10. 9
Machine Learning Project Life cycle
Goal
Definition
Feature
Engineering
Model
Training
Model
Evaluation
Model
Deployment
Model
Maintainance
Model
Serving
Model
Monitoring
Data
Collection &
Preparation
Business
Problem
27. 26
Siloed Team, between Teams Communication
Team A TeamB Team C Team D
Team โฆ Team.. Teamโฆ
Enterprise ML work Teams
Not Everyone Speaks the Same
Language
33. Propritary
Inference
Servers
Using proprietary tools to
perform modeling and
inference
โข SAS
โข SPSS
โข FICO
The Rise of
Open Source
Data Science Tools
โฆ , attempt to wrap the data
science stack in a lightweight web
service framework, and put it into
production.
Phython:
โข SciPy stack
โข Scitkit-learn
โข TensorFlow etc.
R:
โข dplyr
โข ggplot2
โข Etc.
โข Spark, H2O, othersโฆ
Containerization
to-the rescue
containerized the โStone
Ageโ approach, making it
easy to scale, robust, etc.
MLOps
Platform
โข Dockerized open source
ML stacks
โข Deployed them on premise
or in the cloud via
Kubernetes
โข And providing some
manageability(MLOps)
time
Pre-History Age Stone Age Bronze Age
MLOps
Gold Rush Age
2000 2015 2018
https://github.com/adbreind/open-standard-models-2019/blob/master/01-
Intro.ipynb
Adam Breindel
Evolution of MLOps
โ ๊ณ์ ์งํ์ค
34. 33
MLOps Principles โ โContinuous Xโ
MLOps is an ML engineering culture that includes the following practices:
โข Continuous Integration (CI) extends the testing and validating code and components
by adding testing and validating data and models.
โข Continuous Delivery (CD) concerns with delivery of an ML training pipeline that
automatically deploys another the ML model prediction service.
โข Continuous Training (CT) is unique to ML systems property, which automatically
retrains ML models for re-deployment.
โข Continuous Monitoring (CM) concerns with monitoring production data and model
performance metrics, which are bound to business metrics.
38. 37
increase automation and improve the quality of production ML
MLOps looks to
โข Machine Learning
โข DevOps (IT)
โข Data Engineering
Components of MLOps
MLOpsis
defined as โa practice for collaboration and communication
between data scientists and operations professionals to help manage
production ML (or deep learning) lifecycle.
bmc.com/blogs/mlops-machine-learning-ops
An ML engineering culture and practice that aims at unifying ML
system development(Dev) and ML system operation (Ops) - Google
42. 41
MLOps Setup
Components
Description
Source Control ์ฝ๋, ๋ฐ์ดํฐ ๋ฐ ML ๋ชจ๋ธ ์ํฐํฉํธ์ ๋ฒ์ ๊ด๋ฆฌ
Test & Build Services
(1) ๋ชจ๋ ML ์ํฐํฉํธ์ ๋ํ ํ์ง ๋ณด์ฆ ๋ฐ (2) ํ์ดํ ๋ผ์ธ ์ฉ ํจํค์ง ๋ฐ
์คํ ํ์ผ ๋น๋๋ฅผ ์ํด CI ๋๊ตฌ ์ฌ์ฉ
Deployment Services ๋์ ํ๊ฒฝ์ ํ์ดํ ๋ผ์ธ์ ๋ฐฐํฌํ๊ธฐ ์ํด CD ๋๊ตฌ ์ฌ์ฉ
Model Registry ์ด๋ฏธ ํ๋ จ๋ ML ๋ชจ๋ธ์ ์ ์ฅํ๊ธฐ ์ํ ๋ ์ง์คํธ๋ฆฌ
Feature Store
์ ๋ ฅ ๋ฐ์ดํฐ๋ฅผ ๋ชจ๋ธ ํ์ต ํ์ดํ ๋ผ์ธ ๋ฐ ๋ชจ๋ธ ์ ๊ณต ์ค์ ์ฌ์ฉํ ํผ์ฒ๋ก
์ฌ์ ์ฒ๋ฆฌ
ML Metadata Store
๋ชจ๋ธ๋ช , ๋งค๊ฐ ๋ณ์, ํ์ต ๋ฐ์ดํฐ, ํ ์คํธ ๋ฐ์ดํฐ ๋ฐ ๋ฉํธ๋ฆญ ๊ฒฐ๊ณผ์ ๊ฐ์
๋ชจ๋ธ ํ์ต์ ๋ฉํ ๋ฐ์ดํฐ๋ฅผ ์ถ์
ML Pipeline
Orchestrator
ML ์คํ ๋จ๊ณ ์๋ํ
MLOps Setup Components
ml-ops.org/content/mlops-principles
MLOps Principles
43. 42
Design
Model
Development
Operations
Requirements Engineering
ML Use Case Priorization
Data Availability Check
Data Engineering
ML Model Engineering
Model Testing & Validation
ML Model Deployment
CI/CD Pipelines
Monitoring & Triggering
ml-ops.org/content/mlops-principles
MLOps Principles
Iterative-Incremental Process in MLOps
44. 43
Experimentation &
Development
Training Pipeline
Continus Training
Model Serving
Pipeline
Continus
Evaluation
Traceability &
Explainability
Code
Repository
Artifact
Repository
Trained Models &
ML metadata
Deployed
models
Model
Logs
https://medium.com/technoesis/mlops-is-a-practice-not-a-tool-41674c5bdad7
MLOps is a Practice, Not a Tool
Continuous feedback loops
with an MLOps workflow
45. 44
ML Engineering & Operations
Product
manager
Subject matter
Expert
