cross-posted from: https://lemmy.g97.top/post/761
cross-posted from: https://lemmy.g97.top/post/723
Hi! I spawned my own instance of lemmy on my server and I discovered new things about how lemmy and federation works, and I have a lot of doubt. I don't know exactly if those doubts are problems of my implementation of if they are normal, so!
- My main account is on lemmy.world and I see that new posts from communities I follow show up before on lemmy.world and then on my instance. Is it normal?
- With comments happens the same thing and they are slower to "sync". Why?
- If a community has been never discovered from the search form with the full format !community@instance, it will never appear on my instance. This means that is not possible to search for an argument (i.e. steam deck) and finding all the posts and communities about it. Is this normal or a feature that we/you would like to see in future/is adaptable to the concept of the fediverse? Because if I am on a big instance with a lot of users maybe I found that specific community or post, but on smaller instances like mine it will never appear If I don't know the exact name.
- I created a community on my instance and subscribed it from lemmy.world but I don't see any post nor are they in sync. Why? https://lemmy.world/c/[email protected] vs https://lemmy.g97.top/c/announcements.
- From my instance I am unable to follow lemmy.ml communities (they are pending, usually on lemmy.world the pending status is faster)
- I am unable to search for communities on Kbin.social, and when I try I see this log message of type "couldnt_find_object: error decoding response body: missing field
propertiesat line 1 column 206" from my docker instance:
2023-06-20T22:02:16.056226139Z 2023-06-20T22:02:16.055937Z ERROR HTTP request{http.method=GET http.scheme="https" http.host=lemmy.g97.top http.target=/api/v3/ws otel.kind="server" request_id=8211e6a4-2b30-4f8c-98b3-d93843a0e293 http.status_code=101 otel.status_code="OK"}: lemmy_server::api_routes_websocket: couldnt_find_object: error decoding response body: missing fieldpropertiesat line 1 column 206 2023-06-20T22:02:16.056276976Z 0: lemmy_apub::fetcher::search::search_query_to_object_id 2023-06-20T22:02:16.056286500Z at crates/apub/src/fetcher/search.rs:17 2023-06-20T22:02:16.056293804Z 1: lemmy_apub::api::resolve_object::perform 2023-06-20T22:02:16.056300316Z with self=ResolveObject { q: "[[email protected]](/c/[email protected])", auth: Some(Sensitive) } 2023-06-20T22:02:16.056307712Z at crates/apub/src/api/resolve_object.rs:21 2023-06-20T22:02:16.056314152Z 2: lemmy_server::root_span_builder::HTTP request 2023-06-20T22:02:16.056320693Z with http.method=GET http.scheme="https" http.host=lemmy.g97.top http.target=/api/v3/ws otel.kind="server" request_id=8211e6a4-2b30-4f8c-98b3-d93843a0e293 http.status_code=101 otel.status_code="OK" 2023-06-20T22:02:16.056351870Z at src/root_span_builder.rs:16
- I have a lot of warnings in the lemmy log of type "Error encountered while processing the incoming HTTP request: lemmy_server::root_span_builder: Header is expired" such as:
2023-06-20T21:58:12.484449111Z 2023-06-20T21:58:12.484275Z WARN Error encountered while processing the incoming HTTP request: lemmy_server::root_span_builder: Header is expired 2023-06-20T21:58:12.484510012Z 0: lemmy_server::root_span_builder::HTTP request 2023-06-20T21:58:12.484517559Z with http.method=POST http.scheme="https" http.host=lemmy.g97.top http.target=/inbox otel.kind="server" request_id=caf194c5-cac3-4c37-a29c-577d65deb050 http.status_code=400 otel.status_code="OK" 2023-06-20T21:58:12.484525578Z at src/root_span_builder.rs:16 2023-06-20T21:58:12.484530286Z LemmyError { message: None, inner: Header is expired, context: "SpanTrace" }I have more questions/doubt but for now this is enough I think! Thank you!
Header is expired issue is big part of the current federation problem. And whether you know it like it or not, you’ve just made the matter worse. You’re not to blame though. I’ve done it too, along with many other people self hosting our own instance.
