mlct.pdf

Transcript

1. ϝϧΧϦͷMLج൫ MLCT vol.5
 
 hnakagawa

2. ࣗݾ঺հ • Hirofumi Nakagawa (hnakagawa) • 2017೥7݄ೖࣾ • ॴଐ͸SRE • σόΠευϥΠό։ൃ͔Βϑϩϯ τΤϯυ։ൃ·Ͱ΍ΔԿͰ΋԰ • NOT σʔλαΠΤϯςΟετ • https://github.com/hnakagawa
3. ͓࢓ࣄ • ML Platform։ൃ • σʔλαΠΤϯςΟετͱSREͷεΩϧΪϟο ϓΛຒΊΔ • ML Reliability, SysML?, MLOps? • SREͷཱ৔͔ΒMLγεςϜͷࣗಈԽΛߦ͏
4. ML Platform • ಺੡ͷML Platform • kubernetesϕʔε • طଘͷML FrameworkΛ࢖༻͠ ؆୯ʹTraining/ServingΛߦ͏ ؀ڥΛఏڙ
5. ͦͷ͏ͪOSSͰެ։༧ఆ(ଟ෼
6. ϝϧΧϦͷMLར༻ࣄྫ • ײಈग़඼ • ҧ൓ग़඼ݕ஌ • Ձ֨αδΣετ • ΢ΤΠταδΣετ
 ౳ʑ… ̍೔਺ઍສpredictionΛߦ͍ͬͯΔ
7. ML Platform Architecture ,VCFSOFUFT $POUSPMMFS $-* $MVTUFS8PSLGMPX %BTICPBSE 4UPSBHF
   (BUFXBZ .FUSJDT 3VOOFS $PNQPOFOU .FSDBSJ
   .-
   $PNQPOFOU &YUFSOBM
   .JEEMFXBSF
8. Model Training & Serving
 Workflow
9. .-
   1MBUGPSN
   USBJOJOH
   DMVTUFS Workflow for Production $* .-
   1MBUGPSN
   TFSWJOH
   DMVTUFS
   GPS
   UFTU .PEFM
   3FHJTUSZ +PC +PC ɾɾ 3&45
 "1* 4USFBNJOH 5'4FSW JOH ɾɾɾ
10. .-
    1MBUGPSN
    USBJOJOH
    DMVTUFS Training Workflow $* .PEFM
    3FHJTUSZ +PC +PC ɾɾɾ 1. GitHub΁ͷpushΛτϦΨʹtrainingΛىಈ 2. Training͞ΕͨModel͸Model Registry
 ΁্͕Δ
11. Serving Workflow .-
    1MBUGPSN
    TFSWJOH
    DMVTUFS
    GPS
    UFTU .PEFM
    3FHJTUSZ ɾɾ 3&45
 "1* 4USFBNJOH 5'
    4FSWJOH 1. Model RegistryΛ؂ࢹͯࣗ͠ಈͰModel ΛServing 2. Serving&Test͕੒ޭ͢Δͱຊ൪༻k8s manifestΛग़ྗ
12. Container Workflow %BUB4PVSDF
 *NBHF 5FYUɹ 1SFQSPDFT TJOH
    *NBHF &TUJNBUPS
 *NBHF 17 17 1JDUVSF
    1SFQSPDFT TJOH
    *NBHF 17 It’s own implementation
13. 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