Business
Objective
46. 45
ML Engineering & Operations
Data
Acquisition
Exploratory
Data
Analysis
Product
manager
Subject matter
Expert
Business
Objective
Data
Engineer
56. 55
Model Deployment CI/CD
Run Automated Tests
Source
Repository
Deploy to Target
Enviornment
Build Prediction Service
Automated E2E Pipeline
Model
Registry
Reliable & Monitored Serving
57. 56
Model Deployment CI/CD
Run Automated Tests
Source
Repository
Deploy to Target
Enviornment
Build Prediction Service
Automated E2E Pipeline
Model
Registry
Serving Infrastructure
Explain Monitor
Predict
Live Data Evaluate
Log Store
Performance &
Event Logs
ML Metadata
Evaluations,
Data Drift and
Concept Drift
notification
Reliable & Monitored Serving
69. 68
ํจ๊ณผ์ ์ธ MLOps ์๋ฃจ์ ์ด ์ ๊ณตํ๋ ๋ชจ๋ธ ๋ณด์ ์์
โข ์ก์ธ์ค ์ ์ด ๋ฉ์ปค๋์ฆ
โข ๋ชจ๋ธ ์ฌ์ฉ ๋ฐ ์ก์ธ์ค ๊ฐ์ฌ
โข ๋ชจ๋ธ, ํ๋ จ ๋ฐ์ดํฐ ๋ฐ ์ค์ ๋ณดํธ
โข ์ทจ์ฝ์ฑ ๋ถ์
โข ์ฑ๋ฅ์ ์ํฅ์ ๋ฏธ์น๋ ๋ฐ์ดํฐ ๋๋ ์์ ์ ์ค์ํ ๋ณ๊ฒฝ ์ฌํญ์ ๋ํ ๋ณด๊ณ
โข ๋ฐ์ดํฐ ์ ๋ ฅ ์ญ์ (Sanitization)
โข ๋ฐ์ดํฐ ์ต๋ช ํ๋ฅผ ํตํ ๋ฐ์ดํฐ ํ๋ผ์ด๋ฒ์ ๊ฐํ
โข API ๋ฐ ์ก์ธ์ค ๋ชจ๋ํฐ๋ง
ML Model Operationalization Management ์ดํด
Core Components of ML Model Operationalization Management Solutions
Cognilytica Research, ML Model Management & Operations (โMLOpsโ) 2020- Managing the Machine Learning Model Lifecycle, February 28, 2020
70. 69
Machine Learning Model Development
Machine Learning Model Operationalization Management(MLOps)
Model
Dev.
Data
Prep
Model
Training
Model
Evaluation
Model
Packaging
Model
Discovery
Model
Security
Model
Monitoring
Model
Transparency
Model
Governance
Model
Versioning
ML DEV
ML OPS
Components of ML Development and Ops
ML Model Operationalization Management ์ดํด
Core Components of ML Model Operationalization Management Solutions
Cognilytica Research, ML Model Management & Operations (โMLOpsโ) 2020- Managing the Machine Learning Model Lifecycle, February 28, 2020
71. 70
๊ธฐ๊ณํ์ต ๋ชจ๋ธ ์ด์ ๊ด๋ฆฌ ์๋ฃจ์ ์ ํต์ฌ ๊ตฌ์ฑ์์ - ๊ธฐ๋ฅ๋ค
Model Lifecycle Management
Model development processes & artifacts
Output
models
Training
Data Sets
Test Data
Sets
Validation
Data Sets
Hyperpara-
meters
settings
Validation
Outputs
Ensemble
models
Other key
artifacts
Re-training
pipelines
Model
deployment
Version
control (all
model assets)
configuration,
settings management
ensemble hyperpara
meter
Model Versioning
& Iteration
Model Monitoring - Dashboard
Measure management Visibility
Drift
measurement Metrics
Model Governance
Auditability
Model provenance & auditing
Documents
Training
Testing
Deployment
Audit trail
Recording
artifacts
Metadata
Logging
Training, test,
validation sets used
Accuracy
measurements
Version
history &
model
version
usage
Model bias
measure
monitoring Customizable views of data
Operat-
ions
Data
science
Model
development
LOB
Auditing
Compli-
ance
Data
enginee-
ring
Other
roles
Model Discovery(catalogs/registries/marketplaces)
Curated
Listings of
available
models
Narrative
descriptions
of models
Access control
& cost
mechanisms
for model
usage
Visibility
into model
versions
Ability to
segment lists
Potential for
transfer
learning &
model
extension
Model Security
Access control
mechanisms
Auditing of
model use
& access
Protection of
models, training
data, settings
Vulnerability
analyses
Reporting on
significant
changes to
data
Sanitization
of data
inputs
Enforcing
data
privacy
API &
access
monitoring
Model
Model development
Model
latency
Performan
ce time
Quantity
of requests
Prediction
errors
Accuracy
Performance
measure F1 Other
artifacts
Data(sent
to model)
How
model
used
Model
Drift
Data
Drift
across
slices
user
cohorts
Operational
enviornments
Other
segments
Training data sets
Hyperparameter
settings
Output models
Control
Model
Access
authorization
security
category
User access level
Other factors
Cognilytica Research, ML Model Management & Operations (โMLOpsโ) 2020- Managing the Machine Learning Model Lifecycle, February 28, 2020