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
メルカリのマーケット健全化施策を支えるML基盤
Search
Hirofumi Nakagawa/中河 宏文
May 23, 2018
Programming
10
9.1k
メルカリのマーケット健全化施策を支えるML基盤
Hirofumi Nakagawa/中河 宏文
May 23, 2018
Tweet
Share
More Decks by Hirofumi Nakagawa/中河 宏文
See All by Hirofumi Nakagawa/中河 宏文
IoTデバイスでMLモデルを動かす技術
hnakagawa
0
170
Kanazawa_AI.pdf
hnakagawa
0
190
メルカリ写真検索における Amazon EKS の活用事例と プロダクトにおけるEdgeAI technologyの展望
hnakagawa
5
9k
メルカリの写真検索を支えるバックエンド CCSE 2019 version
hnakagawa
0
320
メルカリ写真検索における Amazon EKS の活用事例
hnakagawa
6
29k
メルカリの写真検索を支えるバックエンド
hnakagawa
1
1.2k
Mercari ML Platform
hnakagawa
1
17k
mlct.pdf
hnakagawa
2
2.1k
機械学習によるマーケット健全化施策を支える技術
hnakagawa
0
240
Other Decks in Programming
See All in Programming
マイコンでもRustのtestがしたい その2/KernelVM Tokyo 18
tnishinaga
2
2.3k
AIレビュアーをスケールさせるには / Scaling AI Reviewers
technuma
2
200
KessokuでDIでもgoroutineを活用する / Go Connect #6
mazrean
0
100
Constant integer division faster than compiler-generated code
herumi
2
670
MCP連携で加速するAI駆動開発/mcp integration accelerates ai-driven-development
bpstudy
0
310
Nuances on Kubernetes - RubyConf Taiwan 2025
envek
0
180
バイブコーディング × 設計思考
nogu66
0
120
Scale out your Claude Code ~自社専用Agentで10xする開発プロセス~
yukukotani
9
2.3k
AI OCR API on Lambdaを Datadogで可視化してみた
nealle
0
120
A Gopher's Guide to Vibe Coding
danicat
0
170
The State of Fluid (2025)
s2b
0
180
Flutterと Vibe Coding で個人開発!
hyshu
1
250
Featured
See All Featured
Optimising Largest Contentful Paint
csswizardry
37
3.4k
StorybookのUI Testing Handbookを読んだ
zakiyama
30
6k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
50
5.5k
Principles of Awesome APIs and How to Build Them.
keavy
126
17k
Product Roadmaps are Hard
iamctodd
PRO
54
11k
Statistics for Hackers
jakevdp
799
220k
Agile that works and the tools we love
rasmusluckow
329
21k
The Invisible Side of Design
smashingmag
301
51k
Chrome DevTools: State of the Union 2024 - Debugging React & Beyond
addyosmani
7
810
Building a Scalable Design System with Sketch
lauravandoore
462
33k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
131
19k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
283
13k
Transcript
ϝϧΧϦͷϚʔέοτ݈શԽ ࢪࡦΛࢧ͑ΔMLج൫ Mercari ML Ops Night Vol.1 hnakagawa
ࣗݾհ • Hirofumi Nakagawa (hnakagawa) • 20177݄ೖࣾ • ॴଐSRE •
σόΠευϥΠό։ൃ͔Βϑϩϯ τΤϯυ։ൃ·ͰΔԿͰ • NOT MLΤϯδχΞ • https://github.com/hnakagawa
͓ࣄ • ML Platform։ൃ • MLΤϯδχΞͱSREͷεΩϧΪϟοϓΛຒΊ Δ • ML Reliability,
SysML?, MLOps? • SREͷཱ͔ΒMLγεςϜͷࣗಈԽΛߦ͏
ML Platform • ͷML Platform • kubernetesϕʔε • ϩʔΧϧڥͱΫϥελڥͷ ࠩΛநԽ͢Δ
• ศརAPI܈ • طଘͷML FrameworkΛ༻͠ ؆୯ʹTraining/ServingΛߦ͏ ڥΛఏڙ
ͦͷ͏ͪOSSͰެ։༧ఆ(ଟ
ࣄྫ ϦΞϧλΠϜࢹγεςϜ • ௨শ Lovemachine • ML Platform্ʹ࣮͞Ε͍ͯΔ .-1MBUGPSN USBJOJOHDMVTUFS
-PWFNBDIJOF ($4 GKE PubSub .-1MBUGPSN TFSWJOHDMVTUFS -PWFNBDIJOF
Model Training & Serving Workflow
.-1MBUGPSN USBJOJOHDMVTUFS Workflow for Production $* .-1MBUGPSN TFSWJOHDMVTUFSGPSUFTU .PEFM3FHJTUSZ +PC
+PC ɾɾ 3&45 "1* 4USFBNJOH 5' 4FSWJOH ɾɾɾ
.-1MBUGPSN USBJOJOHDMVTUFS Training Workflow $* .PEFM3FHJTUSZ +PC +PC ɾɾɾ 1.
