I attended the AWS re:Invent at Las Vegas the week after Thanksgiving. It was a great week there, learning AWS technologies while also having fun. The keynotes given by Andy Jassy and Werner Vogels were always exciting: they covered important new services that were launched during this event. Other sessions shared best practices in using AWS services and infrastructure. They also had music bands playing live music at the site and provided great lunches and snacks during the morning and afternoon breaks. Besides, we can come and pick up a swag every day. This year, more than 50,000 people attended the conference and you can see lots of people walking around. This is by far the largest conference I have ever attended and I was amazed by how well they managed to organize it.
AWS annual revenue has grown to 27 billions US dollars, with 47% YoY growth. From the most recent Gartner report, AWS shares 52% market share for the IaaS market. Overall, AWS has ~140 services with 11 database services.
- Support for ARM processors: A1 instances. Good for scale-out application and bring cost saving (~45%)
- V100 GPUs: P3 instances with 8 NVIDIA Tesla V100 GPUs.
- Elastic graphic / Elastic Inference: add GPU on-demand to normal EC2 instances, only for graphic or inference tasks
- 100 Gpbs NIC: c5n instances
Simple Storage Service (S3)
- Intelligent tiering: automatically migrate objects from S3 standard to S3 infrequent access, to save customer S3 storage cost
- Glacier deep archive: a tape replacement solution. Even cheaper than Glacier.
- S3 batch operations: apply the same operation for billions of objects. To use this feature, we first provide a list of objects, either by using the S3 inventory report or generating the list by ourselves. Then, we select a desired action from pre-populated menu of options. S3 will do the job, send the progress and notification when the job is completed.
Elastic File Service (EFS)
- A new class: Infrequent Access. Save cost (~85%).
- FSx for Windows file server: native windows FS; fully managed by AWS; can support 10GB/s throughput
- FSx for Lustre: fully managed; for HPC users
Data Migration Service
- AWS DataSync Service: migrate data between on-premise NFS server and S3/EFS. Need to deply a VM at on-premise datacenter to run the DataSync agent. Mentioned use of compression and multi-part PUTs to improve the transfer performance. Audiences also asked the support of Object store/SMB/HDFS as another source.
New Database services
- AWS Quantum Ledger Database (QLDB): to store immutable, and cryptographically verifiable transaction logs where there is a central trusted authority.
- AWS Timestream: time series database
- IDEs: AWS Toolkits for Visual Studio Code, IntelliJ, and PyCharm
- Supports custom runtime for any Linux compatible language runtime
- Supports Ruby, C++, and Rust
- Supports for Php, cobol, Erlang, and elixir are available from AWS partners
- Lambda layers: allow multiple lambda functions to share libraries
- AWS serverless application repository: store, share, and deploy serverless applications
- Step function service integration: from a step function, one can now talk to Lambda, Batch, DynamoDB, ECS/Fargate, SNS, SQS, Glue, and SageMaker.
- Application load balancer (ALB) support for Lambda: one can trigger lambda functions directly from ALB.
- Firecracker: AWS's lightweight virtualization technology based on KVM. It is the underling container technology powering AWS Lambda and AWS Fargate. Each VM takes only 5 MB of memory. They claimed to achieve startup time as short as 125 ms and can provision 150 microVMs per min/host. They make it open source.
Three layers of AI software stack (from top to bottom)
- AI Services: TextToSpeed, etc
- ML Services: Sagemaker
- ML frameworks: TensorFlow, MXNet, and pyTorch
- SageMaker Groundtruth: provides auto labelling or human to add tags to datasets; For auto labelling, one can set the confidence level: it will only set labels when the confidence from the algorithm is higher than the threshold. human label: mechanical trunk; third party; or your friends.
- AWS Marketplace for machine learning: a place to pick up/sell machine learning packages.
- SageMaker Reforcement Learning: provides RL models.
- SageMaker Neo: open-source compiler to compile a machine model for different architectures (TPU/GPU/Intel CPU/ARM, etc.)
- AWS Inferentia: a machine learning inference chip designed to deliver high performance at low cost; A customized chip by AWS optimized for inference;
Bring AWS hardware to your datacenters; can run AWS software to get consistent management/use experience; AWS's strategy for hybrid cloud; can also run VMWare Cloud on outposts for VMWare customers.
Other new services
- AWS control tower: provides best practice guides/blueprints; prescribed guides for how to best use AWS infrastruture
- AWS security hub
- AWS Lake formation: create data lakes
- AWS Managed Blockchain
- AWS Managed Kafka
- AWS forcast: time series forecasting
- AWS personalize: real-time personalization and recommendation