> What I had missed is that we deployed a new internal service last week that sent less than three GetPostRecord requests per second, but it did sometimes send batches of 15-20 thousand URIs at a time. Typically, we'd probably be doing between 1-50 post lookups per request.
The incredible part about this is because their backend is all TCP/IP they were literally exhausting the ports by leaving all 65k of them in TIME_WAIT, and the workaround was to start randomizing the localhost address to give them another trillion ports or so.
The simple answer is that atproto works like the web & search engines, where the apps aggregate from the distributed accounts. So the proper analogy here would be like yahoo going down in 1999.
Google and MSN Search were already available at this time. Also websites used to publish webrings and there was IRC and forums to ask people about things.
It’s more of a concept of a plan for being distributed. I even went through the trouble of hosting my own PDC and still, I was unable to use the service during the outage
> The timing of these log spikes lined up with drops in user-facing traffic, which makes sense. Our data plane heavily uses memcached to keep load off our main Scylla database, and if we're exhausting ports, that's a huge problem.
Golang's use of a potentially unbounded number of threads is just insane. I used to be fairly bullish on golang, but this, combined with the fact that its garbage collected, makes me feel its just unsuitable for production use.
You can have this problem with any kind of thread -- including OS threads -- if you do an unbounded spawn loop. Go is hardly unique in this.
Goroutines are actually better AFAIK because they distribute work on a thread pool that can be much smaller than the number of active goroutines.
If my quick skim created a correct understanding, then the problem here looks more like architecture. Put simply: does the memcached client really require a new TCP connection for every lookup? I would think you would pool those connections just like you would a typical database and keep them around for approximately forever. Then they wouldn't have spammed memcache with so many connections in the first place...
(edit: ah, it looks like they do use a pool, but perhaps the pool does not have a bounded upper size, which is its own kind of fail.)
Rust's async doesn't have this issue. Or at least, it's the same issue as malloc in an unbounded loop, but that's a more general issue not related to async or threading.
15-20 thousand futures would be trivial. 15-20 thousand goroutines, definitely not.
Why does garbage collection make it unsuitable for production use? A lot of production software is written in garbage collected languages like Java. Pretty much the entire backend for iTunes/Apple Music is written in Java, and it's not doing any kind of fancy bump allocator tricks to avoid garbage. In my mind, kind of hard to argue that Apple Music is not "production use".
There are certainly plenty of projects where garbage collection is too slow, but I don't know that they're the majority, and more people would likely prefer memory safety by default.
Email and the internet don't have "downtime." Certain key infra providers do of course. ISPs can go down. DNS providers can go down. But the internet and email itself can't go down absent a global electricity outage.
You haven't built a decentralized network until you reach that standard imo. Otherwise its just "distributed protocol" cosplay. Nice costume. Kind of like how everybody has been amnesia'd into thinking Obsidian is open source when it really isn't.
The comparison here is to something like TCP/IP. TCP/IP never goes down. TCP/IP is a protocol, the servers may go down and cause disruption, but the protocol doesn't really have the ability to "go down".
Nostr is also a protocol. The communication on top of Nostr is pretty resilient compared to other solutions though, so that's the main highlight here.
If tens of servers go down, then some people may start noticing a bit of inconvenience. If hundreds of servers go down, then some people may need to coordinate out of bound on what relays to use, but it still generally speaking works ok.
Wasn't aware there are ~2k relays now. Have inter-relay sharing situation improved?
When I tried it long time ago, the idea was just a transposed Mastodon model that the client would just multi-post to dozen different servers(relays) automatically to be hopeful that the post would be available in at least one shared relays between the user and their followers. That didn't seem to scale well.
Getting clients to do the right thing is like herding cats, but there has been some progress. Early 2023 Mike Dilger came up with the "gossip model" (renamed "outbox model" for obvious reasons). Here's my write-up: https://habla.news/hodlbod/8YjqXm4SKY-TauwjOfLXS
The basic idea is that for microblogging use cases users advertise which relays their content is stored on, which clients follow (this implies that there are less-decentralized indexes that hold these pointers, but it does help distribute content to aligned relays instead of blast content everywhere).
Also, relays aside, one key difference vs ActivityPub is that no third party owns your identity, which means you can move from one relay to another freely, which is not true on Mastodon.
> What I had missed is that we deployed a new internal service last week that sent less than three GetPostRecord requests per second, but it did sometimes send batches of 15-20 thousand URIs at a time. Typically, we'd probably be doing between 1-50 post lookups per request.
That’ll do it.
Ahh, the three relevant numbers in development: 0, 1, and infinity.
The incredible part about this is because their backend is all TCP/IP they were literally exhausting the ports by leaving all 65k of them in TIME_WAIT, and the workaround was to start randomizing the localhost address to give them another trillion ports or so.
Zero, one, many, many thousands.
And then they fix the issue by using multiple localhost IPs rather than, perhaps, not sending 15-20 thousand URIs at a time
They mentioned it was a temporary fix that they removed after finding and fixing the true root cause, though.
less than ideal if I had to be frank.
I don't really understand this architecture, but I thought Bluesky was distributed like Mastodon? How can it have an outage?
This writeup is useful for backend engineers: https://atproto.com/articles/atproto-for-distsys-engineers
The simple answer is that atproto works like the web & search engines, where the apps aggregate from the distributed accounts. So the proper analogy here would be like yahoo going down in 1999.
