ML/MLOps Engineer
[MLOps] MLOps์˜ ์›์น™๊ณผ ๊ตฌ์„ฑ์š”์†Œ

์•ž์„  ๊ฒŒ์‹œ๊ธ€์—์„œ ML์„ ์ž๋™ํ™”ํ•˜๊ณ  ์šด์˜ํ•˜๋Š”๋ฐ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด MLOps ๊ธฐ์ˆ ์ด ์‚ฌ์šฉ๋œ๋‹ค๊ณ  ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฒˆ ๊ฒŒ์‹œ๊ธ€์—์„œ๋Š” MLOps๊ฐ€ ์ ์šฉ๋˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋ฌด์—‡์ด ํ•„์š”ํ•œ์ง€ ์‚ดํŽด๋ณด๋ ค ํ•ฉ๋‹ˆ๋‹ค.

Machine Learning Operations (MLOps):Overview, Definition, and Architecture (2022)
์œ„์˜ ์—ฐ๊ตฌ์—์„œ MLops์˜ ์›์น™๊ณผ ๊ตฌ์„ฑ์š”์†Œ, ์—ญํ• ๋“ค์— ๋Œ€ํ•ด์„œ ์ƒ์„ธํ•˜๊ฒŒ ์†Œ๊ฐœํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฌธํ—Œ ๊ฒ€ํ† ์™€ MLOps Tool ์กฐ์‚ฌ, ์ „๋ฌธ๊ฐ€์™€ ์ธํ„ฐ๋ทฐ๋ฅผ ํ†ตํ•ด์„œ MLOps๋ฅผ ๊ฐœ๋…ํ™”ํ•˜์˜€์Šต๋‹ˆ๋‹ค. MLOps๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ  ์žˆ๋Š” ์›์น™๊ณผ ๊ธฐ์ˆ , ์—ญํ• ๋“ค์— ๋Œ€ํ•ด์„œ ์•Œ์•„๋ด…๋‹ˆ๋‹ค.

Principles

MLOps๋ฅผ ์‹คํ˜„ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” 9๊ฐ€์ง€ ์›์น™์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

  • P1. CI/CD ์ž๋™ํ™” - ํŠน์ • ๋‹จ๊ณ„์˜ ์„ฑ๊ณต์—ฌ๋ถ€์— ๋Œ€ํ•œ ๋น ๋ฅธ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜์—ฌ ์ „์ฒด ์ƒ์‚ฐ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค.
  • P2. Workflow ์กฐ์ • - DAG(Directed Acyclic Graphs) ์— ๋”ฐ๋ผ ์ž‘์—… ์ˆœ์„œ๋ฅผ ์กฐ์ •ํ•จ. ์ž‘์—…๊ฐ„ ๊ด€๊ณ„์™€ ์ข…์†์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ์‹คํ–‰ ์ˆœ์„œ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค.
  • P3. ์žฌํ˜„์„ฑ - ML ์‹คํ—˜์„ ์žฌํ˜„ํ•˜์˜€์„ ๋•Œ ๋™์ผํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
  • P4. ๋ฒ„์ „ ๊ด€๋ฆฌ - ๋ฐ์ดํ„ฐ์™€ ๋ชจ๋ธ์˜ ๋ฒ„์ „ ๊ด€๋ฆฌ๋ฅผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์žฌํ˜„์„ฑ๊ณผ ์ถ”์ ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
  • P5. ํ˜‘์—… - ๋ฐ์ดํ„ฐ์™€ ๋ชจ๋ธ, ์ฝ”๋“œ์— ๋Œ€ํ•ด ํ˜‘์—…์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
  • P6. ๋ฐ˜๋ณต์ ์ธ ML ํ•™์Šต ๋ฐ ํ‰๊ฐ€ - ์ƒˆ๋กœ์šด feature ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ML ๋ชจ๋ธ์„ ์ฃผ๊ธฐ์ ์œผ๋กœ ์žฌํ•™์Šต ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋ชจ๋‹ˆํ„ฐ๋ง๊ณผ ํ”ผ๋“œ๋ฐฑ, ์ž๋™ํ™”๋œ ํŒŒ์ดํ”„๋ผ์ธ์œผ๋กœ ์ง€์†์ ์ธ ํ•™์Šต์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ์—ฐ์†์ ์ธ ํ•™์Šต์€ ๋ชจ๋ธ ํ’ˆ์งˆ ํ‰๊ฐ€๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
  • P7. ML ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ถ”์ /๋กœ๊น… - ML ์›Œํฌํ”Œ๋กœ์šฐ ์ž‘์—…์— ๋Œ€ํ•ด ์ถ”์  ๊ณผ ๊ธฐ๋ก์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ, ๋ฐ˜๋ณต์ ์ธ ํ•™์Šต ๋‚ด์šฉ๊ณผ ๊ฐ™์€ ๊ฒƒ๋“ค์ด ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค.
  • P8. ์ง€์†์ ์ธ ๋ชจ๋‹ˆํ„ฐ๋ง -์ž ์žฌ์ ์ธ ์˜ค๋ฅ˜์™€ ๋ณ€ํ™”๋ฅผ ํƒ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ, ์ฝ”๋“œ, ๋ฐฐํฌ ์„ฑ๋Šฅ(ex ์ •ํ™•๋„)๋“ฑ์„ ์ฃผ๊ธฐ์ ์œผ๋กœ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
  • P9. ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„ - ํ”ผ์ฒ˜ ์—”์ง€๋‹ˆ์–ด๋ง ๋‹จ๊ณ„๊นŒ์ง€ ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„, ๋ชจ๋ธ ๋ฐฐํฌ ์„ฑ๋Šฅ ๊ด€์ฐฐ ๋ฃจํ”„์™€ ๊ฐ™์ด ํ†ต์ฐฐ๋ ฅ์„ ํ‚ค์šฐ๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

