By Renata Teixeira, Zhi Li, Reenal Mahajan, and Wei Wei
On January 26, 2026, we flipped an vital swap for Stay at Netflix: all Stay occasions are actually encoded utilizing VBR (Variable Bitrate) as a substitute of CBR (Fixed Bitrate). It feels like a small configuration change, however it required us to revisit among the foundational assumptions behind how we ship Stay video at world scale.
VBR lets us tailor the bitrate to the precise complexity of the scene, as a substitute of sending each second of video at roughly the identical bitrate. When a scene is easy, VBR “shaves off” bits that wouldn’t enhance what you see on display screen; when a scene is complicated, it spends extra bits to protect high quality. (The extra basic concept is also known as capped variable bitrate, or capped VBR.) That makes our encodes extra environment friendly and our community extra scalable. But it surely additionally makes site visitors a lot much less predictable: massive bitrate swings can overload servers and CDNs, and our outdated assumptions about “what bitrate equals what high quality” not maintain. In consequence, we have now to rethink each how we handle supply and capability, and which bitrates we provide for every model of the stream. In our stay pipeline, we presently use AWS Elemental MediaLive, the place this “capped” VBR is carried out utilizing the QVBR (High quality‑Outlined Variable Bitrate) setting.
| Why Transfer Stay from CBR to VBR?
Our preliminary Stay encoding pipeline used fixed bitrate (CBR). For every encoded stream, we configured a decision and a nominal bitrate — for instance, a 1080p stream focusing on 5 Mbps — and the precise bitrate stayed near that concentrate on over time. This predictability made each capability planning and day‑to‑day operations simpler. If a server may safely ship round 100 Gbps of Stay site visitors, and every stream averaged near its nominal fee, we may admit on the order of twenty thousand concurrent periods per server and be assured we have been working inside limits. Throughout an occasion, the overall site visitors despatched by a server would change primarily when members joined or left; so long as concurrency was secure, site visitors stayed comparatively flat. The community noticed a clean, straightforward‑to‑cause‑about load profile, and enormous adjustments in throughput virtually all the time mirrored an actual change in utilization, not only a completely different scene on display screen.
The issue is that content material isn’t fixed. A speaking‑head phase in a studio or a easy animation is far simpler to compress than a sequence of fast digicam strikes and quick‑transferring athletes in entrance of a extremely detailed crowd. With CBR, each the straightforward and the onerous segments get the identical bitrate. In easy scenes, we spend extra bits than we’d like; in complicated scenes, we typically don’t spend sufficient.
VBR flips the target. Relatively than aiming for a set bitrate, the encoder goals for a goal high quality and is allowed to boost or decrease the bitrate in line with scene complexity. When the image is simple to encode, VBR can drop the bitrate considerably under the outdated CBR degree whereas retaining high quality fixed. When the motion heats up, it could possibly briefly use extra bits to keep away from seen artifacts.
The determine under exhibits the per‑phase bitrate over time for a similar episode of WWE RAW, encoded as soon as with CBR and as soon as with VBR at a nominal 8 Mbps. With CBR (blue), the bitrate wobbles a bit from phase to phase however stays near the goal; for those who common it over a minute, it’s virtually a straight line. We find yourself spending roughly the identical variety of bits on easy scenes, just like the ready room at the beginning of the stream (shaded area), as on complicated scenes, just like the confetti‑stuffed shot later within the present. With VBR (orange), the encoder can drop the bitrate for the ready room to a small fraction of the nominal fee, whereas permitting the confetti‑stuffed shot to make use of a a lot larger bitrate to protect high quality.
At Netflix scale, the shift to VBR has a number of vital results. Crucial is effectivity on the community: we cut back the common variety of bytes we have to ship a full occasion, which in flip reduces the site visitors wanted to fill all the servers in Open Join, our content material supply community (CDN), and the site visitors wanted to serve segments to members. The second is high quality of expertise. As a result of VBR sends fewer bits for comparable video high quality, we see fewer rebuffers and decrease begin‑up delay. Throughout a number of A/B exams on completely different Stay occasions, we noticed about 5% fewer rebuffers per hour, whereas transferring roughly 15% fewer bytes on common and round a ten% discount in site visitors on the peak minute.
| When Effectivity Fights Stability
The central problem with VBR for Stay isn’t that bitrate varies in any respect (it does underneath CBR as properly), however that VBR can have a lot deeper, longer dips in bitrate, tightly coupled to what’s occurring within the content material.
Underneath CBR, a 5 Mbps stream is successfully that: on a per‑minute foundation, site visitors is remarkably flat. A server that’s comfortably dealing with, say, ten thousand such periods now could be prone to nonetheless be comfy in a minute, barring a wave of recent joins. It’s protected for our steering logic to take a look at present site visitors, see loads of headroom, and route extra periods to that server.
Get Netflix Expertise Weblog’s tales in your inbox
Be a part of Medium free of charge to get updates from this author.
