Skip to content
Open
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
220 changes: 220 additions & 0 deletions proposals/4194-user-redaction.md
Copy link
Copy Markdown
Member

@tulir tulir Sep 15, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Implementation requirements:

  • Server implementing the endpoint
  • Client or bot using the endpoint

Original file line number Diff line number Diff line change
@@ -0,0 +1,220 @@
# MSC4194: Batch redaction of events by sender within a room (including soft failed events)

Targeted attacks are a problem in Matrix. Typically throwaway users
are used to target matrix rooms with abuse in the form of multiple
messages. The intent of the spammer is to fill the room timeline view
in clients with abuse.

Typically, a moderator will be pinged by a bystander, and the
moderator will respond by first banning the throwaway account and then
issuing redactions via a client for each event the throwaway account
sent.

Due to the architecture of various Matrix APIs, there are multiple
issues that can occur in the moderator's flow.

1. Calling `/messages` is required for some clients to accurately get
the full list of `event_id`s the throwaway account sent. This is a
compounding problem because homeservers can be slow to respond to a
filtered `/messages` request and then the client must still process
the response and issue a redaction for each event returned.
This is how both Mjolnir and Draupnir source events for redaction.

2. The client must call `/rooms/redact/event` for each target event,
requiring the client to make multiple requests, further impeding
response time. Homeservers typically rate limit the number of
redactions that a client can send to `/redact` evenly with `/send`,
which further compounds response time.
[MSC2244](https://github.com/matrix-org/matrix-spec-proposals/pull/2244)
tried to alleviate part of the problem by allowing the redactions
to be sent in one event. But there currently are no endpoints that
allow you to create this event[^create-mass-redaction].

3. Soft failure of events sent by the throwaway account after the
moderator bans the throwaway. These events are not visible to the
moderator's matrix client, and there is no way for a moderator to
redact these events. This is a huge problem that occurs far more
frequently than it should[^often-soft-failure].

[^create-mass-redaction]: I think?

[^often-soft-failure]: https://github.com/matrix-org/synapse/issues/9329
## Proposal

This proposal introduces a very simple client-server endpoint to
alleviate all three of the concerns surrounding the moderators
redaction flow.

### Redacting a user's events
Comment thread
Gnuxie marked this conversation as resolved.

A new endpoint is introduced into the client-server spec.
The endpoint is introduced with the intention to allow convenient and
quick bulk redaction of spam events for room moderators. Who may not
necessarily have administrative privileges on their homeserver, and
will also need to redact messages from remote users.

`POST /_matrix/client/v1/rooms/{roomID}/redact/user/{userID}`
Comment thread
Gnuxie marked this conversation as resolved.

This endpoint redacts the target matrix user's unredacted events by
sending redactions on behalf of the requesting user.

The target user's unredacted events are sourced in reverse
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is ordering a MUST in this MSC, or just a SHOULD? Assuming the intention is generally to redact all of a user's events (looking at the client implementation), I feel that this is more suitably an optional thing, however clarification would be nice.

chronological order, starting from the most chronologically-recent
visible event in the room history for the requesting user.
Similarly to `/rooms/{roomID}/messages?dir=b`.

This endpoint is blind to the distinction between normal and
soft-failed events, and will cause redactions to be issued
for any soft-failed events that match the scope of the
request.

Events known to have already been redacted SHOULD be ignored and MUST
NOT contribute to the request limit.

It is left to an implementation detail whether a `m.room.redaction`
event is created for each event, or
[MSC2244](https://github.com/matrix-org/matrix-spec-proposals/pull/2244)
is employed. We expect that for now, most implementations will
issue one `m.room.redaction` event for each event under
the scope of the request. Sent by the requesting user.

#### Rate limiting

This endpoint SHOULD be rate limited, but the number `m.room.redaction`
events emitted SHOULD NOT be the subject of the rate limit.
But instead the number of events that have been redacted overall.
This is to cover semantics of MSC2244.

Servers SHOULD be more liberal in the number of events that can be
redacted in comparison to rate limits for `/send` when the endpoint is
being used to redact another user's events. Rather than the
requester's own events. The intent of being liberal is to allow
moderators to remove spam quickly, rather than struggling behind rate
limits for `/rooms/redact/event` or `/send`.

#### Path parameters

`roomID`: `string` - The room to get events to redact from.

`userID`: `string` - The sender of the events to redact.

#### Query parameters

`limit`: `integer` - The maximum number of events to
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This MSC seems to heavily assume that this endpoint will only be called after a user has been banned, however the pagination defined here doesn't really account for a scenario where that isn't the case. How should pagination work if a user is not banned and continues to send events after the endpoint has already been called? Does it reset position to the latest event from them, working back from there (presumably skipping over events that already have been redacted)? Without a pagination token that's the only way I can see this working, but then, that makes this endpoint kinda ineffective if you're trying to redact someone who hasn't been removed from the room, since they can just send another event to interrupt the redaction progress.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also, bit of a nitpick, but there's no ceiling suggested here - is it reasonable for a client to assume that a server will usually cap out at 100 events, like /messages typically does?

redact. Default: 25.

Client authors should be aware that the server may return less than
the `limit` even when `is_more_events` is `true`, and so should always
check the response.

#### Request body

`reason`: `string` - The reason for redacting the user's events
(Optional).

