦϒδΣΫτετϨʔδ • Datasets • Datalake ΦϒδΣΫτͷࢀরใͱϝλσʔλ
Copyright © 2019 ABEJA, Inc. All rights reserved. σʔληοτͷόʔδϣϯཧ •
Annotation Tool ʹͯ Datalake ͷσʔλʹରͯ͠Ξϊςʔ γϣϯͨ݁͠ՌΛ Datasets ͱͯ͠ग़ྗ %BUBMBLF %BUBTFUT
Copyright © 2019 ABEJA, Inc. All rights reserved. σʔληοτͷόʔδϣϯཧ σʔλΛՃͨ͠߹ɺผͷ
datasets ͱͯ͠࡞Մೳ \^ \^ \^ ɾɾɾ GJMFT BOOPUBUJPOT EBUBTFUT WFSTJPO \^ \^ WFSTJPO
Copyright © 2019 ABEJA, Inc. All rights reserved. σʔληοτͷόʔδϣϯཧ tag
Ͱ datasets Λཧతʹׂ͠ಛఆͷཁૉͷΈΛநग़ ɾɾɾ EBUBTFUT UBH" UBH# \^ " \^ " \^ " \^ # \^ #
Copyright © 2019 ABEJA, Inc. All rights reserved. σʔληοτͷՄࢹԽ σʔληοτࣗମͷ֬ೝ͕Մೳ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ߦڥ Platform
Ͱ Python RuntimeɺओཁͳϑϨʔϜϫʔΫɺϥΠ ϒϥϦશ෦ೖΓͷ Docker Image Λఏڙ
Copyright © 2019 ABEJA, Inc. All rights reserved. ֶशίʔυɾύϥϝʔλ •
ֶशΛ࣮ߦ͢Δ Python ίʔυ • Platform ্Ͱݺͼग़͞ΕΔؔΛ࣮ • Docker Image ʹඞཁͳ Python ϥΠϒϥϦ͕ͳ͍߹ʹ requirements.txt ʹՃ • ༩͑ͨύϥϝʔλڥมͱͯ͠ίʔυͰऔಘՄೳ
Copyright © 2019 ABEJA, Inc. All rights reserved. ༻͢Δσʔληοτɺֶशίʔυɺύϥϝʔλɺ࣮ߦ ڥΛ·ͱΊͯɺ࣮ߦͰ͖Δঢ়ଶͰόʔδϣχϯάͯ͠ཧ
ֶशδϣϒఆٛόʔδϣϯ ֶशίʔυ { } ύϥϝʔλ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ֶशδϣϒఆٛόʔδϣϯͱύϥϝʔλɺΠϯελϯελ ΠϓΛࢦఆֶͯ͠शδϣϒΛ࣮ߦ
ֶशδϣϒ࣮ߦ ֶशίʔυ { } ύϥϝʔλ ֶशδϣϒఆٛόʔδϣϯ { } ্ॻ͖ύϥϝʔλ ֶशδϣϒ σʔληοτ ΠϯελϯελΠϓ ʴ ه ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ֶशδϣϒͷ࣮ߦͱ݁Ռͷཧ •
kubernetes ( EKS ) Λ༻ • Ҏલ kubernetes on EC2 • nvidia-device-plugin Λ༻ͯ͠ GPU Λೝࣝ • spotinst ͰΫϥελΦʔτεέʔϦϯά • ָʹෳͷΠϯελϯεͰͷεέʔϧ͕Մೳ • p2 ܥɺ p3 ܥΠϯελϯε
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ֶशδϣϒ •
k8s ͷ Job ͱֶͯ͠शίʔυʹύϥϝʔλΛ༩࣮͑ͯߦ • SDK Λ༻ͯ͠ɺΤϙοΫ͝ͱͷਫ਼Λߋ৽ ΠϯελϯελΠϓ ࣮ߦڥ 4%, ਫ਼Λอଘ ֶशίʔυ { } ύϥϝʔλ ධՁ݁Ռ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ཧܥίϯςφ ֶशδϣϒͱಉ͡ϊʔυʹஔ͠ɺग़ྗͱͳΔͷΛอଘ
&'4Ͱͷڞ༗ϑΝΠϧετϨʔδ ֶशδϣϒ "HFOU 5FOTPS#PBSE 'MVFOUE Ϛϯτ εςʔλεࢹ ग़ྗϑΝΠϧอଘ ެ։ ϩάΛऔಘ อଘ
Copyright © 2019 ABEJA, Inc. All rights reserved. Fluentd ίϯςφ