14. 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
15. ModelͱίϯςφɾΠϝʔδ • ڊେͳML ModelΛίϯςφɾΠϝʔδʹؚΊ Δ͔൱͔ • ؚΊͳ͍ͷͰ͋Ε͹Կॲʹ഑ஔ͢Δ͔ • ϙʔλϏϦςΟੑͱϩʔυ࣌ؒͷτϨʔυΦϑ • ྑ͍ΞΠσΟΞ͕͋Ε͹ڭ͑ͯԼ͍͞…
16. ௨ৗͷAPIͱಛੑ͕ҧ͏ • ѻ͏ϦιʔεɺModelαΠζ͕େ͖͘ͳΔ৔ ߹͕ଟ͍(਺ඦMBʙ਺GB) • CPUɾϝϞϦϦιʔεͷফඅ͕ܹ͍͠ • ৔߹ʹΑͬͯ͸GPU΋࢖͏
17. ϝϞϦফඅ໰୊ • ҧ൓ݕ஌γεςϜͷPython࣮૷෦෼͸࣮ߦ࣌ ʹ໿2GBϝϞϦΛফඅ͢Δˠࠓޙ͞Βʹ૿͑ Δ༧ఆ΋͋Δ • Scikit-learnͰهड़͞Εͨલॲཧ෦෼͕େ͖͘ ͳΓ͕ͪ
18. Pythonͱฒྻੑ • ౰વThread͕࢖͑ͳ͍(GILͷͨΊ) • ϓϩηεຖʹModelΛϩʔυ͢Δͱඞཁͳϝ ϞϦαΠζ͕େ͖͘ͳΔˠ Blue-Green Deployͷো֐ʹͳΔ
19. ਖ਼௚PythonͰͷServing͸
 Πϯϑϥతʹਏ͍ࣄ͕ଟ͍…
20. ϝϞϦΛݡ͘࢖͏ • fork͢ΔલʹmodelΛϩʔυ͠Copy on Write Λޮ͔͢ • k8sͷone process per containerηΦϦ͸͋ ͑ͯഁ͍ͬͯΔ
21. Copy On Writeͷ෮श ϝϞϦ ਌ϓϩηε ࢠϓϩηε 2.fork 1BHF
    " 1.allocation ಉ͡ྖҬΛࢀর
22. ϓϩηε͕ϝϞϦͷ಺༰Λ
 ॻ͖׵͑Δͱ… ϝϞϦ ਌ϓϩηε ࢠϓϩηε 1BHF
    " 1BHF
    # OS͕ผͷྖҬΛAllocationͯ͠ݩσʔλΛίϐʔ͢Δ ผͷྖҬΛࢀর
23. Current Issues
24. ߴ౓ͳܧଓతϝϯςφϯε͕ඞཁ • MLػೳ͸σʔλͷ܏޲͕มΘͬͨΓɺ༧૝֎ ͷ໰୊͕ൃੜͨ͠Γͯ͠ɺͦΕΒʹରԠ͠ଓ ͚Δඞཁ͕͋Δ MLػೳ͸ϦϦʔεޙ΋େ͖ͳ ίετ͕͔͔Γଓ͚Δ
25. େ෯ͳࣗಈԽ͕ඞਢ
26. In Progress
27. ߴ౓ͳࣗಈԽ • ࣾ಺ͷσʔλ͔ΒFeature Extraction͢Δ࣮૷ ΛίϯϙʔωϯτԽ • ಛఆͷ໰୊Λղܾ͢ΔϞσϧߏஙΛ͋Δఔ౓ ࣗಈԽ • ϦϦʔεޙͷRe-TrainingɺHyper parameter optimizationɺDeploy౳ΛࣗಈԽ
28. AutoFlow 'FBUVSF
    &YUSBDUJPO
    $PNQPOFOUT $MBTTJGJDBUJPO
    $PNQPOFOUT $PODBUFOBUJPO
 $PNQPOFOUT .PEFM
    #VJMEFS $PNQPOFOUT
    3FHJTUSZ Ϋϥελ্ͰϞσϧͷ൒ࣗಈߏஙͱϋΠύʔύϥ ϝʔλͷࣗಈௐ੔Λߦ͏
29. AutoServing %FQMPZ ϦϦʔεޙͷਫ਼౓؂ࢹɾRe-TrainingɾRe-Deploy౳ ΛࣗಈͰߦ͏ .POJUPSJOH &WBMVBUJPO )ZQFS
    QBSBNFUFS
    PQUJNJ[BUJPO 3F5SBJOJOH
30. ·ͱΊ • MLʹ͸গ͠௨ৗͱҧ͏Πϯϑϥ͕ඞཁʹͳΔ
 ˠ·ͩϕετɾϓϥΫςΟε͸෼͔Βͳ͍ • ͦ΋ͦ΋MLͳػೳΛຊ֨ӡ༻͠Α͏ͱ͢Δ ͱɺେ෯ͳࣗಈԽɾ࢓૊ΈԽΛਐΊͳ͍ͱ্ ख͘ߦ͔ͳ͍
31. ͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠!!