Back to the Index

Pandas UDF

What are pandas UDFs?

‍Pandas UDFs (User-Defined Functions) are functions that allow users to perform feature engineering (or any custom transformations) on a Pandas DataFrame using PySpark. Pandas UDFs enable high performance feature engineering using Python functions on data stored in a PySpark DataFrame.

PySpark provides built-in functions for common feature engineering operations, such as filtering, grouping, and aggregating data. However, if you need to write your own custom feature engineering code, you can instead write a Pandas UDF. PySpark will convert your Spark DataFrame into a Pandas DataFrame, apply the user-defined function on the Pandas DataFrame, and then convert the output back into a PySpark DataFrame - without any serialization/deserialization costs as PyArrow is used transfer data between PySpark and Pandas.

‍How do I write a pandas UDF?

You write your UDF and add the pandas_udf decorator provided by PySpark. The decorator allows you to specify the input and output types of the UDF and provides other configuration options. Once the UDF is defined, it can be applied to a PySpark DataFrame using the pandas_udf function.

The case for pandas UDFs for feature engineering:

  1. Performance: Pandas UDFs enable vectorized computation, which leads to improved performance compared to Python UDFs. Vectorized operations in Pandas are implemented using low-level C or Cython code, which significantly reduces the overhead associated with Python's interpreted nature.
  2. Flexibility: Pandas UDFs can be used in various environments, including Spark feature pipelines, Spark batch inference, and Python-based online inference pipelines (as supported by on-demand features in the Hopsworks feature store). This versatility makes it easier to integrate Pandas UDFs across different stages of a machine learning pipeline.
  3. Familiarity: For data scientists and engineers who are already familiar with the Pandas library, working with Pandas UDFs is a natural extension of their existing skill set. This reduces the learning curve and makes it easier to implement custom feature transformations.
  4. Expressiveness: Pandas provides a rich set of data manipulation functions, allowing users to implement complex feature transformations with just a few lines of code. This expressiveness can simplify the feature engineering process and reduce the time spent on implementing custom functions.

Interested for more?

Does this content look outdated? If you are interested in helping us maintain this, feel free to contact us.
P
AI Lakehouse
AI Pipeline
Auto-regressive Models
AutoML
P
Backfill features
Backfill training data
Backpressure for feature stores
Batch Inference Pipeline
P
CI/CD for MLOps
Compound AI Systems
Context Window for LLMs
P
DAG Processing Model
Data Compatibility
Data Contract
Data Lakehouse
Data Leakage
Data Modeling
Data Partitioning
Data Pipelines
Data Quality
Data Transformation
Data Type (for features)
Data Validation (for features)
Data Warehouse
Data-Centric ML
Dimensional Modeling and Feature Stores
Downstream
Query Engines
P
ELT
ETL
Embedding
Encoding (for Features)
Entity
P
Feature
Feature Data
Feature Engineering
Feature Freshness
Feature Function
Feature Groups
Feature Logic
Feature Monitoring
Feature Pipeline
Feature Platform
Feature Reuse
Feature Selection
Feature Service
Feature Store
Feature Type
Feature Value
Feature Vector
Feature View
Filtering
Fine-Tuning LLMs
P
Generative AI
Gradient Accumulation
Grouped Query Attention
P
Hallucinations in LLMs
Hyperparameter
Hyperparameter Tuning
P
Idempotent Machine Learning Pipelines
In Context Learning (ICL)
Inference Data
Inference Logs
Inference Pipeline
Instruction Datasets for Fine-Tuning LLMs
P
LLM Code Interpreter
LLM Temperature
LLMOps
LLMs - Large Language Models
Lagged features
LangChain
Latent Space
P
ML
ML Artifacts (ML Assets)
MLOps
MLOps Platform
MVPS
Machine Learning Infrastructure
Machine Learning Logs
Machine Learning Observability
Machine Learning Pipeline
Machine Learning Systems
Model Architecture
Model Bias
Model Deployment
Model Development
Model Evaluation (Model Validation)
Model Governance
Model Inference
Model Interpretability
Model Monitoring
Model Performance
P
Natural Language Processing (NLP)
P
Offline Store
On-Demand Features
On-Demand Transformation
Online Inference Pipeline
Online Store
Online-Offline Feature Skew
Online-Offline Feature Store Consistency
Orchestration
P
KServe
PagedAttention
Parameter-Efficient Fine-Tuning (PEFT) of LLMs
Point-in-Time Correct Joins
Precomputed Features
Prompt Engineering
Prompt Store
Prompt Tuning
Python UDF
P
RLHF - Reinforcement Learning from Human Feedback
Real-Time Machine Learning
Recommender System
Representation Learning
Retrieval Augmented Generation (RAG) for LLMs
RoPE Scaling
P
SQL UDF in Python
Sample Packing
Schema
Similarity Search
Skew
Splitting Training Data
Streaming Feature Pipeline
Streaming Inference Pipeline
P
Test Set
Theory-of-Mind Tasks
Time travel (for features)
Train (Training) Set
Training Data
Training Pipeline
Training-Inference Skew
Transformation
Two-Tower Embedding Model
Types of Machine Learning
P
Upstream
P
Validation Set
Vector Database
Versioning (of ML Artifacts)
vLLM