María de la Fuente (Solutions Architect Manager for IMEA) @ Databricks
While most companies understand the value creation of leveraging data and are taking on board an AI strategy, only 13% of the data science projects make it to production successfully.
Besides the well-known skills gap in the market, we need to level up our end-to-end approach and cover all aspects involved when working with AI.
In this session, we will discuss the main obstacles to overcome and how we can avoid the major pitfalls to ensure our data science journey becomes successful.
Why do the majority of Data Science projects never make it to production?
1. Why the majority of Data
Science projects never
make it to production?
2. María de la Fuente
Solutions Architect Manager – Israel Middle East &Africa
@Databricks
María de la Fuente | LinkedIn
3. AI is poised to
change the world
$3.9T
Projected Business value
creation by AI in 2022
And most
leaders agree
83%
CEOs say AI is a
strategic priority
But AI doesn’t
make it out the door
at most companies
Of Data Science initiatives
never make it to production
87%
4. Q: Why are these projects struggling?
A: It is mainly because of reliability,
performance and lack of ML end-to-
end tracking mechanism.
7. A typical Machine Learning workflow
Data
Preparation
Feature
Engineering
Model
Training
Model
Evaluation
Model
Deployment
Model
Tuning
Model
Consumption
Data
Ingestion
Users
Data & ML
Engineers
10. Data Lake
The data is not ready for data science & ML
The majority of these projects are failing due to
unreliable data!
Data Science & ML
• Recommendation Engines
• Risk, Fraud Detection
• IoT & Predictive Maintenance
• Genomics & DNA Sequencing
11. ML Lifecycle is Manual, Inconsistent
and Disconnected
● Ad hoc approach to track
experiments
● Very hard to reproduce
experiments
Prep Data
● Multiple tightly coupled
deployment options
● Different monitoring approach
for each framework
Build Model Deploy Model
● Low level integrations for
Data and ML
● Difficult to track data used
for a model
12. Nothing last forever
”Change is the only constant in life¨-Heraclitus, Greek Philosopher
12
One of the main assumptions when creating a model is that future data will
be similar to past data used to build the model
HOWEVER,
Models exists in a dynamic and continually changing environment, when
this environment change, the performance of the model will change too
13. This means…Model Drifting is expected!
13
ML Models will lose their predictive power over time
CONCEPT DRIFT properties of the dependent variable(s) change(s)
DATA DRIFT properties of the independent variable(s) change(s)
14. ML Lifecycle and Challenges
Delta
Tuning Model Mgmt
Raw Data ETL Train
Featurize Score/Serve
Batch + Realtime
Monitor
Alert, Debug
Deploy
AutoML,
Hyper-p. search
Experiment
Tracking
Remote Cloud
Execution
Project Mgmt
(scale teams)
Model
Exchange
Data
Drift
Model
Drift
Orchestration
(Airflow, Jobs)
A/B
Testing
CI/CD/Jenkins
push to prod
Feature
Repository
Lifecycle
mgmt.
Retrain
Update Features
Production Logs
Zoo of Ecosystem Frameworks
Collaboration Scale Governance
16. MLOps: What, why, how?
WHAT: Set of practices for
collaboration and communication
between data scientists and
operations professionals
WHY: Aims to improve the delivery of
machine learning models by
combining the processes of design,
development, testing, and delivery
into a singular process.
● Shortening development cycles, and as a
result, decreasing time to market
● Improving collaboration between teams
across all levels of technical expertise
● Increasing reliability, performance,
scalability, and security of ML systems
● Streamlining operational and governance
processes
● Increasing return on investment of ML
projects
17. MLOps vs DevOps
SAME SAME…
when it comes to continuous integration of source control, unit testing, integration testing, and
continuous delivery of the software module or the package
…BUT DIFFERENT
Continuous Integration (CI) is no longer only about testing and validating code and components, but also
testing and validating data, data schemas, and models
Continuous Deployment (CD) is no longer about a single software package or service, but a system (an ML
training pipeline) that should automatically deploy another service (model prediction service) or roll back
changes from a model
Continuous Testing (CT) is a new property, unique to ML systems, that’s concerned with automatically
retraining and serving the models
18. Tactics for
Successful & Scalable ML in production
● Align business needs & ML Objectives
● Involve right personas
● Lean into the cloud
● Break the silos & support cross-colaboration
● Architect with operations in mind
● Invest & Leverage MLOps
20. End-to-End Data Science and ML on
AutoML
End-to-End ML Lifecycle
ML Runtime and
Environments
Batch
Scoring
Online
Serving
Data Science Workspace
Prep Data Build Model Deploy/Monitor Model
Open,
pluggable
architecture
21. High level Architecture
Unified Analytics Platform
Data Science, Model Training, Test and Selection
APIs
Jobs
Models
Notebooks
Dashboards
ML Runtime
Databricks Runtime
BI
Tool
Connectors
Model Deployment& Monitoring
to the cloud...
to the edge...
ETL / Data Processing
Bronze
Gold
DB
Connect
Tracking Projects Models
End to end ML lifecycle
Registry
Connectors and APIs
for a wide variety of
differentsources...
File
DB/DW