MLOps์˜ ์›์น™์— ๋Œ€ํ•ด์„œ ์‚ดํŽด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ๋Š” MLOps ๊ธฐ์ˆ  ๊ตฌ์„ฑ์š”์†Œ๋ฅผ ์•Œ์•„๋ด…๋‹ˆ๋‹ค.

Technical Components

MLOps๋ฅผ ์ด๋ฃจ๋Š” ๊ธฐ์ˆ  ๊ตฌ์„ฑ ์š”์†Œ๋Š” ์•„๋ž˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ฐ ํ•ญ๋ชฉ ์˜†์— ์žˆ๋Š” P_ ๋Š” ์•ž์„œ ์‚ดํŽด๋ณธ ์–ด๋–ค ์›์น™์— ํ•ด๋‹นํ•˜๋Š” ๊ธฐ์ˆ  ๊ตฌ์„ฑ์š”์†Œ์ธ์ง€ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.

  • C1. CI/CD ๊ตฌ์„ฑ์š”์†Œ (P1, P6, P9) - DevOps์—์„œ๋„ ๋‹ค๋ฃจ๋Š” continuous integration, continuous delivery, and continuous deployment ๊ฐ€ MLOps์—์„œ๋„ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋นŒ๋“œ๋ถ€ํ„ฐ ํ…Œ์ŠคํŠธ, ๋ฐฐํฌ๊นŒ์ง€ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  • C2. ์†Œ์Šค์ฝ”๋“œ ์ €์žฅ์†Œ (P4, P5) - ์†Œ์Šค์ฝ”๋“œ๋ฅผ ์ €์žฅํ•˜๊ณ  ๋ฒ„์ „๊ด€๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. Bitbucket, GitLab, Github, Gitea ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
  • C3. Workflow Orchestration Component (P2, P3, P6) - DAG(Directed Asyclic Graph)๋ฅผ ํ†ตํ•ด์„œ ML ์›Œํฌํ”Œ๋กœ์šฐ์˜ ์ž‘์—… ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Airflow, Kubeflow, Luigi, SageMaker ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
  • C4. Feature Store System(P3, P4) - ๋ฐ์ดํ„ฐ ์ €์žฅ์†Œ์—์„œ ์ง์ ‘ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์ ธ์˜ฌ ์ˆ˜๋„ ์žˆ์ง€๋งŒ, ๋‚ฎ์€ ์˜ˆ์ธก ์ง€์—ฐ ์‹œ๊ฐ„์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•œ ์˜จ๋ผ์ธ ํ”ผ์ฒ˜ ์Šคํ† ์–ด์™€ ์‹คํ—˜์„ ์œ„ํ•œ ์˜คํ”„๋ผ์ธ ํ”ผ์ฒ˜ ์Šคํ† ์–ด๋ฅผ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AWS Feature Store, Tecton.ai, Hopswork.ai ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
  • C5. ๋ชจ๋ธ ํ•™์Šต ์ธํ”„๋ผ(P6) - ๋ชจ๋ธ ํ•™์Šต ์ธํ”„๋ผ์—์„œ๋Š” CPU, RAM, GPU์™€ ๊ฐ™์€ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ„์‚ฐํ˜• ์ธํ”„๋ผ๊ฐ€ ๊ถŒ์žฅ๋ฉ๋‹ˆ๋‹ค. Kubernetes, Red Hat OpenShift ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
  • C6. ๋ชจ๋ธ ๋ ˆ์ง€์ŠคํŠธ๋ฆฌ(P3, P4) - ํ•™์Šต๋œ ML ๋ชจ๋ธ๊ณผ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. ML model artifact์™€ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. MLflow, AWS SageMaker ๋ชจ๋ธ ๋ ˆ์ง€์ŠคํŠธ๋ฆฌ, Azure ML ๋ชจ๋ธ ๋ ˆ์ง€์ŠคํŠธ๋ฆฌ, Neptune.ai ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ €์žฅ์†Œ ์˜ˆ์‹œ๋กœ๋Š” Azure Storage, Google Cloud Storage, S3 ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