#### Response
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should there be a communication of which Events failed to redact? For example due to a db failure. Similar to how s-s api handles failed events where you get a list of Event ids back. Or should the server retry forever to redact the event?

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think it should ever fail to redact individual events 🤔

If the database explodes halfway through, that's the server's problem, it can throw an error on the entire request (but really the server should tell the database to not explode)


`is_more_events`: `boolean` - Whether there are more events outside of
the scope of the request sent by the user that could be redacted, but
have not been because of the `limit` query parameter. If
`is_more_events` is true, then the server should expect that the
client can optionally call the endpoint again, to redact even more events.

`redacted_events`: `object` - An object with the following fields:

* `total`: `integer` - The number of events that have been redacted,
including soft failed events.

* `soft_failed`: `integer` - The number of soft failed events that
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is there a reason for telling the requester how many soft failed events were redacted? especially since they're already included in total?

have been redacted.

```
200 OK
Content-Type: application/json

{
"is_more_events": false,
"redacted_events": {
"total": 5,
"soft_failed": 1
}
}
```

#### Limited response

Servers may wish to limit the maximum size of the limit that can be
specified. This maximum limit may even be dynamic to comply with rate
limiting.

To do this, they should simply redact less events than the limit provided
by the client.

```
POST /.../?limit=1000
200 OK
Content-Type: application/json

{
"is_more_events": true,
"redacted_events": {
"total": 25,
"soft_failed": 3
}
}
```

## Potential issues

### Forward compatibility with `/messages` style query parameters

We probably want to keep the endpoint compatible with `/messages`
style query pamarameters if the endpoint needs to be extended in the
future. It is our current assessment that we are currently forward
compatible. But someone should double check.

### Redacting "future" soft-failed events

Given that a request to the `/rooms/redact/user` endpoint is very
likely to occur after a room ban, then it makes sense that there could
still be soft failed events outside the scope of the request. The
moderator's homeserver is likely to discover soft-failed events from
the target user after the moderator's request has completed.

It could make sense to add a flag to the endpoint to tell the
moderator's homeserver to issue redactions on their behalf for the
newly discovered events. However, this could be complicated for
servers to implement. The flag would have to reset if the target user
is unbanned, and all incoming soft failed events will have to be
checked against a list of flagged servers.

At the moment we will remain forward compatible with a future
proposal to add a flag.
Comment on lines +176 to +192
Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Server implementers: Am I right in making a judgement call that this would be too complicated for the proposal?

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Meowlnir just has a hacky feature where if the last redacted event is <5 minutes old, it'll sleep for a bit and refetch events to redact. The assumption is that the most common soft fail issue is where events race with the ban, which should be covered by waiting a while for the ban to propagate.

A slightly less hacky method might be some way to opt into receiving soft-failed events, but that's probably a different MSC

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, Draupnir could in theory just call the endpoint again after a few minutes. I'd still like to just be able to give a flag to this endpoint, even if servers will only honor it for the next 10minutes or something.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@jevolk suggested that there should be a future proposal to federate the mass-redact action against users to ensure a cleanup of spam that is soft-failed locally but visible on remote servers.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I should probably also add the context of why I wanted this consideration, which is that: it's a really common issue for room mods to get complaints that spam hasn't been redacted from remote users, but the mods can't see it

Copy link
Copy Markdown
Contributor

@timedoutuk timedoutuk Jun 28, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking at implementing this into continuwuity, my ideal way to handle this would be either the client re-calling later, or a MSC4284 policy server would deal with mopping up late-arriving events. Tracking for a limited time in the server would probably come with its own suite of reliability and consistency issues, so having it handled elsewhere is probably both easier and more reliable. Just my two cents.

For context: the continuwuity main room has been dealing with a horrific volume of spam during waves, resulting in the aforementioned inconsistent cleanup. Even with Meowlnir doing the re-checking a few minutes later, we found that events were still arriving days later, primarily via backfill from unsuspecting servers with prev_events and whatnot. We set up a policy server to deal with this, and for a week or two it was still issuing maybe a few redactions per new incoming message, and it has only recently calmed down to maybe one redaction every dozen messages instead. I don't think the server redacting incoming future events, nor re-calling on a timer is going to have the desired effect, aside from in some less aggregious cases.



## Alternatives

* An endpoint could be developed to expose soft-failed events to
clients or appservices such as in
[MSC4109](https://github.com/matrix-org/matrix-spec-proposals/pull/4109).

## Security considerations

### Use case for self redaction

Implementers should be cautious over the use of this API for self
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How about limiting self redaction to non dag critical Events? So messages and custom state events but not for example powerlevel Events or similar.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

DAG critical events are supposed to be redaction exempt already for critical fields i thought.

redaction. We might omit that use case from the MSC if there are concerns.

## Unstable prefix

Until the MSC is accepted, implementations can use `org.matrix.msc4194` as the
unstable prefix and as a flag in the `unstable_features` section of `/versions`:
* `/_matrix/client/unstable/org.matrix.msc4194/rooms/{roomID}/redact/user/{userID}`
instead of `/_matrix/client/v1/rooms/{roomID}/redact/user/{userID}`
Once the MSC is merged, the `org.matrix.msc4194.stable` unstable feature flag
can be used to advertise support for the stable version of the endpoint, until
the spec release with the endpoint is advertised.

## Dependencies

None