Jetstream: Shrinking the AT Protocol Firehose by >99%
jazco.dev
jazco.dev
There's not really a good reason to do compression if the stream is just going to be consumed locally. Instead you can skip that step and broadcast over memory to the other local services.
- Jetstream(1) (no compression) and
- Jetstream(2) (zstd compression).
And my comment means (1) not really needed, except some specific scenarios
FWIW the default mode is uncompressed unless the client explicitly requests compression with a custom header. I tried using per-message-deflate but the support for it in the websocket libraries I was using was very poor and it has the same problem as streaming compression in terms of CPU usage on the Jetstream server.
Would anybody realistically be using those tools with this volume of data, for anything but testing?
If everyone is on one server (remains to be seen), and all the bots blindly trust it because they are cheap and lazy, what the hell is the point?
Reduction first may show the compression is less useful. Verbose, human friendly protocols compressed win out in maintenance tasks, and it’s a marathon not a sprint.
It’s sort of like the right to fork in Open Source; it doesn’t mean people fork all the time or verify every line of code themselves. There’s still trust involved.
I wonder if some security features could be added back, though?
Even internally, BlueSky runs a number of "servers." It's already internally federated. And you can transfer your account to your own server if you want to, though that is very much still beta quality, to be fair.
You're not really "on a server" in the same sense as other things. It's closer to "I create content addressable storage" than "The database on this instance knows my username/password."
DID=$(curl -s https://<username>.bsky.social/.well-known/atproto-did)
curl https://plc.directory/$DID | jq '.service[0].serviceEndpoint'
dig +short _atproto.example.com TXTI'm also currently trying to understand the tradeoffs for did:plc more. It's unclear to me just how centralized it is. Will it always require a single central directory, or is it more like Certificate Transparency? Based on what I've heard about the recover process, I believe it's the latter, but I still need to dig into it more.
I wonder if a rewrite of this in C++ would even bump further the performance and optimise the overall system.
>Jetstream is a streaming service that consumes an AT Proto com.atproto.sync.subscribeRepos stream and converts it into lightweight, friendly JSON.
So let me get this straight. if you did want to run Jetstream yourself you'd still need to be able to handle the 232 GB/day of bandwidth?
This always has been my issue with Bluesky/AT Protocol, For all the talk about their protocol being federated, It really doesn't seem realistic for anyone to run any of the infrastructure themselves. You're always going to be reliant on a big player that has the capital to keep everything running smoothly. At this point I don't really see how it's any different then being on any of the old centralized social media.
If you want large scale social networks, you need to work with a large scale of data. Since federated open queries aren’t feasible, you need big machines.
If you want a smaller scale view of the network, do a crawl of a subset of the users. That’s a perfectly valid usage of atproto, and is how ActivityPub works by nature.
That is good, but it's still a centralized source of truth.
>If you want large scale social networks, you need to work with a large scale of data. Since federated open queries aren’t feasible, you need big machines.
Thats just simply not true. ActivityPub does perfectly without the need of any bulky machine or node acting as a relay for the rest of the network. Every single ActivityPub service only ever interacts with other discovered services. Messages aren't broadcast through a central firehose, they're sent directly to who needs to receive them. This is a fundamental difference with how both protocols work. With ATProto you NEED to connect to some centralized relay that will broker your messages for you. With ActivityPub, there is no middle man, Instances just talk directly to each other. This is why ActivityPub has a discovery problem by the way, but it's just a symptom of real federation.
>and is how ActivityPub works by nature.
It's not. See Above.
> Thats just simply not true.
> [snip]
> This is why ActivityPub has a discovery problem by the way, but it's just a symptom of real federation.
You're actually agreeing with them! The "discovery problem" is because "federated open queries aren't feasible".
> With ATProto you NEED to connect to some centralized relay that will broker your messages for you.
You can connect to PDSs directly to fetch data if you want; this is exactly what the relays do!
If you want to build a client that behaves more like ActivityPub instances, and does not depend on a relay, you could do so:
- Run your own PDS locally, hosting your repo.
- Your client reads your repo via your PDS to see the accounts you follow.
- Your client looks up the PDSs of those accounts (which are listed in their DID documents).
- Your client connects to those PDSs, fetches data from them, builds a feed locally, and displays it to you.
This is approximately a pull-based version of ActivityPub. It would have the same scaling properties as ActivityPub (in fact better, as you only fetch what you need, rather than being pushed whatever the origins think you need). It would also suffer from the same discovery problem as ActivityPub (you only see what the accounts you follow post).
At that point, you would not be consuming any of the _output_ of a relay. You would still want relays to connect to your PDS to pull data into their _input_ in order for other users to see your posts, but that's because those users have chosen to get their data via a relay (to get around the discovery problem). Other users could instead use the same code you're using, and themselves fetch data directly from your PDS without a relay, if they wanted to suffer from the discovery problem in exchange for not depending on a relay.
The core of the issue is, Bluesky's current model is unsustainable. The cost of running the main relay is going to keep rising. The barrier to discovery keeps getting higher and higher. It might cost 150$/month now to mirror the relay, but what's going to happen when it's 1000$?
https://en.wikipedia.org/wiki/Bluesky
By your own numbers it averages 2.7 MB/s. This is manageable with a good cable internet connection or a small vps. This is a small number for 5.5 million active users.
What happens if it expands to 10 times its current active users? Who knows, maybe only the 354 million people around the world with access to gigabit broadband can run a full server at home and the rest of the people and companies that want to run full servers will have to rent a vps.
https://www.telecompetitor.com/gigabit-availability-report-1...
The point here is that this is not a practical problem. How many of these servers do you really need?
It's not. It's a trustless aggregator. The PDSes are the sources of truth, and you can crawl them directly. The relay is just an optimization.