ֶशδϣϒ͕ग़ྗ͢Δඪ४ग़ྗΛอଘ • k8s ͷ DaemonSet ͰίϯςφΛஔ • શͯͷϊʔυʹ̍ͭͷ Fluentd ίϯςφΛ࣮ߦ • جຊతʹ /var/log/containers/*.log Λࢹͯ͠ɺ͜ΕΒ ͷϩάΛ֎෦ͷετϨʔδʹอଘ • Pod ͕ফ͑Δͱϩάফ͑ͯ͠·͏
Copyright © 2019 ABEJA, Inc. All rights reserved. Fluentd ίϯςφ
• RUBY_GC_HEAP_OLDOBJECT_LIMIT_FACTOR ͷઃ ఆ࣍ୈͰɺNoisy Neighbor ʹͳΔ͔ɺResource Limit ʹΑΓ OOM Killer Ͱࡴ͞Εͯ͠·͏
Copyright © 2019 ABEJA, Inc. All rights reserved. TensorBoard ίϯςφ
ֶशδϣϒ͕ग़ྗ͢ΔΠϕϯτϩάͷՄࢹԽ • Inter-Pod Affinity Λ༻ͯ͠ Job ͱಉ͡ϊʔυʹஔ • Job ͱಉ͡ϑΝΠϧγεςϜΛϚϯτ͠ɺϩάΛಡΈ ࠐΈදࣔ • k8s ͷ Service ͷ Node Port Ͱ internal ʹ expose ͠ɺ ͷ Gateway ͕ೝূ͖Ͱެ։
Copyright © 2019 ABEJA, Inc. All rights reserved. Agent ίϯςφ
ֶशδϣϒͷεςʔλεࢹɾ։࢝ / ऴྃ࣌ࠁΛه • Job ͷεςʔλεΛϙʔϦϯάͯ͠ه • Job ͱಉ͡ϑΝΠϧγεςϜΛϚϯτ͠ɺֶशδϣϒ ͷऴྃͱͱʹग़ྗϑΝΠϧΛอଘ ֶशδϣϒ "HFOU εςʔλεࢹɾߋ৽ ग़ྗϑΝΠϧอଘ
Copyright © 2019 ABEJA, Inc. All rights reserved. ࣮ݧཧͷશମ૾ {
} ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved.
Copyright © 2019 ABEJA, Inc. All rights reserved. ML Ops
• ML Engineer ͱ Development ؒͷϓϩηεվળ • Ϗδωεʹద༻Ͱ͖Δਫ਼ΛͭϞσϧΛఏڙ͢Δೳྗ Λ্͛ΔจԽతֶɺϓϥΫςΟεɺπʔϧ
Copyright © 2019 ABEJA, Inc. All rights reserved. ML Engineer
ͱ Development ؒͷϓϩηεվળ ཁٻΛຬͨ͢Ϟσϧ͕Ͱ͖ΔͱଞͷαʔϏε͕ར༻Α͏ʹެ։ • ୭͕ຊ൪͚ͷίʔυΛॻ͔͘ʁ • Data Scientist ͕ॻ͍ͨίʔυΛॻ͖͞ͳ͍ͱ͍͚ͳ͍ • ॻ͖͢ͱਫ਼͕࠶ݱ͠ͳ͍… • Ϟσϧͷߋ৽͕ଟ͗͢ → αʔϏεͷߋ৽ճ૿Ճ • ʑ…
Copyright © 2019 ABEJA, Inc. All rights reserved. ML Engineer
ͱ Development ؒͷϓϩηεվળ Development ଆֶश݁ՌͱਪίʔυͱΈ߹Θͤͯ όʔδϣϯཧՄೳ ਪίʔυ ֶश݁Ռ ॏΈϑΝΠϧ ࣮ߦڥ ධՁ݁Ռ ॏΈϑΝΠϧ ࣮ߦڥ ධՁ݁Ռ δϣϒ̍ δϣϒ̎ ॏΈϑΝΠϧ ࣮ߦڥ Ϟσϧ
Copyright © 2019 ABEJA, Inc. All rights reserved. ML Engineer
ͱ Development ؒͷϓϩηεվળ Ϟσϧͦͷ·· Web API ͱͯ͠ެ։Մೳ Ϟσϧߋ৽࣌ Web API Λ҆શʹߋ৽Մೳ ਪίʔυ Ϟσϧ ॏΈϑΝΠϧ ࣮ߦڥ ॏΈϑΝΠϧ ࣮ߦڥ Ϟσϧ̍ Ϟσϧ̎ ਪίʔυ 8FC"1* 8FC"1* σϓϩΠ ΤϯυϙΠϯτ Γସ͑Մೳ
Copyright © 2019 ABEJA, Inc. All rights reserved. Platform ͰͷϞσϧཧશମ
{ } ֶशίʔυ ύϥϝʔλ ධՁ݁Ռ ॏΈϑΝΠϧ ϩά ࣮ߦ࣌ؒ ֶशδϣϒ σʔληοτ ࣮ߦڥ ਪίʔυ ॏΈϑΝΠϧ ࣮ߦڥ
Copyright © 2019 ABEJA, Inc. All rights reserved. ·ͱΊ •
࣮ݧཧ໘͕ͩɺΒͳ͍ͱޙͰࠔΔ • ֶशͷೖྗͱͳΔ༻͢Δσʔληοτɺֶशίʔυɺ ࣮ߦڥͳͲΛ·ͱΊͯόʔδϣϯཧ • ग़ྗ݁Ռͷอଘग़དྷΔ͚ͩ։ൃऀʹෛ୲Λ͔͚ͳ͍ܗ Ͱ Platform ଆͰ࣮ • αʔϏεԽ͢ΔϞσϧͱֶशδϣϒͷ݁Ռͷඥ͚ͯτ ϨʔαϏϦςΟΛ୲อ