メルカリのマーケット健全化施策を支えるML基盤

Transcript

1. ϝϧΧϦͷϚʔέοτ݈શԽ ࢪࡦΛࢧ͑ΔMLج൫ Mercari ML Ops Night Vol.1
 
 hnakagawa

2. ࣗݾ঺հ • Hirofumi Nakagawa (hnakagawa) • 2017೥7݄ೖࣾ • ॴଐ͸SRE • σόΠευϥΠό։ൃ͔Βϑϩϯ τΤϯυ։ൃ·Ͱ΍ΔԿͰ΋԰ • NOT MLΤϯδχΞ • https://github.com/hnakagawa
3. ͓࢓ࣄ • ML Platform։ൃ • MLΤϯδχΞͱSREͷεΩϧΪϟοϓΛຒΊ Δ • ML Reliability, SysML?, MLOps? • SREͷཱ৔͔ΒMLγεςϜͷࣗಈԽΛߦ͏
4. ML Platform • ಺੡ͷML Platform • kubernetesϕʔε • ϩʔΧϧ؀ڥͱΫϥελ؀ڥͷ ࠩΛந৅Խ͢Δ • ศརAPI܈ • طଘͷML FrameworkΛ࢖༻͠ ؆୯ʹTraining/ServingΛߦ͏ ؀ڥΛఏڙ
5. ͦͷ͏ͪOSSͰެ։༧ఆ(ଟ෼
6. ࣄྫ ϦΞϧλΠϜ঎඼؂ࢹγεςϜ • ௨শ Lovemachine • ML Platform্ʹ࣮૷͞Ε͍ͯΔ .-
   1MBUGPSN
   USBJOJOH
   DMVTUFS -PWFNBDIJOF ($4 GKE PubSub .-
   1MBUGPSN
   TFSWJOH
   DMVTUFS -PWFNBDIJOF
7. Model Training & Serving
 Workflow
8. .-
   1MBUGPSN
   USBJOJOH
   DMVTUFS Workflow for Production $* .-
   1MBUGPSN
   TFSWJOH
   DMVTUFS
   GPS
   UFTU .PEFM
   3FHJTUSZ +PC +PC ɾɾ 3&45
 "1* 4USFBNJOH 5'
   4FSWJOH
 ɾɾɾ
9. .-
   1MBUGPSN
   USBJOJOH
   DMVTUFS Training Workflow $* .PEFM
   3FHJTUSZ +PC +PC ɾɾɾ 1. GitHub΁ͷpushΛτϦΨʹtrainingΛىಈ 2. Training͞ΕͨModel͸Model Registry
 ΁্͕Δ
10. Serving Workflow .-
    1MBUGPSN
    TFSWJOH
    DMVTUFS
    GPS
    UFTU .PEFM
    3FHJTUSZ ɾɾ 3&45
 "1* 4USFBNJOH 5'
    4FSWJOH
 ɾɾɾ 1. Model RegistryΛ؂ࢹͯࣗ͠ಈͰModel ΛServing 2. Serving&Test͕੒ޭ͢Δͱຊ൪༻k8s manifestΛग़ྗ
11. Model Serving APIͷߏ੒ྫ 5FOTPS'MPX
 4FSWJOH 5'
    .PEFM 5'
    .PEFM 'MBTL 4,
 .PEFM 4,
 .PEFM 4,
 .PEFM gRPC .FSDBSJ
    "1* REST FlaskͰલॲཧΛߦ͍
 ཪͷTensorFlow Servingʹ౤͍͛ͯΔ
12. Model Serving API
 Streaming ver ͷߏ੒ྫ 5FOTPS'MPX
 4FSWJOH 5'
    .PEFM 5'
    .PEFM .-
    1MBUGPSN
    'SBNFXPSL
    
