Low latency live streaming using CMAF-based chunked encoding and transfer has, in one form or another, been one – sometimes ‘the’ – subject du jour of the community in recent times. The competing proposals have various strengths and drawbacks that have been well dissected in previous talks and the emerging consensus thus far is that some variant of or co-existence of these approaches is where we will inevitably settle (in spite of the fact that these approaches introduce new complexity, require different adaptive bit rate control in players for good bandwidth adaptation, and in some cases, just can’t be supported without significant changes to origin servers or HTTP implementations).

In this talk, I will describe a different approach to the problem, one that starts from a different assumptions: That is, the root of the problem in realizing low-latency ABR streaming (to standard deployed players over HTTP) is not the streaming protocol per se but the constraints of traditional content distribution and transcoding pipelines. I will dissect the problems of the assumed “standard” architecture that works against the fundamental goals of achieving low latency live output to a large number of concurrent readers (streaming clients) in terms of a) the internal representation of media in the pipeline, b) file I/O constraints, c) application routing and micro service latencies, and d) piecemeal architecture. I will contrast this with an alternative software substrate we’ve built that takes in standard live mezzanine streams (format agnostic) and yields ABR streaming to standard players (DASH, HLS) at scale using CMAF-based encoding, a decentralized real time publishing and routing layer, and just-in-time transcoding with minimal file I/O. I will include learnings regarding its performance for streaming to Internet audiences with high concurrency, in terms of stream bit rate, latency from the live source, infrastructure load and average player bit rate. Presented at Demuxed 2020.