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
150 changes: 150 additions & 0 deletions proposals/4492-directionless-pagination-tokens.md

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.

Implementation not required: the MSC is effectively already implemented.

Original file line number Diff line number Diff line change
@@ -0,0 +1,150 @@
# MSC4492: Clarify that pagination tokens are directionless

The Matrix client-server API has several endpoints which handle pagination over events:

* [`/messages`][messages]
* [`/sync`][sync]
* [`/relations`][relations]
* [`/context`][context]
* [`/search`][search]'s `context` field for each result

Each returns tokens that can be used for further pagination, though the specification is
vague around which tokens can be used with which endpoints.

There is a [spec PR](spec2357) which attempts to fix an inconsistency in the `/relations`
endpoint: the `from` parameter currently lists `next_batch` from `/sync` as an acceptable
value, but this was likely meant to be `prev_batch`. This has raised a broader question:
are pagination tokens directional, or do they represent an opaque position in the event
stream without inherent direction?

Synapse has always treated tokens as opaque positions — a [`StreamToken`][synapse-streamtoken]
can be used for both forward and backward pagination depending on the `dir` parameter.
(Internally it embeds a [`RoomStreamToken`][synapse-roomstreamtoken] with stream (`s`)
and topological (`t`) variants, but this is an implementation detail — the API-visible
type is the same everywhere.)

This MSC aims to codify that understanding: pagination tokens do not have an
inherent direction, even if they happen to be labelled `next_batch` or
`prev_batch`. The direction is always determined by the `dir` parameter.

## Proposal

Pagination tokens produced by `/messages`, `/relations`, `/context`, the `context`
response field of `/search`, and the `prev_batch` response field of `/sync` represent
a posiiton in the event stream and MUST be indistinguishable. These tokens are referred

Check warning on line 34 in proposals/4492-directionless-pagination-tokens.md

View workflow job for this annotation

GitHub Actions / Spell Check with Typos

"posiiton" should be "position".
to as "room event stream tokens" in the rest of the proposal text. (Although it is not
proposed to use that term in the spec itself.)

"Room event stream tokens" MUST only be used within the context of a single room.
(E.g. a request for `/messages` in room `#test:matrix.org` does not make sense to re-use
for `#foo:matrix.org`.) There are no guarantees using a "room event stream token" in a
different room will produce sensible results; servers MAY throw an error if a client attempts
to use one in the wrong room.

`next_batch` tokens produced by [`/sync`][sync] resolve against the server's
global event stream and are safe to use in any room for `/messages`, `/relations`,
and `/context`. The `since` parameter of `/sync` only accepts `next_batch` tokens produced
by `/sync`.

The following table summarises which tokens each endpoint currently accepts and proposed changes:

| Endpoint | Input | Currently accepts | Output | Change |
|---|---|---|---|---|
| `/sync` | `since` | Its own `next_batch` | `next_batch`, `prev_batch` | None |
| `/messages` | `from` | `/sync` (`prev_batch`/`next_batch`), itself (`end`), optionally itself (`start`) | `start`, `end` | `/sync`, `/messages`, `/relations`, `/context`, and `/search`'s `context` |
| `/messages` | `to` | `/sync` (`prev_batch`/`next_batch`), itself (`end`) | — | `/sync`, `/messages`, `/relations`, `/context`, and `/search`'s `context` |
| `/relations` | `from` | Itself, `/messages` (`start`), `/sync` (`next_batch`) | `next_batch`, `prev_batch` | `/sync`, `/messages`, `/relations`, `/context`, and `/search`'s `context` |
| `/relations` | `to` | Itself, `/messages`, `/sync` (vague) | — | `/sync`, `/messages`, `/relations`, `/context`, and `/search`'s `context` |
| `/context` | (none) | Output only | `start` ("backwards"), `end` ("forwards") | Remove directional language |
| `/search` result | (none) | Output only | `start`, `end` (no directional language) | None |

Note that this does not modify the pagination of `/search` results itself, which are
only usable within `/search`.

This does not affect the pagination of any other endpoints, such as `/threads` or
[`/hierarchy`](hierarchy).

## Potential issues

Homeserver implementations differ in how they structure pagination tokens. The rules
above may drastically break how a homeserver generates pagination tokens.

