Designing bitrate ladders for adaptive streaming

A bitrate ladder is a set of encodes at graduated quality. The ladder is only as good as its rung spacing, and the spacing only works if it matches how players actually switch.

Adaptive bitrate delivery works by measuring throughput and buffer health on the client, then requesting a representation whose bitrate the connection can sustain. The player does not know the content; it only knows the manifest that lists rungs and their declared bitrates. Design the ladder badly and the player thrashes between rungs, or spends long stretches at a representation far below the display's capability.

Classic fixed ladders

The oldest approach defines a fixed set of rungs from a low mobile target upward to the maximum the service will ever serve. Rung spacing is typically around 1.5 to 2 times between adjacent bitrates, because smaller steps waste storage while larger steps force the player to choose between visible quality loss and rebuffering. Each rung is encoded at the same resolution grid, commonly derived from a standard definition of ladders published by the major streaming platforms.

Why fixed ladders waste bits

The fixed ladder applies the same bitrate envelope to every title. A slow interview and a detailed landscape therefore receive identical treatment at a given rung, even though their complexity differs by an order of magnitude. Per-title encoding addresses this by analysing each title, measuring the rate-quality curve across several operating points, and choosing the rungs that lie on the efficient frontier of that curve. The result is typically a mix of lower bitrates at equal measured quality, and additional rungs where the content is unusually demanding.

Per-shot encoding goes further by varying operating points within a single title, which can recover additional savings on productions with sharply varying visual complexity. The cost is a more complex manifest and a more demanding analysis stage, so it suits libraries large enough to amortise that engineering.

Codec mix

A modern ladder is often codec-mixed: a widely supported codec for the baseline set and a more efficient codec for clients that advertise support. The manifest can list both, and the player selects the intersection of what it can decode and what the network can sustain. Because a more efficient codec reaches a given quality at a lower bitrate, the codec-mixed ladder effectively shifts every rung leftward for capable clients without editing the baseline ladder.

The player's decision model

Throughput estimation is noisy, so players smooth it and add a safety margin. Buffer occupancy then modulates the choice: a deep buffer permits an aggressive upswitch that improves quality quickly, while a shallow buffer forces a conservative downswitch that avoids a stall. Ladders that place a large bitrate gap between the top two rungs cause oscillation, because the safety margin required to hold the top rung is much larger than the margin for the rung below. Moderate spacing reduces the cost of uncertainty.

Latency changes the calculus as well. Low-latency delivery shortens the available buffer, which makes conservative ladders and smaller segments preferable, at some cost in compression efficiency because shorter segments reset encoder state more often. Every ladder is therefore an optimisation over three variables at once: quality, robustness and cost.

Measuring a ladder

Compare ladders on aggregate quality delivered over a reference network trace, not on instantaneous bitrate. A ladder that delivers a higher average quality at the same total bytes is the better ladder, regardless of how its individual rungs look on paper. Re-run the comparison whenever the codec mix or the segment length changes, because both shift the whole rate-quality surface.