Underneath VBR, the identical stream behaves very in a different way. Throughout a sluggish, straightforward sequence, the encoder may solely generate 2 Mbps for the 5 Mbps stream — and even much less — to keep up its goal high quality, and it could possibly keep at that decrease degree for an prolonged interval. The server then seems to have loads of unused capability: per‑session bitrate is low and mixture site visitors is properly under its limits. Our steering programs naturally interpret this as a sign that the server is underneath‑utilized and might settle for extra periods.
The issue surfaces when the content material adjustments. A struggle begins, confetti begins to fall, or the digicam cuts to a extremely detailed, quick‑transferring shot. To protect high quality, VBR might enhance the bitrate to six, 7, or 8 Mbps on the very subsequent segments. If the server has admitted many extra periods in the course of the previous low‑bitrate interval, the combination site visitors can all of a sudden exceed what the community hyperlink or NIC can maintain. Latency rises, packets are dropped, and gadgets begin to expertise stalls or high quality downshifts. In excessive circumstances, this sample of “bitrate dips adopted by spikes” can destabilize components of the system.
| Making Servers Conscious of Bitrate Variability
These lengthy VBR bitrate dips are nice for effectivity, however — as we simply noticed — they’ll trick our supply programs into pondering servers are underneath‑utilized and protected to load up. Underneath CBR, that habits was predictable sufficient that present site visitors was a very good proxy for a way “full” a server was; underneath VBR, it isn’t.
Our repair was to vary how we resolve whether or not a server can take extra periods. As a substitute of basing that call solely on present site visitors, we reserve capability based mostly on every stream’s nominal bitrate, not simply what it occurs to be utilizing at that second. Even when a VBR stream is presently in a really low-cost, low‑bitrate section, we nonetheless deal with it as one thing that may shortly return to its nominal fee.
This retains our site visitors‑steering habits constant between CBR and VBR and avoids the important thing failure mode the place a server accepts too many periods throughout an extended low‑bitrate interval after which turns into overloaded when bitrate rises once more.
| Tuning VBR Nominal Bitrates to Match CBR High quality
The WWE instance above already hints at why this isn’t computerized. We checked out a single 8 Mbps stream, encoded as soon as with CBR and as soon as with VBR. Each encodes have the identical nominal bitrate, however the determine exhibits how in a different way they behave. The CBR encode stays clustered round 8 Mbps with frequent quick spikes, whereas the VBR encode usually drops far under that degree and solely spikes up when the content material will get complicated. That’s extra environment friendly, however it additionally signifies that “similar nominal bitrate” doesn’t indicate “similar common variety of bits” anymore — so merely reusing our CBR settings dangers giving VBR much less bitrate on common and dropping some high quality.
In follow, after all, we don’t simply encode a single stream; we produce a set of streams at completely different resolutions and nominal bitrates — usually referred to as a bitrate ladder — so gadgets can adapt to their present community situations by switching between them. Once we first utilized VBR utilizing the present CBR ladder, offline evaluation with VMAF (a perceptual video high quality metric) confirmed the priority from the WWE instance: time‑averaged high quality dropped barely on a number of streams, particularly on the lowest bitrates. Early A/B exams confirmed the identical sample: general VMAF about one level decrease than CBR, with a lot of the hole on the backside of the ladder.
To repair this, we in contrast CBR and VBR encodes rung by rung and checked out per‑stream VMAF. Wherever VBR fell greater than about one VMAF level under CBR, we elevated its nominal bitrate simply sufficient to shut the hole. Increased‑bitrate streams, the place VBR high quality was already very near CBR, have been left largely unchanged, together with the 8 Mbps stream from the determine.
The result’s a VBR ladder with barely larger nominal bitrates on a number of low‑finish streams, however decrease general site visitors, as a result of VBR nonetheless drops the bitrate on easy scenes. This lets us match the standard of our CBR ladder whereas retaining the effectivity and stability features that motivated the swap to VBR within the first place.
| What’s Subsequent for Stay VBR
With VBR in manufacturing for all Stay occasions, our focus now could be on utilizing it extra intelligently.
First, we’re testing tips on how to use the precise sizes of upcoming segments in our adaptive bitrate algorithms on gadgets, as a substitute of relying solely on nominal bitrates. This could assist gadgets choose streams that higher match how VBR will behave within the subsequent few seconds, not simply how a stream is labeled on paper.
Second, we’re experimenting with making our capability reservation much less conservative. At this time we reserve based mostly on nominal bitrates to maintain servers protected; by rigorously making use of a “low cost” knowledgeable by actual VBR habits, we hope to release extra headroom with out sacrificing stability.
This work was the results of a broad, cross‑crew effort. We’d prefer to thank Mariana Afonso, Dave Andrews, Mark Brady, Jake Freeland, Te-Yuan Huang, Ivan Ivanov, Yeshwenth Jayaraman, Patrick Kunka, Zheng Lu, Anirudh Mendiratta, Chris Pham, David Pfitzner, Jon Rivas, Garett Singer, Brenda So, Stan Surmay, Bowen Tan, Devashish Thakur, and Allan Zhou for his or her many contributions to creating Stay VBR a actuality at Netflix.
