Deleting a post vs deleting an entire comment tree
Technical Discussion
65
Posts
15
Posters
69
Views
-
I've got a PR for removing an entire comment thread for Lemmy mostly ready, I just need to write the tests and decide on how this exact problem should work, so the timing is honestly impeccable. The things I've been debating between are:
- Same as option 1, a new property like
removeChildren; - Make the
targetof theDeletean array. I'm honestly 50/50 on whether this is actually spec compliant, it's not clear to me that it isn't; - Just sending out an individual
Deletefor everyObject. I like this option the least, but it is very easy to implement and needs no changes in other software to work.
Either of those first two sound fine.
- Same as option 1, a new property like
-
@julian @rimu @nutomic @melroy @BentiGorlich
#hubzilla and #friendica had always have the discussion tree
-
Can you clarify a bit what this means?
I'm imaging the following scenario, i.e.
Givengherkin steps,- Alice on her own server Abel.
- Bob on the forum server Forum.
- Alice created a post in a category of Forum. Let's call it "1000 reasons to hug cows". id
http://abel.example/alice/post. - Bob, a million other trolls, replies with "To tip them over! lol!". Bob's post has id
http://forum.example/bob/post/id.
The
Whenblock is simple. Alice deletes her post, it's a root post (with lots of stupid replies). This means aDeleteactivity is send for the object withhttp://abel.example/alice/post.Now, what does this imply? Obviously, we have "The forum sends an
Announce(Delete).". But do we also have:- The forum sends a
Delete(http://forum.example/bob/post/id)as Bob? - The forum sends a
Announce(Delete(http://forum.example/bob/post/id))as the forum?
How does the forum handle posts not made by its users? It obviously cannot delete them! So is there like an
Undo(Announce(Create(Object)))or aRemove(Object)? -
@thisismissem not even orphaned, a delete leaves behind a tombstone object which should still be dereferenceable from anywhere else.
-
@julian @rimu @nutomic @melroy @BentiGorlich
I'm still confused that so many people have thought that using Announce as a solution for something which is at its core, an addressing problem is a good way to do it.
I guess I might not be aware of some subtleties, but why doesn't the original Delete reach all the interested parties? Why is there a need for Announce-ing it in the first place?
-
@julian @rimu @nutomic @melroy @BentiGorlich
I'm still confused that so many people have thought that using Announce as a solution for something which is at its core, an addressing problem is a good way to do it.
I guess I might not be aware of some subtleties, but why doesn't the original Delete reach all the interested parties? Why is there a need for Announce-ing it in the first place?
Lets say you have a user on instance alpha, and a community on instance beta, with followers on various other instances. The user makes a post in the community, but how can he know who the community followers are to send it directly to them? And how could the followers trust that an activity sent directly was actually approved by the community?
Another solution for this would be inbox forwarding, but by now its too late to change the behaviour of all existing platforms.
-
If that object is the root-level node, and it is deleted, everything below it is also deleted.
This is not exactly true. When Lemmy receives
Delete/Objectfor a (top level) post, it only marks the post itself as deleted. The replies are not marked as deleted in the database, but are (currently) hidden in the UI. For 1.0 this behaviour is already changed in the API and UI so that deleted posts are removed from listings, but accessible by direct link so that comments can be viewed. This doesnt involve any change to federation and is an implementation detail.Adding a new activity type for each such implementation detail is not feasible, and we cannot support all of them. So to maximize compatibility you should definitely keep using
Delete(Object), and add a new property if you need to be explicit about it. -
I've got a PR for removing an entire comment thread for Lemmy mostly ready, I just need to write the tests and decide on how this exact problem should work, so the timing is honestly impeccable. The things I've been debating between are:
- Same as option 1, a new property like
removeChildren; - Make the
targetof theDeletean array. I'm honestly 50/50 on whether this is actually spec compliant, it's not clear to me that it isn't; - Just sending out an individual
Deletefor everyObject. I like this option the least, but it is very easy to implement and needs no changes in other software to work.
This is slightly different from OP as you are talking about deleting a reply (ie
Note) with children, while OP is about deleting a top-level post with children. Nevertheless both can be represented in the same way over federation.Just sending out an individual Delete for every Object. I like this option the least, but it is very easy to implement and needs no changes in other software to work.
This would be terrible for performance when removing dozens or hundreds of comments at once. Rule of thumb should be one activity for one user action.
Make the target of the Delete an array. I’m honestly 50/50 on whether this is actually spec compliant, it’s not clear to me that it isn’t;
The problem with this is that some platforms might get the idea to delete multiple unrelated comments with a single activity, or even comments in different posts. Handling that would make the receiving logic unnecessarily complicated, and would also make it complicated to combine the modlog entries.
