Building data science teams to have an impact at scale

Rajiv Krishnamurthy

Because data science plays an important role at Facebook, we have designed a distinctive model for building unified teams with both infrastructure engineers and data scientists.

Overview

The article discusses how Facebook integrates data science into its infrastructure engineering teams to tackle significant challenges and enhance performance at scale. It highlights the collaborative efforts between data scientists and engineers to optimize systems, improve cache efficiency, and develop forecasting models for data center demand.

What You'll Learn

1

How to build machine learning models for predicting content requests

2

Why collaboration between data scientists and engineers is essential for performance optimization

3

How to develop custom forecasting models for data center demand

Prerequisites & Requirements

Understanding of machine learning concepts(optional)
Experience in infrastructure engineering or data science

Key Questions Answered

How do data scientists improve cache efficiency on Facebook's CDN?

Data scientists collaborate with network engineers to build machine learning models that predict whether content will be requested again. This approach optimizes storage on the CDN, reducing inefficiencies by preventing unnecessary retention of rarely accessed data.

What tools do data scientists use to detect performance degradation in apps?

Infrastructure data scientists developed tools that isolate performance issues from trends in device usage and demographics. By providing controlled comparisons, these tools help mobile engineers identify performance regressions that would otherwise go unnoticed.

What forecasting techniques are used for data center strategy?

Data scientists at Facebook developed custom forecasting models, including Bayesian time series models and deep learning models, to accurately predict data center demand. These models account for complex seasonal fluctuations and varying usage trends across devices and locations.

How does Facebook onboard new data scientists?

Facebook's Bootcamp program provides new data scientists with broad exposure to engineering systems and includes mentorship from partner engineering teams. This approach fosters a deep mutual understanding and shared experience, essential for effective collaboration.

Technologies & Tools

Backend

Machine Learning

Used to predict content requests and optimize cache efficiency.

Backend

Bayesian Time Series Models

Developed for accurate long-term forecasting of data center demand.

Backend

Deep Learning Models

Used in conjunction with Bayesian models for improved prediction accuracy.

Key Actionable Insights

1
Integrate data science with engineering teams to enhance system performance.
Collaboration between data scientists and engineers leads to better optimization of resources and improved performance metrics, making it essential for large-scale applications.

2
Utilize machine learning for predictive modeling in infrastructure.
By predicting future content requests, teams can optimize caching strategies, which is crucial for managing resources effectively in a high-demand environment.

3
Develop custom forecasting models tailored to unique data patterns.
Standard forecasting techniques may not suffice; custom models can provide more accurate predictions, aiding in strategic planning for data center expansions.

Common Pitfalls

1

Failing to account for seasonal fluctuations and trends in data sets can lead to inaccurate forecasting.

This often happens when teams rely solely on conventional forecasting methods without adapting to the unique characteristics of their data.

Related Concepts

Machine Learning

Data Center Strategy

Performance Optimization

Collaboration In Engineering