Independent delivery lab · est. 2026
We take streaming delivery apart and find out where. Cache topology, segment sizing, connection reuse — the unglamorous layer between an encoder and a screen.
Notes here are written while working, not after. They are specific, occasionally wrong, and corrected in place.
Three areas, because they are the three places latency hides.
Deciding what lives at the edge and what does not. Most libraries are long-tail — caching everything costs more than it saves, and the tail is where hit ratios go to die.
Segment length is a trade between startup time and overhead. Two seconds is the reflex answer. It is right less often than people assume, particularly on lossy mobile networks.
Connection reuse, congestion control, how a player recovers when a link degrades mid-stream. Rebuffering is rarely a bandwidth problem. It is usually a recovery problem.
Measured across our own test corpus. Not benchmarks — working baselines.
Short write-ups from work in progress.
A single time-to-live across a catalogue treats a title watched twice a month like one watched every minute. Splitting TTLs by observed request frequency moved our edge hit ratio nearly two points without adding storage.
We expected one-second segments to cut time-to-first-frame. They did, marginally — and then gave it all back in request overhead on high-latency links. The win was in the manifest, not the segments.
Two connections with identical average bandwidth produced very different viewing experiences. The difference was entirely in how quickly each recovered from a brief loss event.
Questions about a delivery setup, or a correction to something written here — both welcome.
We read everything. Replies take a few days when measurement runs are going.