3 ms·

If you're at the point where you decided the 1000 nodes are required, you should probably have already considered the sensible alternatives first, and concluded

If you're at the point where you decided the 1000 nodes are required, you should probably have already considered the sensible alternatives first, and concluded this was somehow better. The network saturation is just a necessary cost of running such a massive cluster. I really have no idea what kind of system would require 1000 nodes, that couldn't be replaced by 100, 10x larger, nodes instead. And at that point, you should probably be thinking of ways to scale the network itself as well.

sausagefeet 2y ago
I don't really understand what this comment is responding to. The comment I responded to hand waved away consuming almost 50% of your pps on heartbeats every 30 seconds as "no big deal".
- ricketycricket 2y ago
  > The comment I responded to hand waved away consuming almost 50% of your pps on heartbeats every 30 seconds as "no big deal". > The network saturation is just a necessary cost of running such a massive cluster. I think this actually answers it perfectly. 1. If you are running 1K distributed nodes, you have to understand that means you have some overhead for running such a large cluster. No one is hand waving this away, it's just being acknowledged that this level of complexity has a cost. 2. If heartbeats are almost 50% of your pps, you are trying to use 1Gbe to run a 1K-node cluster. No one would do this in production and no one is claiming you should. 3. If your system can tolerate it, change the heartbeat interval to whatever you want. 4. Don't use distributed Erlang if you don't have to. Erlang/Elixir/Gleam work perfectly fine for non-distributed workloads as do most languages that can't distribute in the first place. But if you do need a distributed system, you are unlikely to find a better way to do it than the BEAM. Basically, it seems you are taking issue with something that 1) is that way because that's how things work, and 2) is not how anyone would actually use it.
  - gatnoodle 2y ago
    Also, 5. If you plan on deploying such massive clusters, hire a good network engineer. (:
throwawaymaths 2y ago
IIRC a preso by them, WA runs an eye popping number of nodes in distribution and they are indeed fully connected, and it's fine (though they do use a modified BEAM)
- anonymousDan 2y ago
  Interesting, do you have a link by any chance?
  - toast0 2y ago
    From 2014 there's a video link and slides here: https://www.erlang-factory.com/sfbay2014/rick-reed https://www.erlang-factory.com/sfbay2014/rick-reed FYI: I've seen comments from after I left that wandist isn't used anymore; I think a lot of what we gained with that was working around issues in pg2 that stem from global locks not scaling... but the new pg doesn't need global locks at all. There was also some things in wandist to work around difficulties communicating between SoftLayer nodes and Facebook nodes, but that was a transitory need. See the 2024 presentation, 40k nodes! Fairly similar, but smaller numbers in 2012 http://www.erlang-factory.com/conference/SFBay2012/speakers/RickReed http://www.erlang-factory.com/conference/SFBay2012/speakers/... The 2013 presentation is focused on MMS which I don't remember if it was as impressive: http://www.erlang-factory.com/conference/SFBay2013/speakers/RickReed http://www.erlang-factory.com/conference/SFBay2013/speakers/... (note that server side transcoding is from before end to end encryption) I don't think there were similar presentations on Erlang in the large at WhatsApp after that. Big changes between 2014 and 2019 (when I left) were a) chat servers started doing a lot more, and clients per server went down on the big SoftLayer boxes b) hosting moved from SoftLayer to Facebook and much smaller nodes --- also chat servers at SoftLayer were individually addressable, using (augmented) round robin DNS to select, at Facebook the chat servers did not have public addresses, instead everything comes in through load balancers c) MMS was pretty much offloaded into a Facebook storage service (c++); not because the Erlang wasn't sufficient, but because MMS was loosely coupled with the rest of the service, Facebook had a nice enough storage service, a lot of storage, an awful lot of bandwidth, and it wasn't a lot of work for that team to also support WhatsApp's needs; also our Erlang MMS (and the PHP version before it) was built around storing files on specific, addressable nodes, but nodes at Facebook are much more ephemeral and not easy to directly address by clients d) some amount of data storage moved off of mnesia into other Facebook data storage technology; again, not because mnesia wasn't sufficient, but more ephemeral nodes makes it cumbersome (addressable) and the available hardware nodes at FB didn't really match --- there's a very firm bias at FB towards using standard node sizes and the available standard nodes were like a web machine with not much ram or a big database machine with more ram and tons of fast storage; WA mnesia wants lots of ram but doesn't need a lot of storage (all data is in ram, and dumped + logged to disk) so there was a mismatch there --- things that stayed in mnesia needed much larger clusters to manage data size Presentations became less common because of more layers to get approval, and also because it's less fun to share how we built something on top of proprietary layers that others don't really have access to. Anybody could have gotten dual 2690 servers at SoftLayer and run a nice Erlang cluster. Only a few people could run an even bigger chat cluster in a Facebook like hosting environment.
toast0 2y ago
> I really have no idea what kind of system would require 1000 nodes, that couldn't be replaced by 100, 10x larger, nodes instead. Nodes only get so big. Way back when, a quad xeon 4650 v2 was definitely not 2x the throughput of a dual xeon 2690 v2; so you end up with two dual socket systems instead of one quad socket. A quad socket server often costs significantly more than two dual socket servers, and is likely to take longer between order and delivery. There's usually a point where you can still scale up, but scaling out is a better use of resources.

