A perspective on AWS roadmap and selected recent announcements (primarily based on impressions from Re:Invent 2018, slightly updated in December 2019).
8. Beyond Software: Accelerators
● AWS Inferentia: ML inference chip
○ AWS Inferentia will support the TensorFlow, Apache MXNet, PyTorch, as well as ONNX
○ https://aws.amazon.com/machine-learning/inferentia/
○ Each chip provides hundreds of TOPS (tera operations per second) of inference throughput
8
10. Beyond Public Cloud: Outposts
● Extends AWS on-premises, fully managed. GA Dec 2019.
● VMware Cloud or native AWS (connects to closest AWS region control plane)
● EC2, EBS, ELB, single-tenant runtime services (RDS, EMR, ECS/EKS, etc)
○ NO “serverless” (regional) services (for now) - DynamoDB, Aurora, Kinesis, Lambda, etc
■ Opposite direction that the general trend. S3 will be available in 2020.
○ On-prem=reserved+dedicated?
■ These services are priced by EC2/EBS/network resources - not “pure” pay-per-use
10
11. Beyond Public Cloud: Edge Computing
● Develop in the cloud, run anywhere
○ IoT/network/home appliances (Greengrass)
○ Embedded systems (FreeRTOS)
○ DeepLens
○ RoboMaker
○ DeepRacer
○ Ground Station
○ Loca Zones, Wavelength
11
14. Beyond IaaS: Serverless-first
● After 10+ years, PaaS paradigm finally becomes mainstream
○ Motivation:
■ The more application stack layers you
delegate to a 3rd party (as a managed
service), the more you can focus on
delivering value (business logic)
○ Hypothesis:
■ Majority of newly developed applications
will be “serverless” (PaaS, FaaS, SaaS)
● Compute, Workflows, Data
14
15. Beyond IaaS: Serverless-first - Compute, Workflows
● Lambda
○ IDE integration, Layers, Custom runtimes, Provisioned concurrency, Firecracker
○ WebSockets (API Gateway), Notifications for async invocation (“destinations”)
● EKS on Fargate
● AppSync (GraphQL proxy)
○ “Resolvers” for DynamoDB, Elasticsearch, Aurora, Lambda, etc; pipeline
● Step functions (‘state machine as a service’)
○ Seamless integration with Lambda, DynamoDB, Batch, Fargate, SNS, SQS, Glue, SageMaker
○ Transaction vs high-throughput
● Glue (ETL)
○ Serverless Spark jobs, Python jobs (preprocessing)
15
16. Beyond IaaS: Serverless-first - Data
● Glacier Deep Archive
○ Data can be retrieved within 12 hours
○ $1.01 per TB-mo
● FSx for Windows
● FSx for Lustre
● Aurora Serverless, Data API
● DynamoDB transactions, auto-scaling
● Timestream (time-series DB)
● Blockchain (Quantum Ledger Database)
16