Data Council Blog

Here are 10 recent relevant links for data professionals, from blog posts and tutorials to podcast episodes:

1. PyTorch Lightning: a gentle introduction

Former Data Council speaker Will Falcon published an interesting post on PyTorch Lightning, the lightweight PyTorch wrapper born out of his Ph.D. AI research at NYU CILVR and Facebook AI Research (FAIR). Framed as "a gentle introduction", it includes a side-by-side comparison of building a simple MNIST classifier PyTorch and PyTorch Lightning, in order to illustrate how to refactor one into the other. This is highly recommended reading if you are working on AI/ML research, be it as a professional researcher, student or in production.

Your potential salary as a data engineer heavily depends on where you are based; but cost of living also varies around the world. Wondering where you can actually earn more? Let's take a closer look at the United States, Europe and Asia to compare and benchmark data engineering salaries.

Today's companies aim to be data-driven, but data infrastructure is time intensive and costly to build, maintain, and secure.  A coral is the exoskeleton of a small marine animal that attaches and grows on almost anything. Once it starts growing, it can create large reefs, which support a diverse ecosystem of plants and animals. So what happens if you apply that philosophy to the world of data?

Life comes at you fast. Data even more so ...

When the engineering team at Circonus began to feel the pain of systems at scale, there were some common observability tools that provided them with a firehose of operational time series telemetry. However, managing all that data, yet alone making sense of it, was extremely difficult. And the existing tools they tried for managing time series metrics either didn't give mathematical insight, or fell over at modest workloads. They needed a better solution. So they decided to look into other statistical tooling options that had proven themselves for decades in other industries.

Writing endless data transformations wasn't sustainable for an engineering team handling hundreds of inputs. Here's how Clover Health enabled their business users to help.

It's rare to find an ETL system that's completely static. As organizations change and grow they develop new business requirements. Because of this their data pipelines must change and adapt, ultimately becoming more robust and full-featured. Yet constant development can make already brittle ETL systems seem even more fragile.

Furthermore, systems with large numbers of different types of inputs bring special challenges - building, testing and managing an exploding number of data transformations can become a daunting project for the engineering team. 

The Clover Health ETL system supports hundreds of inputs and more than 500 custom transformations in production as well as a large number of custom connections between their different ETL pipelines. When hearing about the magnitude of the system, one might rightfully wonder, "how does Clover guarantee and maintain data quality across so many different inputs and transforms?"

Exploring the development trajectory of Clover's system makes for a fascinating story; hearing about their data team's successes and pitfalls are illustrative lessons to other engineers as they seek to increase the robustness of their own ETL systems.

Recently at a conference, I had the privilege of demonstrating MemSQL processing over a trillion rows per second on the latest Intel Skylake servers.