MSC3030: Jump to date API endpoint

proposals/3030-jump-to-date.md

@@ -0,0 +1,286 @@
+# MSC3030: Jump to date API endpoint
+
+Add an API that makes it easy to find the closest messages for a given
+timestamp.
+
+The goal of this change is to have clients be able to implement a jump to date
+feature in order to see messages back at a given point in time. Pick a date from
+a calender, heatmap, or paginate next/previous between days and view all of the
+messages that were sent on that date.
+
+Alongside the [roadmap of feature parity with
+Gitter](https://github.com/vector-im/roadmap/issues/26), we're also interested
+in using this for a new better static Matrix archive. Our idea is to server-side
+render [Hydrogen](https://github.com/vector-im/hydrogen-web) and this new
+endpoint would allow us to jump back on the fly without having to paginate and
+keep track of everything in order to display the selected date.
+
+Also useful for archiving and backup use cases. This new endpoint can be used to
+slice the messages by day and persist to file.
+
+Related issue: [*URL for an arbitrary day of history and navigation for next and
+previous days*
+(vector-im/element-web#7677)](https://github.com/vector-im/element-web/issues/7677)
+
+
+## Problem
+
+These types of use cases are not supported by the current Matrix API because it
+has no way to fetch or filter older messages besides a manual brute force
+pagination from the most recent event in the room. Paginating is time-consuming
+and expensive to process every event as you go (not practical for clients).
+Imagine wanting to get a message from 3 years ago 😫
+
+
+## Proposal
+
+Add new client API endpoint `GET
+/_matrix/client/v1/rooms/{roomId}/timestamp_to_event?ts=<timestamp>&dir=[f|b]`
+which fetches the closest `event_id` to the given timestamp `ts` query parameter
+in the direction specified by the `dir` query parameter. The direction `dir`
+query parameter accepts `f` for forward-in-time from the timestamp and `b` for
+backward-in-time from the timestamp. This endpoint also returns
+`origin_server_ts` to make it easy to do a quick comparison to see if the
+`event_id` fetched is too far out of range to be useful for your use case.
+
+When an event can't be found in the given direction, the endpoint throws a 404
+`"errcode":"M_NOT_FOUND",` (example error message `"error":"Unable to find event
+from 1672531200000 in direction f"`).
+
+In order to solve the problem where a homeserver does not have all of the history in a
+room and no suitably close event, we also add a server API endpoint `GET
+/_matrix/federation/v1/timestamp_to_event/{roomId}?ts=<timestamp>?dir=[f|b]` which other
+homeservers can use to ask about their closest `event_id` to the timestamp. This
+endpoint also returns `origin_server_ts` to make it easy to do a quick comparison to see
+if the remote `event_id` fetched is closer than the local one. After the local
+homeserver receives a response from the federation endpoint, it probably should
+try to backfill this event via the federation `/event/<event_id>` endpoint so that it's
+available to query with `/context` from a client in order to get a pagination token.
+
+The heuristics for deciding when to ask another homeserver for a closer event if
+your homeserver doesn't have something close, are left up to the homeserver
+implementation, although the heuristics will probably be based on whether the
+closest event is a forward/backward extremity indicating it's next to a gap of
+events which are potentially closer.
+
+A good heuristic for which servers to try first is to sort by servers that have
+been in the room the longest because they're most likely to have anything we ask
+about.
+
+These endpoints are authenticated and should be rate-limited like similar client
+and federation endpoints to prevent resource exhaustion abuse.
+
+```
+GET /_matrix/client/v1/rooms/<roomID>/timestamp_to_event?ts=<timestamp>&dir=<direction>
+{
+    "event_id": ...
+    "origin_server_ts": ...
+}
+```
+
+Federation API endpoint:
+```
+GET /_matrix/federation/v1/timestamp_to_event/<roomID>?ts=<timestamp>&dir=<direction>
+{
+    "event_id": ...
+    "origin_server_ts": ...
+}
+```
+
+---
+
+In order to paginate `/messages`, we need a pagination token which we can get
+using `GET /_matrix/client/r0/rooms/{roomId}/context/{eventId}?limit=0` for the
+`event_id` returned by `/timestamp_to_event`.