  • C7. ML ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ €์žฅ์†Œ(P4, P7) - ML ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ์˜ˆ๋ฅผ๋“ค์–ด์„œ ML ์›Œํฌํ”Œ๋กœ์šฐ ํŒŒ์ดํ”„๋ผ์ธ ์ž‘์—…์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ถ”์ ์ด ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ๋ณ„ ๋ฉ”ํƒ€ ๋ฐ์ดํ„ฐ(ํ•™์Šต ์‹œ๊ฐ„, ๊ธฐ๊ฐ„ ๋“ฑ), ํ•™์Šต ์ž‘์—… ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ, ๋ชจ๋ธ๋ณ„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ(๊ฒฐ๊ณผ, ์„ฑ๋Šฅ, ์ฝ”๋“œ ๋“ฑ)์„ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Kubeflow Pipeline, SageMaker Pipeline, Azure ML, MLflow ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
  • C8. ๋ชจ๋ธ ๋ฐฐํฌ ๊ตฌ์„ฑ์š”์†Œ(P1) - ์‹ค์‹œ๊ฐ„ ์˜ˆ์ธก์„, ๋Œ€๋Ÿ‰ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ์˜ˆ์ธก ๋ฐฐ์ธ  ๋“ฑ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. REST API๋กœ ์ œ๊ณต์ด ๊ฐ€๋Šฅํ•˜๊ณ  Kubernetes, Docker๋ฅผ ์ด์šฉํ•˜์—ฌ ML ๋ชจ๋ธ์„ ์ปจํ…Œ์ด๋„ˆํ™” ํ•˜๋Š” ๊ฒƒ์ด๋‚˜ Flask์™€ ๊ฐ™์€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์žˆ์Šต๋‹ˆ๋‹ค. Kubernetes์˜ Kubeflow์—๋„ ๋ฐฐํฌ ๊ธฐ๋Šฅ์ด ํฌํ•จ๋˜์–ด ์žˆ๊ณ , ๋ฐฐ์น˜ ์˜ˆ์ธก์„ ์œ„ํ•ด์„œ Spark๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํด๋ผ์šฐ๋“œ ์„œ๋น„์Šค์˜ ์˜ˆ๋กœ Microsoft Azure ML REST API, AWS SageMaker Endpoints, IBM Watson Studio, Google Vertex AI prediction service๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
  • C9. ๋ชจ๋‹ˆํ„ฐ๋ง ๊ตฌ์„ฑ์š”์†Œ(P8, P9) - ๋ฐฐํฌ ์„ฑ๋Šฅ๊ณผ ๋ชจ๋ธ infrastructure, CI/CD ๋“ฑ์— ๋Œ€ํ•œ ๋ชจ๋‹ˆํ„ฐ๋ง์„ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. Prometheus with Grafana, ELK stack(Elasticsearch, Logstash, Kibana), TensorBoard ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋‹ˆํ„ฐ๋ง ๊ธฐ๋Šฅ์ด ๋‚ด์žฅ๋œ ์˜ˆ๋กœ๋Š” Kubeflow, MLFlow, AWS SageMaker ๋ชจ๋‹ˆํ„ฐ๋ง ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

์•ž์„œ MLOps์˜ ์›๋ฆฌ์™€ ๊ตฌ์„ฑ์š”์†Œ๋ฅผ ์‚ดํŽด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ๋Š” MLOps๋ฅผ ์œ„ํ•ด ์–ด๋– ํ•œ ์—ญํ• ๋“ค์ด ํ•„์š”ํ•œ์ง€์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์—ญํ•  ๊ด€๊ณ„์ž๋Š” ์˜์–ด ์› ํ‘œํ˜„์„ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.

