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

toward airflow without pain

LINE Developers
May 30, 2019
2.1k

toward airflow without pain

Kenji Yamauchi
Machine Learning Production Pitch #2
https://machine-learning-pitch.connpass.com/event/125115/

LINE Developers

May 30, 2019
Tweet

More Decks by LINE Developers

Transcript

  1. ΞʔΩςΫνϟશମ PostgreSQL (Airflow Backend)  Workers Web Server Celery Scheduler

    Airflow Redis (Celery Broker) DAGs Playbooks Bucket Deployment
  2. ΞʔΩςΫνϟશମ PostgreSQL (Airflow Backend) Workers Web Server Celery Scheduler Airflow

    Redis (Celery Broker) DAGs Playbooks Bucket Deployment M G Q AD Q PS Q
  3. ΞʔΩςΫνϟશମ PostgreSQL (Airflow Backend)  Workers Web Server Celery Scheduler

    Airflow Redis (Celery Broker) DAGs Playbooks Bucket Deployment       
  4. ΞʔΩςΫνϟશମ PostgreSQL (Airflow Backend)  Workers Web Server Celery Scheduler

    Airflow Redis (Celery Broker) DAGs Playbooks Bucket Deployment    
  5. ϏϧυʙධՁ·Ͱͷϑϩʔྫ ੔ܗ Models ֶश ධՁ ఆظతʹ SBXσʔλΛGFUDI ίʔυมԽʹUSJHHFS ࠷৽σʔλͰ ֶशΛUSJHHFS

    ֶशऴྃ࣌ʹUSJHHFS • ֶशɾධՁʹ͍ͭͯ͸ɺQPPMΛ͔࣮ͭͬͯߦϚγϯΛ %"(୯ҐͰࢦఆ ୆Ͱฒྻ࣮ߦ ݁Ռ௨஌ ϞσϧͷετΞ ߴεϖοΫϚγϯͰ࣮ߦ ʜ%"(
  6. ໨ඪ PostgreSQL (Airflow Backend)  DAG Execution Pods Launches when

    DAG executes Web Server Kubernetes Cluster Store logs/training, evaluation data/models Docker Registry S3 like bucket Build Test Publish Deploy (Reload web server & scheduler images) Triggered by developer’s push Scheduler Airflow
  7. ໨ඪ PostgreSQL (Airflow Backend)  DAG Execution Pods Launches when

    DAG executes Web Server Kubernetes Cluster Store logs/training, evaluation data/models Docker Registry S3 like bucket Deploy (Reload web server & scheduler images) Scheduler Airflow Build Test Publish Triggered by developer’s push ࣮ߦ୯Ґ͸7.Πϯελϯε͔Β1PEʹ %"(΋ؚΊͯ"JSGMPX͸શ෦ͭʹ
  8. ໨ඪ PostgreSQL (Airflow Backend)  DAG Execution Pods Launches when

    DAG executes Web Server Kubernetes Cluster Store logs/training, evaluation data/models Docker Registry S3 like bucket Build Test Publish Deploy (Reload web server & scheduler images) Triggered by developer’s push Scheduler Airflow %"(΍%PDLFSGJMFʹมߋ͕͋Δ౓ʹ %SPOFͰϏϧυɾςετ࣮ߦ
  9. ໨ඪ PostgreSQL (Airflow Backend)  DAG Execution Pods Launches when

    DAG executes Web Server Kubernetes Cluster Store logs/training, evaluation data/models Docker Registry S3 like bucket Deploy (Reload web server & scheduler images) Scheduler Airflow Build Test Publish Triggered by developer’s push Ϗϧυ͕੒ޭͨ͠Β ϨδετϦʹίϯςφΛొ࿥"JSGMPXSFMPBE
  10. ΞʔΩςΫνϟมߋʹΑΔվળ • ϚγϯϦιʔε؅ཧίετ࡟ݮ • ݸʑͷ7.Πϯελϯεͷ໘౗Λݟͳͯ͘Α͘ͳͬͨ • 8FCTFSWFS 4DIFEVMFS 8PSLFS͸શͯಉ͡࢓༷ͷίϯς φͳͷͰɺ%PDLFSGJMFͭΛ؅ཧ͢Ε͹ྑ͘ͳͬͨ

    • ͍ͭͰʹ$FMFSZΛ΍ΊͨͷͰ3FEJT؅ཧ΋ͳ͠ʹ • ίϯςφϕʔεͷ؅ཧʹͳΓɺ؀ڥ࠶ݱੑ΋޲্ • .JOJLVCFΛར༻ͨ͠ϩʔΧϧͰͷ؀ڥ࠶ݱ΋ָʹ
  11. %"(ͷςετͰ΍Γ͍ͨ͜ͱ • ཧ૝తʹ͸FOEUPFOEͰ%"(ͷڍಈΛอূ͍ͨ͠ • ࠷௿Ͱ΋5BTL୯ҐͰͷςετ͸͍ͨ͠ ྫ ֶश%"(ʹ͍ͭͯ • &&ςετͰ͸ֶश։࢝τϦΨ σʔλ੔ܗऴྃɾϞσϧ

    ίʔυมߋ ͝ͱʹਖ਼͘͠ىಈ͢Δ͔ɺֶशͷ࠷ޙʹਖ਼͍͠ ϑΥʔϚοτͷϑΝΠϧ໊ͰϞσϧΛग़ྗ͢Δ͔ FUD • 5BTL୯ҐͰ͸ɺਖ਼͍͠όʔδϣϯͷίʔυΛDMPOFͯ͜͠ Α͏ͱ͍ͯ͠Δ͔ FUD