Wonder why nobody asked in my opinion obvious question: What if millions of users connect to a single POP to watch a popular streamer? How BigCache nodes handle it? Obviously people wathing the same stream in one POP are hitting the same BigCache node, does it have replicas? If yes, how many times each cached segment will be replicated? In my head this should be dynamic, I mean for some streams you need 2-3 replicas for balancing, but for some popular ones (for this particular PoP) you may need 10-15. If so how do you do this dynamic replica creation I'm curious. Thanks in advance.
I'm not FB engineer, but here is my take: Proxygen host can keep track of the concurrent requests to a specific streamID and the corresponding cache ID in a shared metadata cache. Certain policy can then be applied to tell the Proxygen host to route new requests to the same streamID to a new cache. Or the cache itself can keep track of the hot streamIDs and ask one of hosts to route requests somewhere else saying you have been requesting too frequently for the same content from me.
Excellent Presentation. I really appreciated his presentation style and with humor mix into this made it more engaging. This was full of nuggets and gave a very clear visual of the process in the forefront as well as the underlying infra. Great job!
Ironically, looking back now after 5 years, how well you've done then eventually Tiktok beats you so hard that end with how many users do you still have now?
"How do you deal with porn?" "We have both people and automated processes monitoring for porn, but yes, it's a _hard_ problem" What a memer. Great talk.
Nice talk, although I've one point, why would a company operating at Facebook scale not allow developers to develop an awesome custom protocol for Live streaming just because of tight deadlines?? I feel these guys should've got enough time to see if rmtps is actually the best protocol or they can create something really cool.
I don't work for FAANG, but I believe the main reason is "time to market" and premature optimization. Business needs to deliver a new feature/product before competitors do and start generating revenue stream from it. I guess it's easier to develop an MVP with whatever technology is available and a good fit for the task. After that you get rid of the bottlenecks in the future if the product is successful.
Really good talk, I watched the whole thing! This guy seems pretty cool too, I wonder if he still works at there at Meta? He doesn't seem evil enough to really fit in there these days? Lol😈
For the people watching it in 2023, RTMPS is no longer preferred since Adobe dropped flash, instead DASH (Dynamic Adapative Streaming over HTTP) or HLS ( HTTP live streaming) is preffered
Disappointing talk. We do not get to watch a talk from FaceBook teams every day, so I wish it was useful. I wish the presenter did not talk offside too much, too many jokes which makes it not even funny. The slides are too dark, the font is too small, not readable at all. Really boring. Instead of insisting on being a geek, I wish he was working on his presentation to be concise and more clear. Out of 51 minutes, only 5-10 minutes was useful.
Infrastructure deep dive starts at 9:10
Great presentation I've ever seen. Simple explanation of complex things with added humor make it awesone. Thank you.
One of the best presentations I've seen. Loved the simplification to explain things while still retaining the tech nitty-gritty :)
Very engaging presentation
Great video and presentation! Very engaging. Thanks for sharing!
Awesome talk. These are insanely hard challenges to overcome and it's really cool to learn from Facebook's example.
Wonder why nobody asked in my opinion obvious question: What if millions of users connect to a single POP to watch a popular streamer? How BigCache nodes handle it? Obviously people wathing the same stream in one POP are hitting the same BigCache node, does it have replicas? If yes, how many times each cached segment will be replicated? In my head this should be dynamic, I mean for some streams you need 2-3 replicas for balancing, but for some popular ones (for this particular PoP) you may need 10-15. If so how do you do this dynamic replica creation I'm curious. Thanks in advance.
I'm not FB engineer, but here is my take: Proxygen host can keep track of the concurrent requests to a specific streamID and the corresponding cache ID in a shared metadata cache. Certain policy can then be applied to tell the Proxygen host to route new requests to the same streamID to a new cache. Or the cache itself can keep track of the hot streamIDs and ask one of hosts to route requests somewhere else saying you have been requesting too frequently for the same content from me.
Why are InfoQ videos only 720p max resolution. The presentations are mostly pixelated.
This is very informative. Awesome Talk!
This presentation is underrated with only 17k views
Excellent Presentation. I really appreciated his presentation style and with humor mix into this made it more engaging. This was full of nuggets and gave a very clear visual of the process in the forefront as well as the underlying infra. Great job!
the protocol comparison in the talk is useless. looks like a go-to winner. No trade off
Ironically, looking back now after 5 years, how well you've done then eventually Tiktok beats you so hard that end with how many users do you still have now?
Great talk, covers enough technical detail but in an accessible way for someone who has no experience with video streaming.
"How do you deal with porn?"
"We have both people and automated processes monitoring for porn, but yes, it's a _hard_ problem"
What a memer. Great talk.
Can someone share what tool was used to make the presentation?
Elegant presentation!
Superb video and awesome talk. Great mix of technology and humor. Explained in plain English. Loved it. Thanks for sharing!!
One of the best presentation I've seen
This is GOLD!
such a good speaker. I really enjoyed the talk
This is one of the best videos I ever watched on youtube! Thank you so much for sharing such an informational video.
Nice talk, although I've one point, why would a company operating at Facebook scale not allow developers to develop an awesome custom protocol for Live streaming just because of tight deadlines?? I feel these guys should've got enough time to see if rmtps is actually the best protocol or they can create something really cool.
I don't work for FAANG, but I believe the main reason is "time to market" and premature optimization.
Business needs to deliver a new feature/product before competitors do and start generating revenue stream from it.
I guess it's easier to develop an MVP with whatever technology is available and a good fit for the task. After that you get rid of the bottlenecks in the future if the product is successful.
We did develop an custom protocol for live and it has been used in production two year ago, so no more rtmp now for fb/ig live.
Presentation link?
Dear future viewer : you can skip first 9 mins
08:14 BTW, Who is President Omaba?
President Lomagaba
great presentation. Its a difficult problem to solve
that's fucking insane you guys ship to a billion users in 8 months man
Really good talk, I watched the whole thing! This guy seems pretty cool too, I wonder if he still works at there at Meta? He doesn't seem evil enough to really fit in there these days? Lol😈
Billions of millions
millions * , not billions.
For the people watching it in 2023, RTMPS is no longer preferred since Adobe dropped flash, instead DASH (Dynamic Adapative Streaming over HTTP) or HLS ( HTTP live streaming) is preffered
Disappointing talk. We do not get to watch a talk from FaceBook teams every day, so I wish it was useful. I wish the presenter did not talk offside too much, too many jokes which makes it not even funny. The slides are too dark, the font is too small, not readable at all. Really boring. Instead of insisting on being a geek, I wish he was working on his presentation to be concise and more clear. Out of 51 minutes, only 5-10 minutes was useful.
Agree
Unfortunately true, very little useful info.