+
+We can always iterate on `/timestamp_to_event` later and return a pagination
+token directly in another MSC ⏩
+
+
+## Potential issues
+
+### Receiving a rogue random delayed event ID
+
+Since `origin_server_ts` is not enforcably accurate, we can only hope that an event's
+`origin_server_ts` is relevant enough to its `prev_events` and descendants.
+
+If you ask for "the message with `origin_server_ts` closest to Jan 1st 2018" you
+might actually get a rogue random delayed one that was backfilled from a
+federated server, but the human can figure that out by trying again with a
+slight variation on the date or something.
+
+Since there isn't a good or fool-proof way to combat this, it's probably best to just go
+with `origin_server_ts` and not let perfect be the enemy of good.
+
+
+### Receiving an unrenderable event ID
+
+Another issue is that clients could land on an event they can't/won't render,
+such as a reaction, then they'll be forced to desperately seek around the
+timeline until they find an event they can do something with.
+
+Eg:
+ - Client wants to jump to January 1st, 2022
+ - Server says there's an event on January 2nd, 2022 that is close enough
+ - Client finds out there's a ton of unrenderable events like memberships, poll responses, reactions, etc at that time
+ - Client starts paginating forwards, finally finding an event on January 27th it can render
+ - Client wasn't aware that the actual nearest neighbouring event was backwards on December 28th, 2021 because it didn't paginate in that direction
+ - User is confused that they are a month past the target date when the message is *right there*.
+
+Clients can be smarter here though. Clients can see when events were sent as
+they paginate and if they see they're going more than a couple days out, they
+can also try the other direction before going further and further away.
+
+Clients can also just explain to the user what happened with a little toast: "We
+were unable to find an event to display on January 1st, 2022. The closest event
+after that date is on January 27th."
+
+
+### Abusing the `/timestamp_to_event` API to get the `m.room.create` event 
+
+Although it's possible to jump to the start of the room and get the first event in the
+room (`m.room.create`) with `/timestamp_to_event?dir=f&ts=0`, clients should still use
+`GET /_matrix/client/v3/rooms/{roomId}/state/m.room.create/` to get the room creation
+event.
+
+In the future, with things like importing history via
+[MSC2716](https://github.com/matrix-org/matrix-spec-proposals/pull/2716), the first
+event you encounter with `/timestamp_to_event?dir=f&ts=0` could be an imported event before
+the room was created.
+
+
+## Alternatives
+
+We chose the current `/timestamp_to_event` route because it sounded like the
+easist path forward to bring it to fruition and get some real-world experience.
+And was on our mind during the [initial discussion](https://docs.google.com/document/d/1KCEmpnGr4J-I8EeaVQ8QJZKBDu53ViI7V62y5BzfXr0/edit#bookmark=id.qu9k9wje9pxm) because there was some prior art with a [WIP
+implementation](https://github.com/matrix-org/synapse/pull/9445/commits/91b1b3606c9fb9eede0a6963bc42dfb70635449f)
+from @erikjohnston. The alternatives haven't been thrown out for a particular
+reason and we could still go down those routes depending on how people like the
+current design.
+
+
+### Paginate `/messages?around=<timestamp>` from timestamp
+
+Add the `?around=<timestamp>` query parameter to the `GET
+/_matrix/client/r0/rooms/{roomId}/messages` endpoint. This will start the
+response at the message with `origin_server_ts` closest to the provided `around`
+timestamp. The direction is determined by the existing `?dir` query parameter.
+
+Use topological ordering, just as Element would use if you follow a permalink.
+
+This alternative could be confusing to the end-user around how this plays with
+the existing query parameters
+`/messages?from={paginationToken}&to={paginationToken}` which also determine
+what part of the timeline to query. Those parameters could be extended to accept
+timestamps in addition to pagination tokens but then could get confusing again
+when you start mixing timestamps and pagination tokens. The homeserver also has
+to disambiguate what a pagination token looks like vs a unix timestamp. Since
+pagination tokens don't follow a certain convention, some homeserver
+implementations may already be using arbitrary number tokens already which would
+be impossible to distinguish from  a timestamp.