Same as option 1, a new property like removeChildren;
This is the simplest and best option.
- Same as option 1, a new property like
-
This is slightly different from OP as you are talking about deleting a reply (ie
Note) with children, while OP is about deleting a top-level post with children. Nevertheless both can be represented in the same way over federation.Just sending out an individual Delete for every Object. I like this option the least, but it is very easy to implement and needs no changes in other software to work.
This would be terrible for performance when removing dozens or hundreds of comments at once. Rule of thumb should be one activity for one user action.
Make the target of the Delete an array. I’m honestly 50/50 on whether this is actually spec compliant, it’s not clear to me that it isn’t;
The problem with this is that some platforms might get the idea to delete multiple unrelated comments with a single activity, or even comments in different posts. Handling that would make the receiving logic unnecessarily complicated, and would also make it complicated to combine the modlog entries.
Same as option 1, a new property like removeChildren;
This is the simplest and best option.
@nutomic > Make the target of the Delete an array
It technically is because of JSON-LD (everything can be an array), however, software may not support it because it's pretty common for devs to treat AP as JSON instead of JSON-LD, so doing an assessment of compatibility would be suggested
-
@julian @rimu @nutomic @melroy @BentiGorlich
#hubzilla and #friendica had always have the discussion tree
@tofeo :verified: @julian ...where "always" means "since before there was even Mastodon".
A side-effect of their model, present at least on Hubzilla and Hubzilla's descendants, including still existing (streams) and Forte, is that comments/replies cannot exist in a stream without a) a parent and b) a start post. On all of them, including Friendica, it isn't a post if it replies to something, very much unlike Mastodon where a thread is a bunch of posts.
Depending on whom you ask, a conversation looks either like this:- Post
- Comment
- Comment
- Comment
- Comment
- Comment
- Comment
- Comment
- Comment
- Comment
- Post
-
Lets say you have a user on instance alpha, and a community on instance beta, with followers on various other instances. The user makes a post in the community, but how can he know who the community followers are to send it directly to them? And how could the followers trust that an activity sent directly was actually approved by the community?
Another solution for this would be inbox forwarding, but by now its too late to change the behaviour of all existing platforms.
@nutomic I solved this by having an the instance be an actor that is an intermediary for all operations. All activities get CC'ed to the followers of the user actor, and of the instance actor. (replace instance with community/group, whatever you use as an aggregate element for your implementation)
This should cover all interested parties imho.
The potential downside is that instances need to explicitly operate between themselves with follow operations (which conveniently also solves the problem of unwanted interactions with less savory corners of the fediverse).
Like I said, a problem of addressing. :P
-
@nutomic I solved this by having an the instance be an actor that is an intermediary for all operations. All activities get CC'ed to the followers of the user actor, and of the instance actor. (replace instance with community/group, whatever you use as an aggregate element for your implementation)
This should cover all interested parties imho.
The potential downside is that instances need to explicitly operate between themselves with follow operations (which conveniently also solves the problem of unwanted interactions with less savory corners of the fediverse).
Like I said, a problem of addressing. :P
@nutomic I realize this explanation probably skips a lot of information that's clear to me from my context, but I'll try to answer if anyone has any questions.
-
@thisismissem not even orphaned, a delete leaves behind a tombstone object which should still be dereferenceable from anywhere else.
mariusor@metalhead.club that works too!
-
thisismissem seems I misspoke, as Nutomic's reply above clarifies: the tree stays but is effectively orphaned. Lemmy v1.0 will allow the reply tree to be accessed post-deletion.
The original query does still remain the same: what would be the best way forward to explicitly signal the deletion (or technically, the removal) of an entire reply tree?
-
This is slightly different from OP as you are talking about deleting a reply (ie
Note) with children, while OP is about deleting a top-level post with children. Nevertheless both can be represented in the same way over federation.Just sending out an individual Delete for every Object. I like this option the least, but it is very easy to implement and needs no changes in other software to work.
This would be terrible for performance when removing dozens or hundreds of comments at once. Rule of thumb should be one activity for one user action.
Make the target of the Delete an array. I’m honestly 50/50 on whether this is actually spec compliant, it’s not clear to me that it isn’t;
The problem with this is that some platforms might get the idea to delete multiple unrelated comments with a single activity, or even comments in different posts. Handling that would make the receiving logic unnecessarily complicated, and would also make it complicated to combine the modlog entries.
Same as option 1, a new property like removeChildren;
This is the simplest and best option.
I do agree that
with_replies, or similar, would be the easiest approach, but I don't think it is the most specific.The bool suggests that all replies to a given object are deleted. However, you do not know whether your idea of what the reply tree is matches that of the originating server (which replies are included, etc.?)
