Low-latency live video streaming is increasingly important for Internet of Things (IoT) applications such as remote surveillance, industrial monitoring, smart-city infrastructure, and live sports analysis, where users must not only be able to watch in real time but also swiftly re-watch or revisit previous events. These requirements impose demands on low end-to-end latency, efficient client-side performance, and responsive digital video recording (DVR) interaction. 

Although existing research has thoroughly investigated protocol-level and network-efficiency-related latency, there is limited empirical understanding of how architectural design choices influence client-side DVR interaction behavior and resource usage. 

This thesis addresses this gap through an experimental study of low-latency streaming architectures for browser-based live video streams with DVR support. A Low-Latency HTTP Live Streaming (LL-HLS) baseline is compared with multiple Web Real-Time Communication (WebRTC)-based variants that use different transport mechanisms and media pipeline designs, including MediaStreamTrack, MediaRecorder, WebCodecs, and DataChannel-based approaches. 

The evaluation considers end-to-end latency, client-side CPU and memory utilization, and client-side DVR interaction behavior, including scrubbing responsiveness, frame-stepping precision, and tag-based replay recall. These metrics inform a trade-off analysis of architectural choices, emphasizing how the placement of responsibilities in the media pipeline affects performance. 

The results show that WebRTC-based delivery achieves significantly lower live latency than LL-HLS, consistent with previous findings in the literature. However, transport latency represents only one aspect of browser-based DVR systems. The evaluated architectures differed substantially in replay-oriented interactions. At 4K, the WebRTC DataChannel + WebCodecs prototype provides the strongest overall trade-off: it achieves the lowest baseline CPU usage of 44.3% compared to 64.2% for LL-HLS, while offering precise frame stepping and 100% accurate tag-jump recall.

The findings indicate that browser-based DVR implementation benefits the most when encoded media units remain visible and addressable at the application level, enabling more precise and responsive DVR interaction.