Roles

  • R1. Business Stakeholder (๋น„์Šทํ•œ ์—ญํ• : Product Owner, Project Manager) - ML๋กœ ๋‹ฌ์„ฑํ•˜๊ณ ์ž ํ•˜๋Š” ๋น„์ฆˆ๋‹ˆ์Šค ๋ชฉํ‘œ๋ฅผ ์ •์˜ํ•˜๊ณ , ML ์ œํ’ˆ์œผ๋กœ ๋ฐœ์ƒํ•˜๋Š” ํˆฌ์ž์ˆ˜์ต๋ฅ  ๋“ฑ ๋น„์ฆˆ๋‹ˆ์Šค ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
  • R2. Solution Architect (๋น„์Šทํ•œ ์—ญํ• : IT Architect) - ์ฒ ์ €ํ•œ ํ‰๊ฐ€๋ฅผ ๊ฑฐ์ณ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๊ณ„ํ•˜๊ณ  ์‚ฌ์šฉํ•  ๊ธฐ์ˆ ์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค.
  • R3. Data Scientist (๋น„์Šทํ•œ ์—ญํ• : ML Specialist, ML Developer) - ๋น„์ฆˆ๋‹ˆ์Šค ๋ฌธ์ œ๋ฅผ ML ๋ฌธ์ œ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ ์กฐ์ •๊ณผ ๊ฐ™์ด ๋ชจ๋ธ ์—”์ง€๋‹ˆ์–ด๋ง์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
  • R4. Data Engineer (๋น„์Šทํ•œ ์—ญํ• : DataOps Engineer) - ๋ฐ์ดํ„ฐ ๋ฐ ํ”ผ์ณ ์—”์ง€๋‹ˆ์–ด๋ง ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ตฌ์ถ•ํ•˜๊ณ  ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. feature store system์˜ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์— ์ ์ ˆํ•˜๊ฒŒ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.
  • R5. Software Engineer - ์†Œํ”„ํŠธ์›จ์–ด ์„ค๊ณ„ ํŒจํ„ด, ๊ฐœ๋ฐœ ์ง€์นจ ๋ฐ ๋ชจ๋ฒ” ์‚ฌ๋ก€๋ฅผ ์ ์šฉํ•˜์—ฌ ์ œํ’ˆ์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค.
  • R6. DevOps Engineer - ๊ฐœ๋ฐœ๊ณผ ์šด์˜ ์‚ฌ์ด์˜ ๊ฐ„๊ทน์„ ๋ฉ”์šฐ๊ณ  ์ ์ ˆํ•œ CI/CD์ž๋™ํ™”, ML ์›Œํฌํ”Œ๋กœ์šฐ ์กฐ์ •, ๋ชจ๋ธ ๋ฐฐํฌ, ๋ชจ๋‹ˆํ„ฐ๋ง์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
  • R7. ML Engineer/MLOps Engineer - R3, R4, R5, R6 ์˜ ๋„“์€ ์˜์—ญ์—์„œ ๋น„์Šทํ•ฉ๋‹ˆ๋‹ค. ML ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  ์šด์˜ํ•˜๋ฉฐ, ์ž๋™ํ™”๋œ ML ์›Œํฌํ”Œ๋กœ์šฐ ํŒŒ์ดํ”„๋ผ์ธ๊ณผ ๋ชจ๋ธ์˜ ํ”„๋กœ๋•์…˜ ๋ฐฐ์น˜๋ฅผ ๊ด€๋ฆฌํ•˜๊ณ , ๋ชจ๋ธ๊ณผ ML์ธํ”„๋ผ๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•ฉ๋‹ˆ๋‹ค.

์ด๋ฒˆ ๊ฒŒ์‹œ๊ธ€์—์„œ๋Š” MLOps์˜ ์›์น™๊ณผ ๊ธฐ์ˆ  ์š”์†Œ, ์—ญํ• ์— ๋Œ€ํ•ด์„œ ์‚ดํŽด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ํ•„์š”ํ•œ ๊ธฐ์ˆ  ๊ตฌ์„ฑ์š”์†Œ๋„ ๋งŽ๊ณ  ์—ญํ• ๋“ค๋„ ๋‹ค์–‘ํ•œ๋ฐ์š”. ๋‹ค์Œ ๊ฒŒ์‹œ๊ธ€์—์„œ๋Š” ๊ฐ ์š”์†Œ๋“ค์ด ์–ด๋–ป๊ฒŒ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ๋Š”์ง€ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ†ตํ•ด ์ดํ•ดํ•ด๋ด…๋‹ˆ๋‹ค. ML ์ƒ๋ช… ์ฃผ๊ธฐ์˜ ๊ฐ ๋‹จ๊ณ„๊ฐ€ ์–ด๋–ป๊ฒŒ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ๋Š”์ง€ ์•Œ์•„๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.


reference
Dominik Kreuzberger and Niklas Kühl and Sebastian Hirschl (2022) Machine Learning Operations (MLOps): Overview, Definition, and Architecture