Upgrade to Pro — share decks privately, control downloads, hide ads and more …

What we need for MLOps

What we need for MLOps

CCSE 2019登壇
https://ccse.jp/2019/

Keigo Hattori

July 13, 2019
Tweet

More Decks by Keigo Hattori

Other Decks in Technology

Transcript

  1. Copyright © ABEJA, Inc. All rights reserved whoami 2 Copyright

    © ABEJA, Inc. All rights reserved Keigo Hattori Software Engineer ~2009.3 Tohoku Univ / M.S. (Information Science) 2009.4~2017.10 Fuji Xerox / Researcher (ML x NLP), etc. 2017.11~2019.5 LINE / Senior Software Engineer / Clova 2019.6~ ABEJA / Software Engineer / Platform 2015.11~2019.5 Apitore / Founder Marketplace for algorithm 2017.12~ Rekcurd / Owner Flexible managing system for ML model @keigohtr
  2. Copyright © ABEJA, Inc. All rights reserved What is MLOps?

    3 Sculley et al. 2015 MLOps = Machine Learning Workflow + DevOps Cameras forbidden
  3. Copyright © ABEJA, Inc. All rights reserved MLOps Pipeline 4

    Data Collection Annotation Modeling Test Serving Service Infrastructure
  4. Copyright © ABEJA, Inc. All rights reserved MLOps Pipeline 5

    Data Collection Annotation Modeling Test Serving Data Engineer Annotator ML Engineer QA Engineer Site Reliability Engineer Service Infrastructure Infra Engineer
  5. Copyright © ABEJA, Inc. All rights reserved CASE: cookpad 6

    Cameras forbidden cookpad. PyData.Tokyo #17
  6. Copyright © ABEJA, Inc. All rights reserved CASE: Comcast 8

    Cameras forbidden Comcast. SPARK+AI SUMMIT2019
  7. Copyright © ABEJA, Inc. All rights reserved CASE: Comcast 9

    Cameras forbidden Comcast. SPARK+AI SUMMIT2019
  8. Copyright © ABEJA, Inc. All rights reserved What do we

    need for MLOps? Machine Learning Workflow • Reproducibility • Feedback DevOps • Controllability • Resource Management • Stability • Automation • Measurement • Observability 10
  9. Copyright © ABEJA, Inc. All rights reserved What do we

    need for MLOps? Machine Learning Workflow • Reproducibility = 機械学習結果を再現したい • Feedback = 機械学習モデルを継続的に改善したい DevOps • Controllability = 出力 / 結果 を制御したい • Resource Management = 計算資源を有効活用したい • Stability = 安定的に運用したい • Automation = 自動化して運用コストを下げたい • Measurement = ROI / KPI を計測したい • Observability = 監視したい 11
  10. Copyright © ABEJA, Inc. All rights reserved What should we

    do for MLOps? (My Answer) 12 Training Serving
  11. Copyright © ABEJA, Inc. All rights reserved Reference • How

    to Utilize MLflow and Kubernetes to Build an Enterprise ML Platform. https://youtu.be/cDtzu4WBzWA • Hidden Technical Debt in Machine Learning Systems. https://bit.ly/2RF9EbS • How to Deploy Machine Learning Models. https://bit.ly/2K1MQBq • クックパッドの機械学習基盤 2018. https://bit.ly/2JuSBFU 13
  12. Copyright © ABEJA, Inc. All rights reserved Reference • Komodo

    Health, https://bit.ly/2G76CIY • PFN, https://bit.ly/2XZznRV • LINE, https://bit.ly/2NMCKrt • TIS, https://bit.ly/2THs8ru • RECRUITライフスタイル, https://bit.ly/2O8aj7i • RECRUITテクノロジー, https://bit.ly/2LeqOfW • リブセンス, https://bit.ly/32llTPH • DeNA, https://bit.ly/30t9Qhr • メルカリ, https://bit.ly/2YNHfD3 and so on… 14
  13. Copyright © ABEJA, Inc. All rights reserved We are hiring

    • https://hrmos.co/pages/abeja • https://www.wantedly.com/companies/abeja/projects 15