Streaming Workflows on AWS

Executive Summary

• Two AWS Media reference architectures describe workflows for (1) personalized server-side ad insertion (SSAI) at high concurrency and (2) AI-assisted live highlight extraction and vertical-video creation for social platforms.
• The SSAI design centers on AWS Elemental MediaTailor integrated with Google Ad Manager (GAM) using VAST, producing per-session HLS manifests and supporting asynchronous ad tracking via MediaTailor PING.
• The vertical-video prototype uses AWS Elemental Inference for key-moment detection, event-driven orchestration with Amazon EventBridge, and clip packaging and conversion workflows spanning HLS segmentation, Amazon S3 storage, and MP4 generation.

Key Industry Developments

• Personalized SSAI at high concurrency using HLS manifests
• Sony LIV delivered personalized advertisements at scale to a peak of over 10 million concurrent viewers during Asia Cup Cricket 2025.
• Multiple live inputs were ingested and encoded using AWS Elemental MediaLive into adaptive bitrate (ABR) HLS renditions, establishing the base stream format for downstream ad insertion and delivery.
• AWS Elemental MediaTailor acted as the SSAI layer by dynamically generating per-session manifests with ad insertions, enabling personalization at the manifest level rather than requiring client-side stitching.
• Ad decisioning and tracking integrations designed for scale
• The SSAI architecture leveraged AWS Elemental MediaTailor integrated with Google Ad Manager (GAM), using VAST responses from GAM for ad decisions.
• MediaTailor’s pre-fetch capability was used to reduce request timeouts and throttling issues at the Ad Decision Server (ADS) during high viewership periods, indicating an explicit mitigation for ADS pressure under load.
• MediaTailor’s PING feature enabled asynchronous ad tracking for impression reporting without affecting playback or performance, separating tracking calls from the critical playback path.
• AI-driven, event-based highlight creation and vertical formatting
• AWS Elemental Inference is described as a fully managed AI service that automatically detects key moments and transforms broadcasts into vertical video content, targeting 9:16 vertical social platforms.
• The prototype automatically detects key moments from live sports broadcasts, extracts clips, reformats them for 9:16 vertical platforms, and surfaces them in a review portal within seconds, emphasizing low-latency editorial availability.
• Each detection emits an event to Amazon EventBridge, enabling an event-driven workflow where detections trigger downstream clip harvesting and storage actions.

Real-World Use Cases

• Live sports SSAI with multi-CDN distribution
• A live sports streaming workflow used AWS Elemental MediaLive to ingest and encode multiple live inputs into ABR HLS renditions, then relied on AWS Elemental MediaTailor to generate per-session HLS manifests with personalized ad insertions.
• The SSAI workflow integrated with Google Ad Manager (GAM) for ad decisioning via VAST responses, aligning ad selection with a standard ad response format used by ad servers.
• Final HLS manifests and segments were distributed across multiple CDNs in a multi-CDN environment including Amazon CloudFront to handle regional traffic spikes and provide low-latency delivery.
• Operational support for live streaming workflows
• AWS Unified Operations for Media provided architectural guidance, workflow reviews, partner coordination, and real-time support during live streaming, reflecting an operations-focused layer around the live event workflow.
• Live highlight extraction, clip management, and MP4 outputs
• A live workflow used AWS Elemental MediaLive for ingest and encoding and AWS Elemental MediaPackage V2 to segment content and maintain a rolling window for clip extraction, supporting near-real-time harvesting from the live stream.
• AWS Elemental Inference emitted events to Amazon EventBridge containing PTS values, descriptive tags, and an AI-generated description, which triggered clip harvesting and storage in Amazon S3.
• Video editing and MP4 download workflows used asynchronous processing with Amazon SQS queues and AWS Lambda functions, storing outputs in S3 and updating records in Amazon DynamoDB; HLS content was converted to MP4 format using AWS Elemental MediaConvert.

Why It Matters

• SSAI performance depends on controlling load across ad decisioning and tracking paths
• Per-session manifest generation enables individualized ad insertion while keeping playback in HLS, but it increases the importance of resilient ad decisioning and tracking workflows under concurrency.
• Pre-fetching ad decisions can reduce ADS timeouts and throttling during high viewership periods, and asynchronous tracking via MediaTailor PING can preserve playback performance while still supporting impression reporting.
• Event-driven AI highlight workflows formalize low-latency editorial pipelines
• Emitting structured detection events (including PTS values and tags) to Amazon EventBridge supports deterministic clip harvesting from a rolling HLS window and enables downstream automation without requiring manual timecode discovery.
• Asynchronous processing patterns (SQS + Lambda) and durable storage (S3) support parallelizable clip conversion and distribution workflows, including MP4 generation for download and sharing.

Sources

• https://aws.amazon.com/blogs/media/how-sony-pictures-networks-india-delivered-personalized-ad-insertion-at-scale-on-aws/
• https://aws.amazon.com/blogs/media/how-aws-built-a-live-ai-powered-vertical-video-capability-for-fox-sports-with-aws-elemental-inference/