Slide 1
Slide 1 text
From notebook to
production in
Vertex AI
Wessel Huising & Daniel van der Ende
Slide 2
Slide 2 text
● Data & AI Consultancy and Training
● We help organizations be successful with
Data and AI
● Mix of Data Engineers, Data Scientists,
Machine Learning Engineers, Analytics
Engineers & Analytics Translators
● Based in Amsterdam & Eindhoven
● Part of Xebia Data & AI
● The best Payment Service Provider out
there
● Founded in 2004 by Adriaan Mol
● Mission to simplify financial services by
creating world-class products
● Currently active for merchants in European
Economic Area (EEA), Switzerland, and the
United Kingdom
Slide 3
Slide 3 text
The challenge
Slide 4
Slide 4 text
● Low effort
● Highly depending on
one specific Data
Scientist
● Prone to human errors
● Labor intensive
Manual
predictions are
convenient
Data Scientists
≠ Software
Engineers
● Data Scientist tend not
to have traditional
Software Engineering
background
● Tend to lack
understanding of
DevOps
ML models
are something
else
● Different than regular
software artifacts
● There is significant
overlap
Machine Learning models to production is hard
Slide 5
Slide 5 text
The classic ML routine
Slide 6
Slide 6 text
INPUT
DATA
The classic ML routine
Slide 7
Slide 7 text
INPUT
DATA
The classic ML routine
Lack of a centralized data source
Slide 8
Slide 8 text
INPUT
DATA
The classic ML routine
Code not living in a Python package
Slide 9
Slide 9 text
INPUT
DATA
The classic ML routine
Missing lineage tracking of artifacts
Slide 10
Slide 10 text
INPUT
DATA
The classic ML routine
No version control of the code
Slide 11
Slide 11 text
INPUT
DATA
The classic ML routine
No automatic or scheduled
predictions
Slide 12
Slide 12 text
INPUT
DATA
The classic ML routine
Predictions all over the place
Slide 13
Slide 13 text
What is MLOps?
and should you want it?
Slide 14
Slide 14 text
● Bringing a model into
production state
● Reliable & automated service
but comes with extra costs
● MLOps is a End-to-End
process
● Results in great ML software
MLOps = Data + ML + DevOps
Slide 15
Slide 15 text
The big decision
Slide 16
Slide 16 text
No content
Slide 17
Slide 17 text
The big decision
Open Source solutions Managed solutions
Slide 18
Slide 18 text
What is Vertex AI
Slide 19
Slide 19 text
“Build, deploy, and scale ML models faster, with pre-trained and
custom tooling within a unified artificial intelligence platform.”
What does Google tell us?
Slide 20
Slide 20 text
Vertex AI has a few (important) components
The Buzzword Bingo
And more…
Metadata
Endpoints
Models
Pipelines
Workbench Features Datasets
Slide 21
Slide 21 text
Vertex AI has a few (important) components
The Buzzword Bingo
And more…
Metadata
Endpoints
Models
Pipelines
Workbench Features Datasets
Slide 22
Slide 22 text
From notebook to
production
Slide 23
Slide 23 text
No content
Slide 24
Slide 24 text
Goal: Use machine learning to create a model that predicts which passengers survived
the Titanic shipwreck
Slide 25
Slide 25 text
In a Workbench we can do “traditional” Data Science without any limitations
or strings attached.
Each DS has their own Virtual Machine
Each VM:
● Has access to data
● Is persistent
● Can be configured to work with VSCode or PyCharm on your local
machine
● Can have the specs you need/want
Step 1: Let’s explore the
problem! Workbench
Slide 26
Slide 26 text
Time to move our code out of notebooks!
In this step, we’ll:
● Use the Pipeline components that Google provides out of
the box to setup a training pipeline.
● Train our model and output Datasets and Models.
● Deploy our model to an Endpoint so our model is available
for consumption by downstream users.
Step 2: Deploy it as if
you’re Google
Metadata
Models
Pipelines
Datasets
Endpoints
Slide 27
Slide 27 text
Let’s take a step back…
Good Not so good
❌ Automated Train/Test splits are nice,
but also opaque and not very configurable
❌ Where is the model evaluation step?
❌ Everything disappears into one big
“train the model” step, including
preprocessing.
✅ We have an ML Pipeline
✅ It’s all code
✅ It can be scheduled and kicked off
automatically
✅ Everything now is traceable
Slide 28
Slide 28 text
To mitigate the downsides, while keeping the upsides, we use the
Kubeflow API
It integrates well with Vertex AI and can be easily customized
Mollievert is our package with customized components to simplify and
clarify our ML Pipelines
Step 3: Now Mollie-fy it!
Metadata
Models
Pipelines
Datasets
Endpoints
Slide 29
Slide 29 text
Wrapping up
Mollie’s ML Platform enables robust MLOps practices by:
● Making it easier to go to production without lowering the bar
● Empowering Data Scientists and ML Engineers with tooling
● Defining a ‘Golden Path’ to production, but allowing customization if desired.
Slide 30
Slide 30 text
No content