How Netflix keeps its data infrastructure cost-effective
acceldataio.medium.com
acceldataio.medium.com
Or maybe I'm way underestimating the size of the Netflix library, but my impressions is it extremely tiny by comparison to YouTube.
In the end this read like a marketing piece for AccelData.
If there is one thing that Netflix technically excel, it is their 800Gbps Serving Appliance [1] with FreeBSD. ( I cant wait until they do 1.6Tbps )
I wish they did, their current encoding for some shows is absolutely atrocious. I don't know the terminology for what I'm seeing, but for example in Stranger Things, it seems like the brightness and contrast is boosted for every n-th frame. In general, smooth motion seems to get a stuttering quality whenever Netflix touches it. I don't observe this so much with other video sources on the same device...
Tangent if they really cared about user experience and Ecology they would implement h265 and h266 support in chromium (aka a mere function call to the embedded ffmpeg). H266 is revolutionary but it's adoption is of zero because of AV1 lobbyists a bit like so called environmentalists once launched a RPG rocket on the superphenix thorium mankind saving reactor.
Adoption on the Web. Most broadcasting companies are getting ready to move to H.266.
I wish we could get a JPEG XL based Video codec.
> I wish we could get a JPEG XL based Video codec. Well i'm curious about that but FLIF, the predecessor (and maybe XL too) is based on MANIAC which is a variation of H264 encoding.. https://en.wikipedia.org/wiki/Free_Lossless_Image_Format#:~:...
Oh they are definitely on the roadmap ( Qualcomm and Mediatek ). I think a lot of people forget there are long lead time in hardware development. Especially with Mobile where energy usage is a key issue. The conformance test of H.266 only finished in mid 2021 if I remember correctly and continued being refined in 2022. If you are going to test your silicon in 2021, your earliest chance of any mobile solution with VVC is going to be 2023.
I would guess both AV1 and VVC lands in Qualcomm SoC in ~2023. MediaTek and Apple coming in 2024.
This doesn’t sound accurate to me. What about their 400Gb/800Gb FreeBSD boxes serving up content? Surely those are in installed in Netflix managed collocation facilities? Unless the author is suggesting that any on-prem hardware is a ‘private’ cloud. :barf:
https://openconnect.netflix.com/en/#sample-architectures
If one of these appliances installed in an ISP's cabinet dies, content to end users will get served directly by an appliance within Netflix's infrastructure.
But I imagine it's been a challenge, since video sizes have grown considerably as well. Hence why new Youtube algos heavily favor videos that can be easily monetized and thus profitable for them.
Note: this mostly applies to dark scenes. bright scenes are not nearly as bad, but still don't look as good as youtube.
The VMAF-approved banding seems particularly telling!
But stuff like this intro just look incredibly good to me : https://www.youtube.com/watch?v=Yeh4rsXhsUg
Their source says Netflix has 50K titles which means you can probably store entirety of it on the edge.
Most numbers I've seen suggest prime is significantly larger. (Though in my opinion its quantity over quality unfortunately).