GitHubͷpushΛτϦΨʹtrainingΛىಈ 2. Training͞ΕͨModelModel Registry ্͕Δ
Serving Workflow .-1MBUGPSN TFSWJOHDMVTUFSGPSUFTU .PEFM3FHJTUSZ ɾɾ 3&45 "1* 4USFBNJOH 5'
4FSWJOH ɾɾɾ 1. Model RegistryΛࢹͯࣗ͠ಈͰModel ΛServing 2. Serving&Test͕ޭ͢Δͱຊ൪༻k8s manifestΛग़ྗ
Model Serving APIͷߏྫ 5FOTPS'MPX 4FSWJOH 5' .PEFM 5' .PEFM 'MBTL
4, .PEFM 4, .PEFM 4, .PEFM gRPC .FSDBSJ"1* REST FlaskͰલॲཧΛߦ͍ ཪͷTensorFlow Servingʹ͍͛ͯΔ
Model Serving API Streaming ver ͷߏྫ 5FOTPS'MPX 4FSWJOH 5' .PEFM
5' .PEFM .-1MBUGPSN 'SBNFXPSL PS "QBDIF#FBN 4, .PEFM 4, .PEFM 4, .PEFM gRPC PubSub
TensorFlow Serving • TensorFlow project͕ఏڙͯ͠ ͍ΔServingڥ • PythonॲཧܥΛհͣ͞ʹTFͷ modelΛservingͰ͖Δ •
ඪ४ͷ࣮ͰgRPCͰAPIΛ ఏڙ
ModelͱίϯςφɾΠϝʔδ • ڊେͳML ModelΛίϯςφɾΠϝʔδʹؚΊ Δ͔൱͔ • ؚΊͳ͍ͷͰ͋ΕԿॲʹஔ͢Δ͔ • ϙʔλϏϦςΟੑͱϩʔυ࣌ؒͷτϨʔυΦϑ •
ྑ͍ΞΠσΟΞ͕͋Εڭ͑ͯԼ͍͞…
௨ৗͷAPIͱҧ͏ • ѻ͏ϦιʔεɺModelαΠζ͕େ͖͘ͳΔ ߹͕ଟ͍(ඦMBʙGB) • CPUɾϝϞϦϦιʔεͷফඅ͕ܹ͍͠ • ߹ʹΑͬͯGPU͏
ϝϞϦফඅ • LovemachineͷPython࣮෦࣮ߦ࣌ʹ 2GBϝϞϦΛফඅ͢Δˠࠓޙ͞Βʹ૿͑Δ༧ ఆ͋Δ • Scikit-learnͰهड़͞ΕͨTF-IDFͷલॲཧ෦ ͕େ͖͘ͳΔࣄ͕ଟ͍
Pythonͱฒྻੑ • વThread͕͑ͳ͍(GILͷͨΊ) • ϓϩηεຖʹModelΛϩʔυ͢Δͱඞཁͳϝ ϞϦαΠζ͕େ͖͘ͳΔˠ Blue-Green DeployͷোʹͳΔ
ਖ਼PythonͰͷServing Πϯϑϥతʹਏ͍ࣄ͕ଟ͍…
ϝϞϦΛݡ͘͏ • fork͢ΔલʹmodelΛϩʔυ͠Copy on Write Λޮ͔͢ • k8sͷone process per
containerηΦϦ͋ ͑ͯഁ͍ͬͯΔ
Copy On Writeͷ෮श ϝϞϦ ϓϩηε ࢠϓϩηε 2.fork 1BHF" 1.allocation ಉ͡ྖҬΛࢀর
ϓϩηε͕ϝϞϦͷ༰Λ ॻ͖͑Δͱ… ϝϞϦ ϓϩηε ࢠϓϩηε 1BHF" 1BHF# OS͕ผͷྖҬΛAllocationͯ͠ݩσʔλΛίϐʔ͢Δ ผͷྖҬΛࢀর
Current Issues • ਓؒͷߦಈΛ૬खʹ͍ͯ͠Δҝɺσʔλͷ ͕มΘΓ͔ͬͨ͢Γɺ༧֎ͷ͕ൃ ੜͨ͠Γͯ͠ɺରԠ͠ଓ͚Δඞཁ͕͋Δ ˠ ML Model࡞ऀʹෛ୲ֻ͕͔Γଓ͚Δ ˠ
SREͱͯࣗ͠ಈԽΛؚΜͩΈͰղܾ ͍ͨ͠
In Progress • ࣾͷσʔλ͔ΒEmbedding͢Δ࣮Λίϯ ϙʔωϯτԽ • ಛఆͷΛղܾ͢ΔϞσϧߏஙΛ͋Δఔ ࣗಈԽ ˠࣾͷղܾʹಛԽͨ͠ઐ༻ͷAutoMLత ͳԿ͔
AutoFlow(Ծ) 'FBUVSF&YUSBDUJPO $PNQPOFOUT $MBTTJpDBUJPO $PNQPOFOUT $PODBUFOBUJPO $PNQPOFOUT .PEFM #VJMEFS $PNQPOFOUT
3FHJTUSZ Ϋϥελ্ͰϞσϧͷࣗಈߏஙͱϋΠύʔύϥ ϝʔλͷࣗಈௐΛߦ͏
·ͱΊ • MLʹগ͠௨ৗͱҧ͏Πϯϑϥ͕ඞཁʹͳΔ ˠ·ͩϕετɾϓϥΫςΟε͔Βͳ͍ • ͦͦMLͳػೳΛຊ֨ӡ༻͠Α͏ͱ͢Δ ͱɺେ෯ͳࣗಈԽɾΈԽΛਐΊͳ͍ͱ্ ख͘ߦ͔ͳ͍
͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠!!
We are Hiring!!
SRE ML Reliability • SysML? MLOps? ৽͍͠Job description • SREεΩϧ+MLͷجૅࣝ
• MLΠϯϑϥͷࣗಈԽɾΈԽΛਪ͠ਐΊͯ ͘ΕΔਓࡐ • ͪΖΜଞͷ৬छઈࢍืूத!!
ৄࡉͪ͜Β https://careers.mercari.com/