### Synapse

Synapse uses a single [`StreamToken`][synapse-streamtoken] type for all pagination
tokens across `/sync`, `/messages`, `/relations`, `/context`, and `/search`. It
does not distinguish direction or origin in the token itself. (Internally, the
`room_key` sub-field is a [`RoomStreamToken`][synapse-roomstreamtoken] which can
represent a stream or topological position, but this is an implementation detail
— the API-visible token envelope is the same everywhere.) No changes are required.

### Conduit lineage (conduwuit / continuwuity / Tuwunel)

These implementations use a [`PduCount`][conduit-pducount] type (a signed integer
with `Normal(u64)` / `Backfilled(i64)` variants) across `/messages`, `/relations`,
and `/context`. Tokens from these endpoints do not encode direction.

`/sync` parses its `since` parameter as a `u64` — it accepts **only**
non-negative integers. A `PduCount::Backfilled` (negative) token from `/relations`
or `/messages` would fail to parse in `/sync`. The reverse direction works: sync
tokens are always non-negative and parse cleanly as `PduCount::Normal`.

### Dendrite lineage (Dendrite / Zendrite)

These implementations use **three separate token types** where each type is tied to
a specific endpoint, encoding both direction and origin in the token format:

* [`StreamingToken`][dendrite-streaming] (`s{...}`) — used by `/sync`.
* [`TopologyToken`][dendrite-topology] (`t{...}`) — used by `/messages`, `/context`,
and `/search`'s `context`.
* [`StreamPosition`][dendrite-streampos] (bare integer) — used by `/relations`.

`/relations` parses its tokens via `strconv.Atoi`, which rejects both `s`- and
`t`-prefixed strings. Conversely, `/sync` expects only `StreamingToken` and
rejects bare integers and `t`-prefixed strings. Only `/messages` has dual-parsing
logic that tries `TopologyToken` first and falls back to `StreamingToken`,
converting server-side.

**Changes needed:** At minimum, `/relations` should accept any token format (not
just bare integers). A pattern like `/messages` already uses (try one format,
fall back to another, converting server-side) could be applied to the other endpoints.

## Alternatives

The main alternative is to leave the current inconsistencies as-is. The [spec PR](spec2357)
could land its minimal fix, though the review discussion showed it is unclear which
direction is correct without addressing the broader question of interchangeability.

Another alternative is to leave each endpoint describing its own accepted token
types independently, with no general principle. This is the status quo.

## Security considerations

None.

## Unstable prefix

N/A

## Dependencies

N/A

[messages]: https://spec.matrix.org/v1.18/client-server-api/#get_matrixclientv3roomsroomidmessages
[sync]: https://spec.matrix.org/v1.18/client-server-api/#get_matrixclientv3sync
[relations]: https://spec.matrix.org/v1.18/client-server-api/#get_matrixclientv1roomsroomidrelationseventid
[context]: https://spec.matrix.org/v1.18/client-server-api/#get_matrixclientv3roomsroomidcontexteventid
[search]: https://spec.matrix.org/v1.18/client-server-api/#post_matrixclientv3search
[threads]: https://spec.matrix.org/v1.18/client-server-api/#get_matrixclientv1roomsroomidthreads
[hierarchy]: https://spec.matrix.org/v1.18/client-server-api/#get_matrixclientv1roomsroomidhierarchy

[spec2357]: https://github.com/matrix-org/matrix-spec/pull/2357

[synapse-streamtoken]: https://github.com/element-hq/synapse/blob/76b4fdceed0739d83ac79588416dc88f25d8d14e/synapse/streams/config.py#L39-L45
[synapse-roomstreamtoken]: https://github.com/element-hq/synapse/blob/76b4fdceed0739d83ac79588416dc88f25d8d14e/synapse/types/__init__.py#L722

[conduit-pducount]: https://github.com/continuwuation/continuwuity/blob/main/src/core/matrix/pdu/count.rs#L13-L17

[dendrite-streaming]: https://github.com/element-hq/dendrite/blob/main/syncapi/types/types.go#L106-L118
[dendrite-topology]: https://github.com/element-hq/dendrite/blob/main/syncapi/types/types.go#L213-L218
[dendrite-streampos]: https://github.com/element-hq/dendrite/blob/main/syncapi/types/types.go#L41-L52
Loading