This is a fantastic write-up, thanks for sharing!
Google and MSN Search were already available at this time. Also websites used to publish webrings and there was IRC and forums to ask people about things.
It’s more of a concept of a plan for being distributed. I even went through the trouble of hosting my own PDC and still, I was unable to use the service during the outage
Mastodon infra can have outages, too.
It's just confined to one instance if it goes down, not all of Mastodon.
It's not really distributed. It's a centralised service that pulls some parts of 0.01% of user profiles from their own servers.
A web interface and home server can have an outage. Bluesky is just a web interface and home server.
At least they aren't hiding and transparent about it unlike the big tech corps with so called SLAs
> The timing of these log spikes lined up with drops in user-facing traffic, which makes sense. Our data plane heavily uses memcached to keep load off our main Scylla database, and if we're exhausting ports, that's a huge problem.
I expect this is common.
Tell us more about this buggy "new internal service" that's scraping batch data :P
Golang's use of a potentially unbounded number of threads is just insane. I used to be fairly bullish on golang, but this, combined with the fact that its garbage collected, makes me feel its just unsuitable for production use.
You can have this problem with any kind of thread -- including OS threads -- if you do an unbounded spawn loop. Go is hardly unique in this.
Goroutines are actually better AFAIK because they distribute work on a thread pool that can be much smaller than the number of active goroutines.
If my quick skim created a correct understanding, then the problem here looks more like architecture. Put simply: does the memcached client really require a new TCP connection for every lookup? I would think you would pool those connections just like you would a typical database and keep them around for approximately forever. Then they wouldn't have spammed memcache with so many connections in the first place...
(edit: ah, it looks like they do use a pool, but perhaps the pool does not have a bounded upper size, which is its own kind of fail.)
Rust's async doesn't have this issue. Or at least, it's the same issue as malloc in an unbounded loop, but that's a more general issue not related to async or threading.
15-20 thousand futures would be trivial. 15-20 thousand goroutines, definitely not.
Why does garbage collection make it unsuitable for production use? A lot of production software is written in garbage collected languages like Java. Pretty much the entire backend for iTunes/Apple Music is written in Java, and it's not doing any kind of fancy bump allocator tricks to avoid garbage. In my mind, kind of hard to argue that Apple Music is not "production use".
There are certainly plenty of projects where garbage collection is too slow, but I don't know that they're the majority, and more people would likely prefer memory safety by default.
Everything is understood by comparison. Unsuitable for production use, compared to what is the more apt question.
Distributed social media goes down? hrmmm.
Email and the internet don't have "downtime." Certain key infra providers do of course. ISPs can go down. DNS providers can go down. But the internet and email itself can't go down absent a global electricity outage.
You haven't built a decentralized network until you reach that standard imo. Otherwise its just "distributed protocol" cosplay. Nice costume. Kind of like how everybody has been amnesia'd into thinking Obsidian is open source when it really isn't.
Bluesky is a provider. Blacksky didn’t go down.
Is there anything running on Blacksky other than Bluesky with more than say, 100 active users?
AOL never even got to that level of dominance in the internet 1.0 era.
The point is it's not a distributed network if one node is 99.9% of all traffic.
Great write up... curious about the RCA. Thanks!
Thank you for the post mortem on this outage.
nostr never goes down
If nostr went down would people even notice?
If any major nostr relay goes down, no one notices. That has happened many times, the network is very resilient to that.
probably not
All support to other decentralizers but nothing never goes down.
The comparison here is to something like TCP/IP. TCP/IP never goes down. TCP/IP is a protocol, the servers may go down and cause disruption, but the protocol doesn't really have the ability to "go down". Nostr is also a protocol. The communication on top of Nostr is pretty resilient compared to other solutions though, so that's the main highlight here.
If tens of servers go down, then some people may start noticing a bit of inconvenience. If hundreds of servers go down, then some people may need to coordinate out of bound on what relays to use, but it still generally speaking works ok.
1000x redundancy makes it vanishingly unlikely. Although I know we're due for a pole shift so all bets are off I suppose.
Wasn't aware there are ~2k relays now. Have inter-relay sharing situation improved?
When I tried it long time ago, the idea was just a transposed Mastodon model that the client would just multi-post to dozen different servers(relays) automatically to be hopeful that the post would be available in at least one shared relays between the user and their followers. That didn't seem to scale well.
Getting clients to do the right thing is like herding cats, but there has been some progress. Early 2023 Mike Dilger came up with the "gossip model" (renamed "outbox model" for obvious reasons). Here's my write-up: https://habla.news/hodlbod/8YjqXm4SKY-TauwjOfLXS
The basic idea is that for microblogging use cases users advertise which relays their content is stored on, which clients follow (this implies that there are less-decentralized indexes that hold these pointers, but it does help distribute content to aligned relays instead of blast content everywhere).
Also, relays aside, one key difference vs ActivityPub is that no third party owns your identity, which means you can move from one relay to another freely, which is not true on Mastodon.
Did all 3 users notice?
Naw, only one did. Turns out the other two were his socket accounts he used to upvote and comment on his own content.
Okay, nuff trolling for today
Lite Blue on a dark Blue background. That is a new one, I have seen grey text on lite grey, but blue on blue ?
The article does work in lynx, at least I can read it.