+
+A related alternative is to use `/messages` with a `from_time`/`to_time` (or
+`from_ts`/`to_ts`) query parameters that only accept timestamps which solves the
+confusion and disambigution problem of trying to re-use the existing `from`/`to`
+query paramters. Re-using `/messages` would reduce the number of round-trips and
+potentially client-side implementations for the use case where you want to fetch
+a window of messages from a given time. But has the same round-trip problem if
+you want to use the returned `event_id` with `/context` or another endpoint
+instead.
+
+
+### Filter by date in `RoomEventFilter`
+
+Extend `RoomEventFilter` to be able to specify a timestamp or a date range. The
+`RoomEventFilter` can be passed via the `?filter` query param on the `/messages`
+endpoint.
+
+This suffers from the same confusion to the end-user of how it plays with how
+this plays with `/messages?from={paginationToken}&to={paginationToken}` which
+also determines what part of the timeline to query.
+
+
+### Return the closest event in any direction
+
+We considered omitting the `dir` parameter (or allowing `dir=c`) to have the server
+return the closest event to the timestamp, regardless of direction. However, this seems
+to offer little benefit.
+
+Firstly, for some usecases (such as archive viewing, where we want to show all the
+messages that happened on a particular day), an explicit direction is important, so this
+would have to be optional behaviour.
+
+For a regular messaging client, "directionless" search also offers little benefit: it is
+easy for the client to repeat the request in the other direction if the returned event
+is "too far away", and in any case it needs to manage an iterative search to handle
+unrenderable events, as discussed above.
+
+Implementing a directionless search on the server carries a performance overhead, since
+it must search both forwards and backwards on every request. In short, there is little
+reason to expect that a single `dir=c` request would be any more efficient than a pair of
+requests with `dir=b` and `dir=f`.
+
+### New `destination_server_ts` field
+
+Add a new field and index on messages called `destination_server_ts` which
+indicates when the message was received from federation. This gives a more
+"real" time for how someone would actually consume those messages.
+
+The contract of the API is "show me messages my server received at time T"
+rather than the messy confusion of showing a delayed message which happened to
+originally be sent at time T.
+
+We've decided against this approach because the backfill from federated servers
+could be horribly late.
+
+---
+
+Related issue around `/sync` vs `/messages`,
+https://github.com/matrix-org/synapse/issues/7164
+
+> Sync returns things in the order they arrive at the server; backfill returns
+> them in the order determined by the event graph.
+>
+> *-- @richvdh, https://github.com/matrix-org/synapse/issues/7164#issuecomment-605877176*
+
+> The general idea is that, if you're following a room in real-time (ie,
+> `/sync`), you probably want to see the messages as they arrive at your server,
+> rather than skipping any that arrived late; whereas if you're looking at a
+> historical section of timeline (ie, `/messages`), you want to see the best
+> representation of the state of the room as others were seeing it at the time.
+>
+> *-- @richvdh , https://github.com/matrix-org/synapse/issues/7164#issuecomment-605953296*
+
+
+## Security considerations
+
+We're only going to expose messages according to the existing message history
+setting in the room (`m.room.history_visibility`). No extra data is exposed,
+just a new way to sort through it all.
+
+
+
+## Unstable prefix
+
+While this MSC is not considered stable, the endpoints are available at `/unstable/org.matrix.msc3030` instead of their `/v1` description from above.
+
+```
+GET /_matrix/client/unstable/org.matrix.msc3030/rooms/<roomID>/timestamp_to_event?ts=<timestamp>&dir=<direction>
+{
+    "event_id": ...
+    "origin_server_ts": ...
+}
+```
+
+```
+GET /_matrix/federation/unstable/org.matrix.msc3030/timestamp_to_event/<roomID>?ts=<timestamp>&dir=<direction>
+{
+    "event_id": ...
+    "origin_server_ts": ...
+}
+```
+
+Servers will indicate support for the new endpoint via a non-empty value for feature flag
+`org.matrix.msc3030` in `unstable_features` in the response to `GET
+/_matrix/client/versions`.