PyCon_Ireland_2022 = {
"Getting Started":
"Feature Stores"
}
# CHAPTER 0
Introduction
Currently 🤓🤓
- Data Science Consultant
- Instructor @ Decoded
Agenda
WHAT?
...is a feature store and what will it enable at my organization, team, client, etc.?
1.
2.
WHY?
...do feature stores matter? Why were they created in the first place? Why should I care?
3.
HOW?
...do we get started with feature stores, and hat options do we have? Build or Buy?
# CHAPTER 0
# CHAPTER 2
What?
We offer a variety of services and plans tailored to business needs of any kind and of any size.
Why?
1.You need to ingest features that are unknown to the client.
2. Data scientists, machine learning engineers, and others, are duplicating features to train different models.
3.You are using hard-to-compute and frequently updated features to constantly train new models.
# CHAPTER 2
1. Model ingests features unknown to the client.
# CHAPTER 1
Why?
Imagine having a centralized mall of data where DB Tables = Stores...
Table 1
Table 2
Table 4
Table 3
Table 5
Table 6
Table 7
Model 1
Recommendation System
User
User
User
User
User
User
User
User
Model 1
Table 1
Table 2
...
Table 3
Simplicity = {
"The Why again but with Code"
}
// This slide uses Auto-Animate to animate between
// two different code blocks
const distanceBetween = ( p1, p2 ) => {
// TODO
}
distanceBetween([10,10], [50,50])# PRESENTING CODE
Get Historical Features and Train
How = "?"
feast init -m feature_store# PRESENTING CODE
Starting a Project
feast init -m feature_store
# output
Creating a new Feast repository in pycon_ireland22/home_credit# PRESENTING CODE
cd feature_store/feature_repo/
touch feature_definitions.pyvim feature_repo
# output
project: home_credit
registry: /path/to/registry.db
provider: local
online_store:
path: /path/to/online_store.db
entity_key_serialization_version: 2# PRESENTING CODE
Adjusting the Configurations
vim feature_repo
# output
project: home_credit
registry: data/registry.db
provider: local
online_store:
path: data/online_store.db
entity_key_serialization_version: 2vim feature_definitions.py# PRESENTING CODE
Defining Entities and Sources
from datetime import timedelta
import pandas as pd
from feast import Entity, FeatureView, Field, FileSource, ValueType
from feast.types import Float32, Int64
from pathlib import Path
data_path = Path().cwd().parents[1]/"data"
usr_id = Entity(
name="user_id", join_keys=["SK_ID_CURR"],
value_type=ValueType.INT64, description="Eeach Customer's ID"
)
where_cc_data_is = FileSource(
path=str(data_path/"processed/cc_df.parquet"),
event_timestamp_column="event_timestamp", name="target_df"
)# PRESENTING CODE
Defining Feature Views
# ...continued
cc_fv = FeatureView(
name="credi_card",
entities=[user_id],
ttl=timedelta(days=365),
schema=[
Field(name="NAME_CONTRACT_TYPE", dtype=Int64),
Field(name="FLAG_OWN_CAR", dtype=Int64),
Field(name="FLAG_OWN_REALTY", dtype=Int64),
Field(name="AMT_INCOME_TOTAL", dtype=Int64),
Field(name="AMT_CREDIT", dtype=Int64),
...
],
source=cc_source
)# CHAPTER 2
Resources
We offer a variety of services and plans tailored to business needs of any kind and of any size.
ACME Offices
| Blog Posts | Tutorials | Websites |
|---|---|---|
| USA | Los Angeles | +1 555 0194 |
| USA | New York | +1 555 0142 |
| Sweden | Stockholm | +46 555 0077 |
| UK | London | +44 555 0211 |
| South Korea | Seoul | +82 555 0138 |
Blog Posts
1. What is a Feature Store?
2. Feature Platform vs Feature Stores
3. Do You Really Need a Feature Store?
4. Feature Stores Explained: Three Common Architectures
5. Feature Stores Explained: Three Common Architectures
6. Feature Stores Explained: Three Common Architectures
7. Feature Stores Explained: Three Common Architectures
8. Feature Stores Explained: Three Common Architectures
Tutorials
1. Crating a Feature Store with Feast P1, P2, P3
Build
Price $Free.99?
You own it
Customised to your needs
Time Consuming
Little-to-no references
Buy - Integrate
Price $199 /mo
Faster roll up (potentially)
Support
Save on headcount
Very new
NEW
Buy - End-to-End
Solutions
Price $599 /mo
Up to five projects
Five designs/project
3-day turnaround
24/7 support