How Edge Containers Transform Global Audio‑Video Services: Real‑World Lessons

At the 2023 WOT Global Technology Innovation Conference in Shenzhen, Xiao Shao, Director of Backend Technology at iConstruct, presented "Exploring and Practicing Edge Containers in Global Audio‑Video Scenarios". He explained why real‑time interactive services naturally fit edge computing, highlighting benefits such as reduced latency, lower costs, higher concurrency, and improved fault tolerance.

Motivation and Goals

Initially iConstruct rented large‑scale VMs and physical machines worldwide, leading to low edge resource utilization, high costs, and fragmented operations. The team aimed for three improvements with cloud‑native edge containers: maximize compute and bandwidth reuse, boost operational efficiency and reliability through elastic scaling and version management, and build a unified cloud‑edge native infrastructure.

Challenges of Deploying Edge Containers

Key obstacles include the lack of a standard edge‑container definition, handling stateful audio‑video services, diverse service specifications, multi‑process pod updates, network interruptions, traffic cost reduction, and operational efficiency.

Global Audio‑Video Cloud Architecture

The solution combines multi‑cloud infrastructure, edge containers, a global multi‑center layout, and the Massive Serial Data Network (MSDN) for massive ordered data transmission. A central hub remains essential because a full‑mesh of 500+ data centers would be inefficient.

Technical Choices

Host‑network mode is used to eliminate container network virtualization overhead, with a DaemonSet managing port allocation.

OpenKruise CloneSet replaces the default Kubernetes workload to support in‑place upgrades, targeted scaling, sidecar deployment, and image pre‑warming.

Sidecar containers are injected via webhook, enabling independent gray‑release of the main audio‑video engine and its auxiliary processing container.

For zero‑downtime upgrades beyond image changes, a custom CloneSetMigration operator coordinates migration without pod recreation, preserving IP and port stability.

Image pre‑warming is preferred over P2P distribution (e.g., Dragonfly) because it avoids complex network requirements and bandwidth throttling, meeting the need for rapid image pulls during in‑place upgrades.

Elastic Scaling for Stateful Audio‑Video Workloads

Scaling decisions consider bandwidth, PPS, stream count, CPU, and memory. Scaling‑down is delayed until active streams finish, using a state machine with states such as "initialized", "allocated", "tombstone", and "awaiting deletion" to ensure near‑zero impact on users.

Quality and Operational Efficiency Enhancements

The proprietary MSDN provides real‑time network quality monitoring, enabling intelligent global routing, sub‑second fault recovery, and higher transmission reliability. Machine‑learning models predict edge‑site utilization to automate capacity planning, node procurement, and decommissioning. Multi‑cluster management unifies global resource control.

Cost Optimization Strategies

Edge resource pooling allows different business clusters to share the same pool, reducing redundancy.

A hierarchical resource‑pool scheduler (Spread → BinPack) prefers packing pods onto heavily used nodes, improving overall utilization.

OpenYurt optimizations reduce Service and EndpointSlice usage, filter kubelet watch lists, and employ a Pool‑Coordinator/Yurthub design to keep a single copy of pool‑scope data per node, cutting cloud‑edge traffic.

Multi‑cloud serverless clusters provide burst capacity for edge spikes, combining edge bandwidth with serverless cost‑effective scaling.

Future Outlook

iConstruct plans to advance in three areas: develop more suitable scheduling algorithms (including CPU topology‑aware and GPU scheduling), build richer toolchains to lower edge‑container onboarding time to minutes, and expose additional capabilities directly on the edge side.

The company will continue to push edge containers forward in global audio‑video scenarios.