AWS / reInvent 2020 / AWS Partner Network / devops /

re:Invent 2020 Week Two re:Cap

Aaron Walker

14 December 2020

3 Minute Read

With the second week now behind us, I thought I put together my 2nd post on the highlights from week 2, which included the first-ever Machine Learning Keynote by Swami Sivasubramanian and the Infrastructure Keynote by Peter DeSantis.

Machine Learning Keynote - Swami Sivasubramanian

The fact that AWS split out ML/AI related announcements into it's own keynote is definite proof that they are doubling down on the investment in this space. Swami described their commitment to providing a very broad and deep set of ML/AI offerings. He then went on to highlight that they have released over 250 new features across the suite of AWS ML/AI related services since re:Invent 2019.

The ML/AI keynote was also packed full of new product launches and updates and here is a link to the full list

Here's my highlight list:

  • Introducing Amazon SageMaker Pipelines - The first purpose built CI/CD service for machine learning
  • Introducing Amazon SageMaker Data Wrangler – The fastest and easiest way to prepare data for machine learning
  • Announcing new capabilities for Amazon SageMaker Debugger with real-time monitoring of system resources and profiling training jobs
  • Announcing Amazon Lookout for Metrics

Introducing Amazon SageMaker Pipelines - The first purpose built CI/CD service for machine learning

Many of the teams I've worked with, who are building ML/AI features into products, struggle with the disconnect between how they build and deploy application code and how they do that for training and deploying machine learning models. CI/CD processes for application code are well understood and there are plenty of processes and tools to help. When it comes to building and deploying ML models every team has to solve essentially the same problems. They often start with the CI/CD tools they know but they often fall short due to specialised requirements of training and deploying ML models. So the introduction of Amazon SageMaker Pipelines looks to close the gap between data science and engineering teams, and application development teams. Providing a way to collaborate seamlessly on ML projects and streamline building, automating and scaling of end-to-end ML workflows.

Introducing Amazon SageMaker Data Wrangler – The fastest and easiest way to prepare data for machine learning

Good training data is probably the most critical factor when building ML models. Teams can spend weeks or months aggregating and preparing data from different sources. Typically, 90% of the time and cost of ML projects is spent just preparing data. Amazon SageMaker Data Wrangler is designed to reduce the time it takes to aggregate and prepare data. It provides a single interface to complete each step of the data preparation workflow, including data selection, cleansing, exploration and visualization.

Amazon SageMaker Data Wrangler provides a single click data selection tool that allows you to import data from different sources such as Amazon S3, Amazon Athena, Amazon Redshift, AWS Lake Formation and Amazon SageMaker Feature Store. It is also fully integrated with Amazon SageMaker Pipelines allowing you to build automated ML workflows.

Announcing new capabilities for Amazon SageMaker Debugger with real-time monitoring of system resources and profiling training jobs

The announcement of Amazon SageMaker Debugger continues the theme of AWS making ML capabilities more accessible to developers. When Amazon SageMaker was first announced it was primarily used by data scientists who had a good understanding of ML concepts and just needed a way to build and train models at scale. As usage of SageMaker increased, developers found the lack of transparency into the training process challenging. The new Amazon SageMaker Debugger capability provides richer insights into real-time training metrics such as learning gradients and weights. It also automatically monitors system resources such as CPU, GPU, network, I/O and memory providing a complete resource utilization view of training jobs.

Announcing Amazon Lookout for Metrics

The announcement of Amazon Lookout for Metrics was a little puzzling after last weeks announcement of DevOps Guru. But after digging a little deeper, Amazon Lookout for Metrics focuses on anomaly detection of business health metrics. It automatically connects to 19 different data sources which include AWS services such as S3, CloudWatch, Redshift and RDS but also supports SaaS applications such as Salesforce, Marketo and Amplitude. It detects anomalies, claiming high accuracy, by using the ML algorithm that’s best suited for your data. It allows you to group related anomalies and rank them by severity. It can then alert you and provide actionable results highlighting potential root causes.

Infrastructure Keynote - Peter DeSantis

Peter DeSantis is the SVP, AWS Infrastructure & Support. He heads up anything Infrastructure related at AWS and is a bit of an AWS institution. Typically there aren't many announcements during Peter's keynotes but they generally follow a theme and provide insights into how AWS can do what it does at Amazon scale. This year he focused on sustainability and AWS' commitment to 100% renewable energy by 2025, which is 5 years ahead of their original target.

He started talking about AWS data center power management and complexities ensuring continuous power delivery to the data-center. He highlighted that AWS now uses their own software to control data-center generator and switch gear. They have simplified the overall design of the UPS by using in-rack pluggable UPS devices that are also running their own software. All of this is aimed to reach their sustainability goal and helps them reduce the power conversion losses by 35%.

Peter then did a deep-dive into the AWS Custom Silicon and talked about how Nitro has enabled them to deliver all the new innovations in EC2, including showing a nitro card connected to a Mac mini which is how they are delivering the new EC2 Mac instances. They moved onto Graviton2 (AWS custom ARM chips) and customers are saving up to 40% on compute costs by switching. He highlighted how Graviton2 consumes considerably less power which helps them reduce overall power consumption.

Finally, he wrapped up by going over the renewable energy projects they are working on and how they have added more renewable energy in the last 12 months since they started this incentive in 2014.

This concluded the second week of re:Invent. One more week and one more keynote to go.



More Blog Posts