sausagefeet 2y ago

I don't really understand what this comment is responding to. The comment I responded to hand waved away consuming almost 50% of your pps on heartbeats every 30 seconds as "no big deal".

ricketycricket 2y ago
> The comment I responded to hand waved away consuming almost 50% of your pps on heartbeats every 30 seconds as "no big deal". > The network saturation is just a necessary cost of running such a massive cluster. I think this actually answers it perfectly. 1. If you are running 1K distributed nodes, you have to understand that means you have some overhead for running such a large cluster. No one is hand waving this away, it's just being acknowledged that this level of complexity has a cost. 2. If heartbeats are almost 50% of your pps, you are trying to use 1Gbe to run a 1K-node cluster. No one would do this in production and no one is claiming you should. 3. If your system can tolerate it, change the heartbeat interval to whatever you want. 4. Don't use distributed Erlang if you don't have to. Erlang/Elixir/Gleam work perfectly fine for non-distributed workloads as do most languages that can't distribute in the first place. But if you do need a distributed system, you are unlikely to find a better way to do it than the BEAM. Basically, it seems you are taking issue with something that 1) is that way because that's how things work, and 2) is not how anyone would actually use it.
- gatnoodle 2y ago
  Also, 5. If you plan on deploying such massive clusters, hire a good network engineer. (:

throwawaymaths 2y ago

IIRC a preso by them, WA runs an eye popping number of nodes in distribution and they are indeed fully connected, and it's fine (though they do use a modified BEAM)

anonymousDan 2y ago
Interesting, do you have a link by any chance?
- toast0 2y ago
  From 2014 there's a video link and slides here: https://www.erlang-factory.com/sfbay2014/rick-reed https://www.erlang-factory.com/sfbay2014/rick-reed FYI: I've seen comments from after I left that wandist isn't used anymore; I think a lot of what we gained with that was working around issues in pg2 that stem from global locks not scaling... but the new pg doesn't need global locks at all. There was also some things in wandist to work around difficulties communicating between SoftLayer nodes and Facebook nodes, but that was a transitory need. See the 2024 presentation, 40k nodes! Fairly similar, but smaller numbers in 2012 http://www.erlang-factory.com/conference/SFBay2012/speakers/RickReed http://www.erlang-factory.com/conference/SFBay2012/speakers/... The 2013 presentation is focused on MMS which I don't remember if it was as impressive: http://www.erlang-factory.com/conference/SFBay2013/speakers/RickReed http://www.erlang-factory.com/conference/SFBay2013/speakers/... (note that server side transcoding is from before end to end encryption) I don't think there were similar presentations on Erlang in the large at WhatsApp after that. Big changes between 2014 and 2019 (when I left) were a) chat servers started doing a lot more, and clients per server went down on the big SoftLayer boxes b) hosting moved from SoftLayer to Facebook and much smaller nodes --- also chat servers at SoftLayer were individually addressable, using (augmented) round robin DNS to select, at Facebook the chat servers did not have public addresses, instead everything comes in through load balancers c) MMS was pretty much offloaded into a Facebook storage service (c++); not because the Erlang wasn't sufficient, but because MMS was loosely coupled with the rest of the service, Facebook had a nice enough storage service, a lot of storage, an awful lot of bandwidth, and it wasn't a lot of work for that team to also support WhatsApp's needs; also our Erlang MMS (and the PHP version before it) was built around storing files on specific, addressable nodes, but nodes at Facebook are much more ephemeral and not easy to directly address by clients d) some amount of data storage moved off of mnesia into other Facebook data storage technology; again, not because mnesia wasn't sufficient, but more ephemeral nodes makes it cumbersome (addressable) and the available hardware nodes at FB didn't really match --- there's a very firm bias at FB towards using standard node sizes and the available standard nodes were like a web machine with not much ram or a big database machine with more ram and tons of fast storage; WA mnesia wants lots of ram but doesn't need a lot of storage (all data is in ram, and dumped + logged to disk) so there was a mismatch there --- things that stayed in mnesia needed much larger clusters to manage data size Presentations became less common because of more layers to get approval, and also because it's less fun to share how we built something on top of proprietary layers that others don't really have access to. Anybody could have gotten dual 2690 servers at SoftLayer and run a nice Erlang cluster. Only a few people could run an even bigger chat cluster in a Facebook like hosting environment.

toast0 2y ago

> I really have no idea what kind of system would require 1000 nodes, that couldn't be replaced by 100, 10x larger, nodes instead. Nodes only get so big. Way back when, a quad xeon 4650 v2 was definitely not 2x the throughput of a dual xeon 2690 v2; so you end up with two dual socket systems instead of one quad socket. A quad socket server often costs significantly more than two dual socket servers, and is likely to take longer between order and delivery. There's usually a point where you can still scale up, but scaling out is a better use of resources.