Helicopter Racing League (GCP Case Study — Architectural Solution)

In this series of GCP enterprise solutions, I will discuss another case study, the Helicopter racing league. Below you can see the requirements for this case study.

[embed]https://services.google.com/fh/files/blogs/master_case_study_helicopter_racing_league.pdf

This use case presents different challenges, primarily using Anthos, Vertex AI and other video streaming APIs. If we read the executive statement, based on fan engagement, the organisation is interested in enhanced video streaming and real-time predictions.

Given the above requirements, I have plotted this solution which fulfils most of the expectations.

Let's bifurcate the solution to understand the rationale behind choosing GCP products and design architecture.

Live Streaming & Storage

HRL offers a paid service to stream the races all over the world with live telemetry and predictions throughout each race. Additionally, as new fans engage with the sport, particularly in emerging regions, they want to move the serving of their content, both real-time and recorded, closer to their users. Video recording and editing is performed at the race tracks, and the content is encoded and transcoded, where needed, in the cloud. Enterprise-grade connectivity and local compute is provided by truck-mounted mobile data centers.

For video streaming solutions, google has a built-in API i.e. Live Streaming API. This API does both encoding and transcoding of the videos. Standard video encoding with support for input streams like RTMP, and SRT using FFMPEG encoding transcoding for mobile, web etc to formats like Apple HLS with fMP4 and MPEG2-TS segments, MPEG-DASH with fMP4 segments. This API has out-of-the-box integration with storage to google cloud storage.

https://cloud.google.com/livestream/docs https://cloud.google.com/storage

Video Transcoding & Delivery

Additionally, as new fans engage with the sport, particularly in emerging regions, they want to move the serving of their content, both real-time and recorded, closer to their users. Video encoding and transcoding is performed on VMs created for each job. Minimize operational complexity. Increase transcoding performance. Create a data mart to enable the processing of large volumes of race data. Reduce viewer latency. Enhance global availability and quality of the broadcasts. Increase the number of concurrent viewers.

One thing to consider is, live streaming API does transcode as well, however, for live streams that transcoded videos can be persisted on cloud storage. However, the requirement also states that HRL wants to create a datamart, which means customers can stream old recorded videos and want to transcode them on demand. Google offers an API integration for this as well, transcoder API.

Another important requirement for this solution is reducing latency. GCP provides a cloud CDN to allow loading content faster and closer to users with a multi-region HTTPS load balancer.

https://cloud.google.com/transcoder/docs/transcode-video https://cloud.google.com/cdn/docs/overview https://cloud.google.com/load-balancing/docs

Analytics & Real-Time Predictions

HRL offers a paid service to stream the races all over the world with live telemetry and predictions throughout each race. HRL wants to migrate their existing service to a new platform to expand their use of managed AI and ML services to facilitate race predictions. Race predictions are performed using TensorFlow running on VMs in the current public cloud provider. Support ability to expose the predictive models to partners. Increase predictive capabilities during and before races: Race results Mechanical failures Crowd sentiment Increase telemetry and create additional insights. Measure fan engagement with new predictions. Create real-time analytics of viewer consumption patterns and engagement.

This is where it gets very interesting. A couple of key requirements that HRL has is live telemetry and real-time predictions. GCP does provide APIs generating video telemetry i.e. video intelligence API, however real-time predictions, I will be discussing this in detail.

GCP provides this API called Live Streaming API which works in the streaming mode which means we can generate live telemetry such as video analysis, rich metadata, video annotations, object-based event triggers, highlight reels and recommendations along with video streaming.

Interesting part now, HRL also wants to get race predictions and wants to share predictive models with their partners so they can also generate predictions for their own criteria. After a lot of reading, Vertex AI is the right choice for this requirement.

Vertex AI offers services in 3 different ways. Predefined Models (Not applicable for this use case as its API like Vision) Auto ML (Applicable for this use case as users/partners can gain different predictions as per the requirements) Custom Models (Use of deep learning containers or images to generate predictions on user-defined model)

For HRL requirements, I have used both AutoML & Deep learning VM images. If we read the existing technical requirements, HRL already has a few prediction models running on VMs. As they are custom models, generating predictions specific to HRL predefined models and AutoML won't help here. It will be a lift-and-shift approach to generate predictions using custom models with VM images.

On the other hand, they are also required to open predictive models to partners, which is only possible with AutoML as these models are capable of unsupervised learning and generate predictions based on patterns. It helps to generate predictions with both streaming and batch modes, which means, old content that has been recorded over years can be used to generate predictions.

From a storage perspective, I think big query is a well-suited database that can persist telemetry and allows data to be fetched in both API & query format.

https://cloud.google.com/video-intelligence#section-5 https://cloud.google.com/vertex-ai/docs/training-overview https://cloud.google.com/vision-ai/docs/connect-bigquery https://cloud.google.com/automl https://cloud.google.com/vertex-ai/docs/general/deep-learning

Multi Cloud Container Solution

Their race prediction services are hosted exclusively on their existing public cloud provider. Minimize operational complexity.

This requirement could be easily solved by migrating the existing prediction service to GCP in container format. However, going by the executive statement, they don't have an immediate need to migrate the solution from the existing cloud provider service to GCP. Additionally, this gives us a chance to describe one of the amazing products Anthos, which abstracts out the complexity of managing containers by providing a common control plane.

Let's understand more in detail, why ANTHOS. I like this technical overview of ANTHOS which explains a lot.

The idea is to have all clusters in one place so management overhead is less like common configuration and policy management, a common set of networking rules, a common set of IAM and security policies and easier operations and resiliency features.

Anthos has this concept of the fleet, which allows multiple Kubernetes clusters to be part of the fleet. The fleet has sameness, which means within a fleet of clusters, some Kubernetes objects such as namespaces in different clusters are treated as if they were the same thing when they have the same name, providing a central user interface in the GCP console for managing fleet of clusters.

https://cloud.google.com/anthos/docs/concepts/overview https://cloud.google.com/anthos/multicluster-management/fleet-overview

Disclaimer

All these opinions are my sole responsibility based on learning GCP products and do not represent the only option for meeting these requirements. The solution can also be achieved with other products as well.