Remove(Context), on the other hand does imply both that the container is deleted, and all of its replies, which are dereferenceable by resolving the context directly. It also has the benefit of being able to provide a pointer to where it was removed from, which is useful.So to me it's not just a matter of preference, but that there are additional benefits to
RemoveI will of course concede that it is more work to deliver
Remove.cc thisismissem
-
mariusor@metalhead.club let's ask helge@mymath.rocks what implementor support for
Delete(Array)looks like LOL -
mariusor@metalhead.club let's ask helge@mymath.rocks what implementor support for
Delete(Array)looks like LOL@julian as long as it's in the spec, I don't really care.
If we all bow to inferior implementations the ecosystem will stagnate. Mastodon's quirks have done enough damage in my opinion. My choice is to be brave and build for the future.
PS. Not to brag or anything (🤞) but my implementation can operate on activities with arrays as object, actor, etc.
One thing where you can lead the way (because the threadiverse would really benefit from it) is to accept arrays in inReplyTo (where you put all the ancestors of the current post, not just the parent).
-
@nutomic I solved this by having an the instance be an actor that is an intermediary for all operations. All activities get CC'ed to the followers of the user actor, and of the instance actor. (replace instance with community/group, whatever you use as an aggregate element for your implementation)
This should cover all interested parties imho.
The potential downside is that instances need to explicitly operate between themselves with follow operations (which conveniently also solves the problem of unwanted interactions with less savory corners of the fediverse).
Like I said, a problem of addressing. :P
I'm sure that approach works as well. This would have been worth discussing 4 or 5 years ago when I was just implementing federation in Lemmy for the first time. By now FEP-1b12 is already an established standard which is used by various platforms, and it would be completely unfeasible to replace it with something else.
-
I fail to see what the fundamental difference is. If you are unsure about the target with
Delete/Object, you can also resolve thecontextof Object to figure that out. Anyway the instance where theGroupis hosted is always the authority, so the state there is the correct one.Actually I would rather think of this from a different perspective, namely from the perspective of the mod who clicks the remove button. That would happen when a post is offtopic or violates the rules, and then the intent clearly is to remove all replies as they are not useful. It wouldnt make sense to leave up a single reply two levels deep just because it wasnt included in the context for some reason.
Gli ultimi otto messaggi ricevuti dalla Federazione
-
@reiver i think the disjunction between Object and Link was actually unnecessary. https://github.com/w3c/activitystreams/issues/666
i also think there's too much emphasis on types when there really shouldn't be -- it's the *properties* that you end up using almost all of the time. pretty much the only types that actually matter are the Activity types (because you can't infer those).
-
-
@reiver Did you and I discuss queryfy a while ago, or was it one of my other projects?
Just wondering whether I owe you a heads up since queryfy has been bumped up to v0.3.0
-
With ActivityPub / ActivityStreams...
To me, it feels like there should have been something that is a common parent of both 'Object' and 'Link'.
That just had the "name", "nameMap", and "preview" fields (along with "id" and "type, of course) — since that is what 'Object' and 'Link' share in common.
I'll just call this common parent: 'Entity'.
...
It could have even been an opportunity to talk about how to handle unknown types.
-
@soapdog@toot.cafe hmm... just thinking aloud here.
You posit in another post that the network effects inflate exponentially:
> Push models are resource hogs that approach exponential growth in a large network like the fediverse
That's not true. If you post a message then it sends a copy to each follower. That's linear growth. If you collapse recipients via shared inboxes you can reduce that further.
If you're referring to the torrent of requests that happen if your post is shared (the "thundering herd" problem) then that's actually a PULL happening from those requesting instances!
Secondly, in a pull model of AP, you would need to continually poll servers of all your followers so as to approach a real-time effect. You'd be polling servers over and over again, and many of them would have nothing new, with so much wasted traffic.
If your expectations include semi real-time updates, the push model is much more performant, in my humble opinion.
-
@evan @mariusor @silverpill i think we probably need to revisit the user story of creating multiple objects at once, or more accurately, the user story of minting and binding multiple identifiers at once.
-
@evan my suggestion: https://github.com/w3c/activitypub/issues/533
-
@evan @mariusor @silverpill re: ids though the RDF ecosystem (and jsonld) doesn't use "null", it uses blank node identifiers (those prefixed with _: are special cased by the prefix expansion algorithm). this can allow for "transient" activities or "anonymous" objects (and the graph data model auto assigns _:b1, _:b2 and so on when "id" is missing; the canonicalization algorithm assigns _:c14n0 and _:c14n1 and so on)
this is maybe not the best way to create replies collections though...