The way federation currently works is each write action must be federated outwards to each federated instance. A comment reply, such as this one, must be federated outwards by the hosting instance. An instance receiving a federation event must also discard messages that are older than 10 seconds.
Here lies the problem… popular instances like lemmy.world and lemmy.ml has thousands of users, and thousands of federated servers. Yesterday, when I checked, lemmy.world had 3600 users per day and 2200+ federated servers. If there’s a really popular post on a very popular community, and 10% of the users comments on it? Lemmy.world server must send 360x2200 = 700K+ outbound federation event messages. Each one of these are sent over HTTPS via TCP so they can’t send all of them at the same time, and the messages are put into a queue where the federation workers will send them out. Each worker will send the message and because HTTPS is over TCP, it is not fire and forget, the worker must wait for acknowledgement for the packets. If an instance owner gets bored because they’re not getting all the messages and shuts down? Now the worker needs to wait for that to error out and thereby delaying messages further down the queue. If it had to wait more than 10 seconds? Everyone down the queue will just get expired messages because the event is already outdated.
So now you’ve already created an instance and adding to the load of the network, just like me, what can you do? Keep your server online in a fast data center. Use Cloudflare to reduce latency. That way at least your server isn’t going to introduce too much latency to other servers down the queue. Hopefully the devs figure out something to make the process better. I’ve put in a more scalable notification fleet architecture change on GitHub already. Lets see if they can implement that or change other requirements on the system.
Can you link your proposed change? I am interested
EDIT: for what concerns the queue stuff and the 10 second expiration. Is this part of the activityPub protocol or it has been choose by Lemmy's devs during the implementation of the protocol?
Also this is my GitHub issue ticket that will allow even larger servers to exist after the 5 minutes extension limit exhausts.
https://github.com/LemmyNet/lemmy/issues/3230
As an alternative, I opened up a discussion for a better purposed architecture design.
https://github.com/LemmyNet/lemmy/issues/3245
I believe the 10 seconds expiration is Lemmy thing. The ActivityPub protocol doesn’t dictate 10 seconds expiry. There is currently a change in activitypub-federation-rust (line 70), I don’t know when that will be released but it may help.
As someone who has just enough knowledge to know how big the task of creating a performant way to propagate updates through the federation is, I really hope there are some smart people working on a solution. That is the biggest advantage reddit has over lemmy: Known and centralized hardware standards. Lemmy needs to find a way to make propagation work when half of all instances are hosted at home on consumer-grade hardware.
Isn't there a mechanism to remove timing out servers? Or a way to unregister your instance ? Otherwise the model could never scale properly as servers get retired every now and then, even within the same instance.
If there is an option to adjust/disable it, I wasn't able to find it.
This commit changes the timeout to 1 day. I assume 0.18 will ship it, though I haven't checked.
Do you happen to know how the server-to-server connections are managed? I'm not too familiar with it, but it seems like HTTP/3 might provide some benefits for server-to-server communication.
Also, regarding queuing federation messages, I'm curious if packages like Kafka or Pulsar have been considered? They aren't typically used over HTTP, but it doesn't seem like it would be too hard to adapt, and the stream retention policy could be set to allow consumers to pick up older records as they have capacity (to avoid the issue around servers getting out of sync. The consumer would know the queue offset for each stream it was consuming and could pick up records as it has capacity, provided it doesn't fall so far behind that the records expire). Publishers could provide separate topics for different message types to allow consumers to prioritize activity types (for example, prioritizing receiving replies over up/down votes). Also servers could potentially use cluster replication (Mirror Maker) to handle moving activity records from one server to another (again, HTTP-only would be an issue here), and each server could then consume the federation activity messages locally from its own queue.
Kafka/Pulsar support have strong scaling support, so adding capacity for federation messages should be fairly straightforward.
I've only used Kafka once, and I'm completely unqualified to operate an instance of any complexity, by in general my experience with it was pretty good.
What is making the matter worse, is everyone clobbering together on lemmy.ml, and lemmy.world. This causes those particular servers to be vastly overloaded.
If, say, people created communities ELSEWHERE, the load could be spread-around.
Not- saying the architecture of federation isn't a problem, as indeed, it is a huge problem- but, in the interim time, this can be helped out by people spreading out.