Scaling Push Messaging for Millions of Devices @Netflix

InfoQ Dev Summit Boston, a two-day conference of actionable advice from senior software developers hosted by InfoQ, will take place on June 24-25, 2024 Boston, Massachusetts.
Deep-dive into 20+ talks from senior software developers over 2 days with parallel breakout sessions. Clarify your immediate dev priorities and get practical advice to make development decisions easier and less risky.
Register now: bit.ly/47tNEWv
--------------------------------------------------------------------------------------------------------------------------------------
Download the slides & audio at InfoQ: bit.ly/2SKpUIC
Susheel Aroskar talks about Zuul Push, a scalable push notification service that handles millions of "always-on" persistent connections from all the Netflix apps running. He covers the design of the Zuul Push server and reviews the design details of the back-end message routing infrastructure that lets any Netflix microservice push notifications to any connected client.
This presentation was recorded at QCon New York 2018: bit.ly/2JFHitG
#SoftwareArchitecture #Netflix #ZuulPush #InfoQ #QConNewYork

[@prudvim3513]

Much better than watching "Design a Notification Service" System design interview videos.

[@zkwentz]

The “ask client to close” due to TCP wait avoidance, is brilliant. Well done!

[@lalanashree3962]

One of the best 'Notification Design videos' seen so far.

[@bedtimestoriesforkids9755]

I am watching this right before a system design session! :)

[@vibhorsharma6172]

Same lol

[@ThanosVassilakis]

This a great talk that really warms up after 20 minutesin when the talk goes into operation details.

[@audi88]

This talk is so good. Especially the mention of benchmarking numbers. m4.large ~ 8G,2vCPU = 84000 concurrent connections. ❤️

[@mat822]

yes it is very good talk but it not clear what is the throughput of the system with 84K clients, I guess it is very low if he was saying the CPU was barely used.

[@faizhalde2836]

@@mat822 the system does look io intensive (mostly network calls) so low cpu makes sense i think

[@helloworld7313]

@@mat822 125 million users, 1 push notification per day, each notification takes 1KB 125000000*1024/86400 < 2MB/s negligible throughput. Of course they kill connections periodically, but still very low throughput system.

[@Nemosaur]

This has been so insightful Thank you so much for this talk!

[@km2411_]

Really neat presentation with much informative content.

[@RanjithVj404]

Excellent presentation. Thanks for getting into implementation details!

[@pothiq]

It's truly comprehensive. Thanks for the nice presentation in a very understandable way.

[@huhu5375]

Amazing talks! I learn a lot from this video. "Ask client to close" is really brilliant. Randomized the connection time is also very impressive. I have seen random backoff algorithm, don't know it can also be used in this scenario.

[@sasaa4908]

Thank you so much for this talk !! very well explained ...

[@PankajKumar-vi4tn]

Very good talk. I also liked some of the questions in the end.

[@viveksai9353]

Every bit of applause deserved for such a verbose explaination.

[@8Trails50]

This is an insanely good talk.

[@at_tap]

Extremely well conducted video, very informative and explained without any wrinkles. Kudos!!

[@Dipesh_Sol]

Reminds me of backpressure, ofc not the same here.

[@AbhijeetJadhav24]

Super comprehensive and very inspirational! Thank you.

[@amitbaijal421]

Great presentation.

[@jmitesh01]

nit pick here, but how does sharding for availability help in push registry feature checklist? Its because it protects by only some portion of connection server goes down?

[@madeeshafernando8496]

wow, what a great talk, explains really well!!

[@stillyi]

Thank you! Very informative presentation.

[@unmeshchougule5666]

Question : when you say client opens connection and stays connected to that server, wouldn't it talk to load balancer and not the actual server here? So even if you deploy new server behind that load balancer there won't be any issue of the connection closing and re-opening as you mentioned? Please shed some light if I am missing something here.

[@pragmaticsoftwareengineer7965]

Brilliant Talk.. Learned so much.. Thanks Susheel.

[@codetolive27]

Excellent presentation!! Lots to learn

[@attention_shopping]

what a great talk. so dense, no fluff

[@wyjiaonju]

Great talk! Very inspirational!

[@lukeyd13]

What an amazing speaker this guy is

[@mukeshprajapati9642]

Really good explanation. Literally clapped at the end :)

[@deepakkumarnd]

Great talk, very detailed

[@BHARATKUMAR-le6eq]

I think some confusion at ELB. ELB operates at Layer 4 which means at the TCP layer already.

[@renon3359]

TCP TIMED WAIT avoidance was brilliant

[@allanvalooran6604]

Great Talk

[@praveenX]

Just Amazing! 👏

[@NK-ju6ns]

Neat presentation and insightful.

[@ankitjindal7059]

Really thought through presentation

[@dhrumilparmar9294]

Awesome talk! Enjoyed it.

[@VinayYadav-xy7dk]

Brilliant!!

[@hutofrock]

Excellent presentation with enough in-depth details and nice insights!

[@VishalThakur-wo1vx]

Love this talk

[@PratikNandoskar]

Very Good. thanks for sharing.

[@unmeshchougule5666]

great talk, thank you!

[@conquerorsanthosh]

Such an awesome content , thankyou

[@hyper.charged_]

Excellent

[@himanshuupadhyay6749]

Very nice presentation

[@gharatmayuresh15]

Excellent presentation !!!

[@BHARATKUMAR-le6eq]

I have only one request why the client depends on the server. I mean why this client-server mapping. What if the server is down then we need to migrate all the client-specific servers to a new server which is costly and may create issues like thundering. It's not we should make the server stateless with client requests and maybe only we should store client information in the push registry?? and when push notifications just randomly pick the server and send a message to the client id.

[@smonkey001]

How can they "push messaging" to Android with WebSocket while Google simply banned any push notification methods but FCM?

[@lahmerilyas4430]

This push mechanism is for users who are using the app. FCM is used for both forground/background usage.

[@TheNikhilbhoyar]

what is worse than thundering herd? its recurring thundering herd

[@ridhwaans]

what is the difference between push vs kafka messages/ message queues?

[@maximecaron3133]

with push client subscribe to a small set of keys( movie name …). while with kafka there is a single log and client need to read all message that have been written to the log

[@sonicjetson6253]

Hearing all this talk of massively scalable is getting so nauseous man, fk this shit

[@harshitpandey6664]

Tip : Watch at 1.25x speed.

[@greyreynyn]

man this is cool and all but all this work is just so you can say "hey, we think you might like this anime" ?