The 2024 edition of re:Invent has just ended in Las Vegas. As anticipated, AI was a key focus of the conference, with Amazon Nova and a new version of Sagemaker among the most significant highlights. However, the announcement that generated the most excitement in the community was the preview of Amazon Aurora DSQL, a serverless, distributed SQL database with active-active high availability.
The conference marked Matt Garman's debut keynote as AWS CEO, while Andy Jassy, now Amazon CEO, returned to the stage in Vegas to introduce the Amazon Nova foundation models. Notably, the release pattern this year was unusual, with no announcements made during Werner Vogels' keynote.
Starting with why he joined Amazon 20 years ago, Vogels instead focused on six lessons on "simplexity". Centering on distributed systems, he analyzed timestamps and their role in solving synchronization challenges in building distributed databases. On the same day, Vogels released his tech predictions for 2025 and beyond. Rajesh Pandey, principal engineer at AWS, comments:
Werner's keynote was truly thought-provoking. 'Simplexity' might be the most impactful concept I'm taking away from this re:Invent (...) His framework for managing complexity while maintaining operational simplicity is invaluable.
Below is a review of this year's main announcements impacting computing, databases, storage, machine learning, and development. InfoQ will cover these and other updates as separate news pieces soon.
Compute
Garman highlighted the widespread adoption of Graviton, Amazon's ARM processor. Without providing specific numbers, he noted that Graviton instances now account for more EC2 usage than the total EC2 usage in 2019, just five years ago. However, no new major ARM version was released this year, with Graviton4 remaining the latest version.
Designed for AI/ML training and inference, Trainium2 instances are now generally available, with the cloud provider claiming that Apple, Adobe, and Qualcomm have already achieved their best price performance for training using them. Anthony Charbonnier, senior account manager at AWS, highlights:
While the name suggests "just" an upgrade, it's actually way more than that. Each TRN2 chip delivers up to 1.3 petaflop (dense FP8) with 96GByte of HBM3e memory capacity, and supports 2.9 TB/second of HBM bandwidth. This is huge! In addition to the 16 Trainium2 chips, each Trn2 instance features 192 vCPUs, 2 TiB of memory, and 3.2 Tbps of Elastic Fabric Adapter (EFA) v3.
Connected with NeuroLink, the EC2 Trm2 UltraServers provide up to 64 Tranium2 chips and up to 83.2 FP8 petaflops.
Databases
Aurora DSQL introduced a serverless, distributed SQL database claiming unlimited scale, 99.99% single-region availability, and 99.999% multi-region active-active availability. By separating the transaction logic from the storage layer, and processing transactions with strong consistency, Amazon claims that the new PostgreSQL-compatible database delivers 4x faster reads and writes compared to Google Cloud Spanner.
Source: AWS blog
Rehan Van Der Merwe, cloud engineer at DataChef and AWS Hero, summarizes how DSQL compares to other managed options and writes:
Aurora DSQL promises an OLTP database that scales to zero but also has OLAP features for powerful querying. It's like having a car that can go from 0 to 60 in 3 seconds, but also has a trunk big enough to fit a couch.
Not everyone in the community agrees. Similar to Aurora DSQL, Dynamo DB introduced the preview of multi-region strong consistency for global tables. MemoryDB multi-region is now generally available, providing a cross-region, Valkey and Redis compatible in-memory database.
Storage
"Is S3 becoming a data lakehouse?" was a common sentiment among the community when the new Amazon S3 Tables bucket was announced during the first keynote. Due to the huge growth of tabular data in S3, with the de-facto open standard Apache Parquet and customers having billions of Parquet files, Apache Iceberg has become the way to access the data over the years.
AWS' answer is a storage-optimized solution for analytics workloads, claiming up to 3x faster query performance and up to 10x higher transactions per second for Apache Iceberg tables compared to standard S3 storage. Rahul Chakraborty, engineering manager for data & analytics at Roche, writes:
AWS S3 Tables represent a significant advancement in cloud storage, blending S3's flexibility with Apache Iceberg's structured data management capabilities. (...) It will be worth watching how Databricks and Snowflake respond to this new development!
As metadata makes data more discoverable, Amazon S3 Metadata is a new option now in preview that automatically updates object metadata on S3. The new feature is designed to curate, identify, and use S3 data for business analytics and real-time inference applications. Read more on InfoQ.
With AWS sunsetting many Snow portable devices, the new AWS Data Transfer Terminals, currently available in Los Angeles and New York only, allow customers to upload to the cloud faster, relying on secure physical locations with high-throughput connections. The upload speed is up to 400 Gbps, and pricing is per hour. Paul Power, data platform lead at Aer Lingus, comments:
An interesting offering somewhere between accelerated upload and AWS Snowball. I remember as an intern bringing backup tapes to an offsite safety deposit box. Some things change, some things stay the same.
Machine Learning
Machine learning and artificial intelligence were major topics at the conference, with over 800 talks on the subject. Andy Jassy's vision, as with relational databases or analytics, is that there's not going to be a single LLM that will "rule the world." He announced Amazon Nova, a set of new foundation models available only on Amazon Bedrock, claiming "industry-leading price-performance."
Nova Micro, Nova Lite, and Nova Pro are already generally available models, accepting text, image, or video inputs and generating text outputs. Nova Premier, the most capable of Amazon's multimodal models, is expected in Q1 2025. Nova Canvas, an image generation model, and Nova Reel, a video generation model, were also announced but are not yet available. Simon Willison, creator of Datasette and co-creator of the Django Web Framework, completed an early analysis of the new LLMs, writing:
My initial impressions from trying out the models are that they're mainly competitive with the Google Gemini family. They are extremely inexpensive — Nova Micro slightly undercuts even the previously cheapest model Gemini 1.5 Flash-8B — can handle quite a large context and the two larger models can handle images, video, and PDFs. (...) With this release, I think Amazon may have earned a spot among the top tier of model providers.
Bedrock model distillation, which transfers knowledge from a large, complex model to a smaller one, Bedrock Automated Reasoning checks, which prevent factual errors from LLM hallucinations, and Bedrock multi-agent collaboration were the other significant additions in the generative AI space.
The cloud provider unveiled what it called the "next generation" of Amazon SageMaker, the cloud-based machine-learning platform, aiming to unify data engineering, analytics, and generative AI in a single studio with enhanced capabilities. Additionally, the new SageMaker Lakehouse integrates S3 data lakes and Redshift warehouses, enabling unified analytics and AI/ML on a single data copy through open Apache Iceberg APIs.
Development
Amazon Q, the generative AI–powered assistant for software development, received a revamp a year after its initial announcement, with agent capabilities including generating documentation, code reviews, and unit tests. While the earliest features were primarily Java-focused, this year's announcements target transformation capabilities for .NET, mainframe, and VMware workloads.
Source: AWS blog
Recap
How did this edition compare to past ones? While there was plenty of excitement and promising new services, Amazon used the keynotes to celebrate the 10th anniversary of AWS Lambda, Amazon ECS, and Amazon Aurora. These three major pillars of the largest hyper-scale platform were first unveiled at the 2014 conference, a hard feat to match.
The AWS editorial team summarized their top announcements from the AWS re:Invent 2024 in an article. Luc van Donkersgoed, principal engineer at PostNL, provided curated re:Invent feeds at aws-news.com. Ran Isenberg, principal software architect at CyberArk, wrote a summary of serverless takeaways.
All the keynotes and many sessions are already freely available on the dedicated YouTube channel.
