Non-italicised points represent notes taken from Google’s official case study documentation, with keywords highlighted in bold. Points in italics represent additional insights and possible solutions.
Overview
- Global sports league for helicopter racing
- World championship and regional events
- Need to cater for global and regional traffic
- Points of presence in many regions—close to end users
- CDN
- Offers paid service to stream races with live telemetry and predictions throughout
- Uptime important
- Real-time analytics
- Wish to migrate to a new platform
- Currently on AWS or Azure—Storage Transfer Service
- Increase use of AI/ML services
- More serving of real time and recorded content closer to users in emerging markets
Existing Technical Environment
- Public cloud-first company
- Video recorded/edited at racetracks
- Encoding and transcoding performed in the cloud
- VMs
- Enterprise connectivity and local compute in truck mounted DCs
- Mobile DCs—dedicated interconnect not an option
- Race prediction service hosted on cloud
- Content stored in object storage
- Migration—Storage Transfer Service
- Video encoding/transcoding performed on VMs for each job
- Lift/shift with Migrate for Compute Engine
- GPUs
- Machine types:
- Tau T2D (option for scale-out architecture, no GPU support)
- C2, C2D (ultra-high performance, no GPU support)
- A2 (optimized for GPU usage)
- Transcoder API—managed service
- Race predictions via TensorFlow models running on VMs
Business Requirements
- Expand predictive capabilities
- Reduce latency in emerging markets
- Cloud CDN
- Global Load Balancing
- Regional Managed Instance Groups (MIGs)
- Regional Cloud Storage buckets
- Expose models to partners
- Private connectivity – VPN
- API Gateway / Apigee
- Increase predictive capabilities (race results, mechanical failures, crowd sentiment)
- Previous race results—batch data analysis
- Sentiment analysis:
- Natural Language API—predefined categories inc. sentiment analysis
- AutoML Natural Language—custom categories
- Vertex AI—Video Intelligence API
- Increase telemetry and insights
- Cloud IoT Core
- Cloud Bigtable, Dataflow, BigQuery, Looker
- Measure engagement with new predictions
- Cloud Firestore
- AI/ML models—Vertex AI
- Enhance global availability and quality of broadcasts
- Higher quality video encoding—high spec VMs (Tau, GPUs)
- Increase number of concurrent viewers
- Increase capacity
- Dynamic horizontal scaling—Autoscaling Groups
- Minimize operational complexity
- Managed services where possible
- Reduce number of platforms
- Ensure regulatory compliance
- New revenue streams e.g. merchandise
- Online store—SaaS
- Cloud Run, App Engine, Cloud Firestore
Technical Requirements
- Maintain/increase prediction throughput/accuracy
- TPUs, Vertex AI
- Reduce viewer latency
- Content closer to viewers
- Multi-regional GCS buckets
- Cloud CDN
- Global Load Balancers
- HSL protocol—HTTP Live Streaming, Apple protocol for on-demand auto/video
- Increase transcoding performance
- Vertical and horizontal scaling
- Transcoder API—batch processing
- Real-time analytics of viewer consumption
- Data mart—processing large volumes of race data
- AutoML Vision
- Vertex AI
- BigQuery ML
Potential Solution Design
