Overview
The article introduces Michelangelo, Uber's internal machine learning platform designed to democratize machine learning and streamline the process of building, deploying, and operating ML solutions at scale. It discusses the platform's architecture, workflow, use cases, and future enhancements.
What You'll Learn
1
How to build and deploy machine learning models using Michelangelo
2
Why standardizing ML workflows is crucial for scalability
3
When to leverage batch vs. online prediction services
Prerequisites & Requirements
- Understanding of machine learning concepts and workflows
- Familiarity with data processing tools like Spark and Hadoop(optional)
Key Questions Answered
What is Michelangelo and how does it function?
Michelangelo is Uber's internal machine learning platform that simplifies the process of building, deploying, and managing machine learning models at scale. It integrates various components for data management, model training, evaluation, and deployment, enabling teams to leverage machine learning effectively across the organization.
How does UberEATS utilize Michelangelo for delivery time predictions?
UberEATS employs Michelangelo to predict meal delivery times using gradient boosted decision tree regression models. These models incorporate various features, including historical data and real-time metrics, to provide accurate delivery time estimates at different stages of the order process.
What are the key components of Michelangelo's architecture?
Michelangelo's architecture includes a mix of open source components like HDFS, Spark, and Cassandra, along with in-house tools for data management and model serving. This architecture supports both online and offline predictions, ensuring scalability and efficiency in processing data.
What is the machine learning workflow in Michelangelo?
The machine learning workflow in Michelangelo consists of six key steps: managing data, training models, evaluating models, deploying models, making predictions, and monitoring predictions. This structured approach ensures that machine learning processes are standardized and reproducible across teams.
Key Statistics & Figures
Predictions served per second
250,000
This figure represents the capacity of the highest traffic models deployed using Michelangelo.
P95 latency for models not needing Cassandra features
less than 5 milliseconds
This latency applies to models that can operate without accessing additional features from the Cassandra store.
P95 latency for models requiring Cassandra features
less than 10 milliseconds
This latency is observed for models that depend on features stored in Cassandra.
Number of features in Feature Store
10,000
These features are utilized across various machine learning projects within Uber.
Technologies & Tools
Some links below are affiliate links. We may earn a commission if you make a purchase.
Data Storage
Hdfs
Used for storing Uber's transactional and log data.
Data Processing
Spark
Facilitates batch processing and model training.
Database
Cassandra
Stores features for online model predictions.
Streaming
Kafka
Aggregates logged messages from Uber's services.
Machine Learning
Tensorflow
Utilized for building and training deep learning models.
Key Actionable Insights
1Standardizing machine learning workflows can significantly enhance productivity and model performance across teams.By implementing a unified workflow, teams can avoid common pitfalls such as inconsistent data handling and model deployment issues, leading to more reliable outcomes.
2Utilizing a shared feature store can reduce redundancy and improve data quality in machine learning projects.When teams share features, they can leverage existing data insights, which accelerates the development process and fosters collaboration.
3Monitoring model predictions is essential for maintaining accuracy over time.By logging predictions and comparing them to actual outcomes, teams can identify model drift and take corrective actions to ensure continued performance.
Common Pitfalls
1
Failing to monitor model predictions can lead to undetected model drift.
Without ongoing monitoring, models may become less accurate over time as the underlying data distribution changes, resulting in poor performance.
2
Not standardizing data pipelines can result in inconsistent feature generation.
When teams use different methods for feature generation, it can lead to discrepancies in model training and prediction, ultimately affecting the reliability of outcomes.