> Messages aren't broadcast through a central firehose
ATProto works like the web does. People publish information on their servers, and then relays crawl them and emit their crawl through a firehose.
> ActivityPub does perfectly without the need of any bulky machine or node acting as a relay for the rest of the network
ActivityPub doesn't do large scale aggregated views of the activity. The peer-wise exchanges means that views get localized; this is why there's not network-wide search, metrics, or algorithms.
> This is why ActivityPub has a discovery problem by the way,
right
> but it's just a symptom of real federation.
"real" ?
This is such a massive understatement. the relay is the single most important piece in the entire Bluesky stack.
Let me ask you this, is it possible for me to connect to a PDS directly, right now, via the bluesky app? or is this something that will be possible in the future?
>ATProto works like the web does. People publish information on their servers, and then relays crawl them and emit their crawl through a firehose.
>ActivityPub doesn't do large scale aggregated views of the activity.
So are relays really just an optimization or an integral part of how ATProtocol is supposed to work? ActivityPub doesn't require relays to function properly. This is why I say it's real federation. You can't truly be federated if you require centralization.
I’m not so sure: isn’t the certificate transparency log a pretty good example of a federated group of disparate members successfully sharing a view of the world?
That requires some form of centralization to be useful (else it’s not really a log, more of a series of disconnected scribbles), and it’s definitely a true federated network.
Well, yes, that's what you do when you log in. If you open devtools you'll see that the app is communicating with your PDS.
> So are relays really just an optimization or an integral part of how ATProtocol is supposed to work?
I think the issue here is that you're mentally slicing the stack in a different way that atproto does. You expect each node the be a full instance of the application, and that the network gets partitioned by a bunch of applications exchanging peerwise.
A better mental model of atproto is that it's a network of cross-org microservices. https://atproto.com/articles/atproto-for-distsys-engineers gives a decent intuition about it.
That's pretty cool. Does bsky.app do a DID resolve or does it have a faster back channel for determining PDS addresses for bsky.social?
If someone wants to run a server, they probably would pay for a VPS with a gigabit connection, which would be able to do 120 MB/s.
You might need to pay for extra bandwidth, but it is probably less than a night out every month.
I run a custom rust-based firehose consumer on the main firehose using a cheap digitalocean droplet and don't cross 15% CPU usage even during the peak 33 Mb/s bandwidth described in this article.
The core team seem to put a lot of intention into being able to host different parts of the network. The most resource intensive of which is the relay which produces the event stream that Jetstream and others consume from. One of their devs did a breakdown of that and showed you could run it at $150 per month which is pricey but not unattainable with grants or crowdfunding https://whtwnd.com/bnewbold.net/entries/Notes%20on%20Running...
It wasn't even a criticism, just an observation for anyone else who was thinking the same or was interested in another popular project with a similar name (and seemingly similar functions? didn't look too hard).
Naming things is hard and we all kinda share one global tech namespace, so this is gonna inevitably happen.
The firehose is all public data going into the network right?
Isn't that pretty tiny for a worldwide social network?
And the fact one country can cause a 10x surge in traffic also suggests its worldwide footprint must be tiny...
Every year or so, you plop in one more 10TB SSD.
I aggregated some stats when we hit 10M users here - https://bsky.app/profile/did:plc:q6gjnaw2blty4crticxkmujt/po...
Lots of these numbers are public and the impact of the surge can be seen here: https://bskycharts.edavis.dev/static/dynazoom.html?plugin_na...
Note the graphs in that link only show users that take a publicly visible action (i.e. post, like, follow, etc.) and won't show lurkers at all.
It's the "main subset" of the public data, being "events on the network": for the Bluesky app that's posts, reposts, replies, likes, etc. Most of those records are just metadata (e.g. a repost record references the post being reposted, rather than embedding it). Post / reply records include the post text (limited to 300 graphemes).
In particular, the firehose traffic does _not_ include images or videos; it only includes references to "blob"s. Clients separately fetch those blobs from the PDSs (or from a CDN caching the data).
Introduce your jargon before expositing, please.
https://en.wikipedia.org/wiki/Iridium_satellite_constellatio...
https://en.wikipedia.org/wiki/Bluesky
And the protocol is documented here:
Hope this helps with the confusion.
SSE Benefits: Standard HTTP protocol, Built-in gzip compression, Simpler client implementation
As for application development, I think web socket APIs are generally exposed much better and used much easier than SSEs. I agree that SSEs are a more appropriate technology to use here, but they're used so little that I don't think the tooling is good. Just about every language has a dedicated websocket client library, but SSEs are usually implemented as a weird side effect of a HTTP connection you need to keep alive manually.
The stored ZSTD objects make sense, as you only need to compress once rather than compress for every stream (as the author details). It also helps store the data collected more efficiently on the server side if that's what you want to do.
The dictionary allows for better compression without needing a large amount of data, and sending every client the same compressed binary data saves a lot of CPU time in compression. Streams, usually, require running the compression for each client.
Also, this was a fun thing I built mostly in my free time :)
Also, it's just a good look towards the community. Remember, ideally not all of ATProto's infrastructure is run by BlueSky themselves, but by a diverse set of folks who want to be able to control their own data. These are more likely to be individuals or small groups, not capitalized startups. While the protocol itself is designed to be able to do this, "what happens when BlueSky is huge and running your own infra is too expensive" is a current question that some folks are reasonable skeptical about: it's no good being a federated protocol if nobody can afford to run a node. By doing stuff like this, the BlueSky team is signaling that they're aware of and responsive to these concerns, so don't look at it as trying to save money on bandwidth: look at it as a savvy way to generate some good marketing.
That's my take anyway.
EDIT: scrolling down, you can see this kind of sentiment here: https://news.ycombinator.com/item?id=41637307