 PS
 "QBDIF
    #FBN
 4,
 .PEFM 4,
 .PEFM 4,
 .PEFM gRPC PubSub
13. TensorFlow Serving • TensorFlow project͕ఏڙͯ͠ ͍ΔServing؀ڥ • PythonॲཧܥΛհͣ͞ʹTFͷ modelΛservingͰ͖Δ • ඪ४ͷ࣮૷Ͱ͸gRPCͰAPIΛ ఏڙ
14. ModelͱίϯςφɾΠϝʔδ • ڊେͳML ModelΛίϯςφɾΠϝʔδʹؚΊ Δ͔൱͔ • ؚΊͳ͍ͷͰ͋Ε͹Կॲʹ഑ஔ͢Δ͔ • ϙʔλϏϦςΟੑͱϩʔυ࣌ؒͷτϨʔυΦϑ • ྑ͍ΞΠσΟΞ͕͋Ε͹ڭ͑ͯԼ͍͞…
15. ௨ৗͷAPIͱ͸ҧ͏ • ѻ͏ϦιʔεɺModelαΠζ͕େ͖͘ͳΔ৔ ߹͕ଟ͍(਺ඦMBʙ਺GB) • CPUɾϝϞϦϦιʔεͷফඅ͕ܹ͍͠ • ৔߹ʹΑͬͯ͸GPU΋࢖͏
16. ϝϞϦফඅ໰୊ • LovemachineͷPython࣮૷෦෼͸࣮ߦ࣌ʹ໿ 2GBϝϞϦΛফඅ͢Δˠࠓޙ͞Βʹ૿͑Δ༧ ఆ΋͋Δ • Scikit-learnͰهड़͞ΕͨTF-IDF౳ͷલॲཧ෦ ෼͕େ͖͘ͳΔࣄ͕ଟ͍
17. Pythonͱฒྻੑ • ౰વThread͕࢖͑ͳ͍(GILͷͨΊ) • ϓϩηεຖʹModelΛϩʔυ͢Δͱඞཁͳϝ ϞϦαΠζ͕େ͖͘ͳΔˠ Blue-Green Deployͷো֐ʹͳΔ
18. ਖ਼௚PythonͰͷServing͸
 Πϯϑϥతʹਏ͍ࣄ͕ଟ͍…
19. ϝϞϦΛݡ͘࢖͏ • fork͢ΔલʹmodelΛϩʔυ͠Copy on Write Λޮ͔͢ • k8sͷone process per containerηΦϦ͸͋ ͑ͯഁ͍ͬͯΔ
20. Copy On Writeͷ෮श ϝϞϦ ਌ϓϩηε ࢠϓϩηε 2.fork 1BHF
    " 1.allocation ಉ͡ྖҬΛࢀর
21. ϓϩηε͕ϝϞϦͷ಺༰Λ
 ॻ͖׵͑Δͱ… ϝϞϦ ਌ϓϩηε ࢠϓϩηε 1BHF
    " 1BHF
    # OS͕ผͷྖҬΛAllocationͯ͠ݩσʔλΛίϐʔ͢Δ ผͷྖҬΛࢀর
22. Current Issues • ਓؒͷߦಈΛ૬खʹ͍ͯ͠Δҝɺσʔλͷ܏ ޲͕มΘΓ΍͔ͬͨ͢Γɺ༧૝֎ͷ໰୊͕ൃ ੜͨ͠Γͯ͠ɺରԠ͠ଓ͚Δඞཁ͕͋Δ
 ˠ ML Model࡞੒ऀʹෛ୲ֻ͕͔Γଓ͚Δ
 ˠ SREͱͯ͠͸ࣗಈԽΛؚΜͩ࢓૊ΈͰղܾ ͍ͨ͠
23. In Progress • ࣾ಺ͷσʔλ͔ΒEmbedding͢Δ࣮૷Λίϯ ϙʔωϯτԽ • ಛఆͷ໰୊Λղܾ͢ΔϞσϧߏஙΛ͋Δఔ౓ ࣗಈԽ
 ˠࣾ಺ͷ໰୊ղܾʹಛԽͨ͠ઐ༻ͷAutoMLత ͳԿ͔
24. AutoFlow(Ծ) 'FBUVSF
    &YUSBDUJPO
    $PNQPOFOUT $MBTTJpDBUJPO
    $PNQPOFOUT $PODBUFOBUJPO
 $PNQPOFOUT .PEFM
    #VJMEFS $PNQPOFOUT
    3FHJTUSZ Ϋϥελ্ͰϞσϧͷ൒ࣗಈߏஙͱϋΠύʔύϥ ϝʔλͷࣗಈௐ੔Λߦ͏
25. ·ͱΊ • MLʹ͸গ͠௨ৗͱҧ͏Πϯϑϥ͕ඞཁʹͳΔ
 ˠ·ͩϕετɾϓϥΫςΟε͸෼͔Βͳ͍ • ͦ΋ͦ΋MLͳػೳΛຊ֨ӡ༻͠Α͏ͱ͢Δ ͱɺେ෯ͳࣗಈԽɾ࢓૊ΈԽΛਐΊͳ͍ͱ্ ख͘ߦ͔ͳ͍
26. ͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠!!
27. We are Hiring!!
28. SRE ML Reliability • SysML? MLOps? ৽͍͠Job description • SREεΩϧ+ML෼໺ͷجૅ஌ࣝ • MLΠϯϑϥͷࣗಈԽɾ࢓૊ΈԽΛਪ͠ਐΊͯ ͘ΕΔਓࡐ • ΋ͪΖΜଞͷ৬छ΋ઈࢍืूத!!
29. ৄࡉ͸ͪ͜Β
 https://careers.mercari.com/