a763cbb0e1
* Put federation client functions into their own file * Look for missing auth events in RS input * Remove retrieveMissingAuthEvents from federation API * Logging * Sorta transplanted the code over * Use event origin failing all else * Don't get stuck on mutexes: * Add verifier * Don't mark state events with zero snapshot NID as not existing * Check missing state if not an outlier before storing the event * Reject instead of soft-fail, don't copy roominfo so much * Use synchronous contexts, limit time to fetch missing events * Clean up some commented out bits * Simplify `/send` endpoint significantly * Submit async * Report errors on sending to RS input * Set max payload in NATS to 16MB * Tweak metrics * Add `workerForRoom` for tidiness * Try skipping unmarshalling errors for RespMissingEvents * Track missing prev events separately to avoid calculating state when not possible * Tweak logic around checking missing state * Care about state when checking missing prev events * Don't check missing state for create events * Try that again * Handle create events better * Send create room events as new * Use given event kind when sending auth/state events * Revert "Use given event kind when sending auth/state events" This reverts commit 089d64d271b5fca8c104e1554711187420dbebca. * Only search for missing prev events or state for new events * Tweaks * We only have missing prev if we don't supply state * Room version tweaks * Allow async inputs again * Apply backpressure to consumers/synchronous requests to hopefully stop things being overwhelmed * Set timeouts on roomserver input tasks (need to decide what timeout makes sense) * Use work queue policy, deliver all on restart * Reduce chance of duplicates being sent by NATS * Limit the number of servers we attempt to reduce backpressure * Some review comment fixes * Tidy up a couple things * Don't limit servers, randomise order using map * Some context refactoring * Update gmsl * Don't resend create events * Set stateIDs length correctly or else the roomserver thinks there are missing events when there aren't * Exclude our own servername * Try backing off servers * Make excluding self behaviour optional * Exclude self from g_m_e * Update sytest-whitelist * Update consumers for the roomserver output stream * Remember to send outliers for state returned from /gme * Make full HTTP tests less upsetti * Remove 'If a device list update goes missing, the server resyncs on the next one' from the sytest blacklist * Remove debugging test * Fix blacklist again, remove unnecessary duplicate context * Clearer contexts, don't use background in case there's something happening there * Don't queue up events more than once in memory * Correctly identify create events when checking for state * Fill in gaps again in /gme code * Remove `AuthEventIDs` from `InputRoomEvent` * Remove stray field Co-authored-by: Kegan Dougal <kegan@matrix.org> |
||
---|---|---|
.. | ||
consumers | ||
internal | ||
notifier | ||
routing | ||
storage | ||
streams | ||
sync | ||
types | ||
README.md | ||
syncapi.go |
Sync API Server
This server is responsible for servicing /sync
requests. It gets its data from the room server output log. Currently, the sync server will:
- Return a valid
/sync
response for the user represented by the providedaccess_token
. - Return a "complete sync" if no
since
value is provided, and return a validnext_batch
token. This contains all rooms the user has been invited to or has joined. For joined rooms, this includes the complete current room state and the most recent 20 (hard-coded) events in the timeline. - For "incremental syncs" (a
since
value is provided), as you get invited to, join, or leave rooms they will be reflected correctly in the/sync
response. - For very large state deltas, the
state
section of a room is correctly populated with the state of the room at the start of the timeline. - When you join a room, the
/sync
which transitions your client to be "joined" will include the complete current room state as per the specification. - Only wake up user streams it needs to wake up.
- Honours the
timeout
query parameter value.
Internals
When the server gets a /sync
request, it needs to:
- Work out which rooms to return to the client.
- For each room, work out which events to return to the client.
The logic for working out which rooms is based on Synapse:
- Get the CURRENT joined room list for this user.
- Get membership list changes for this user between the provided stream position and now.
- For each room which has membership list changes:
- Check if the room is 'newly joined' (insufficient to just check for a join event because we allow dupe joins). If it is, then we need to send the full room state down (and 'limited' is always true).
- Check if user is still CURRENTLY invited to the room. If so, add room to 'invited' block.
- Check if the user is CURRENTLY left/banned. If so, add room to 'archived' block.
- Add joined rooms (joined room list)
For each room, the /sync response returns the most recent timeline events and the state of the room at the start of the timeline.
The logic for working out which events is not based entirely on Synapse code, as it is known broken with respect to working out
room state. In order to know which events to return, the server needs to calculate room state at various points in the history of
the room. For example, imagine a room with the following 15 events (letters are state events (updated via '
), numbers are timeline events):
index 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 (1-based indexing as StreamPosition(0) represents no event)
timeline [A, B, C, D, 1, 2, 3, D', 4, D'', 5, B', D''', D'''', 6]
The current state of this room is: [A, B', C, D'''']
.
If this room was requested with ?since=14&limit=5
then 1 timeline event would be returned, the most recent one:
15
[ 6 ]
If this room was requested with ?since=9&limit=5
then 5 timeline events would be returned, the most recent ones:
11 12 13 14 15
[5, B', D''', D'''', 6]
The state of the room at the START of the timeline can be represented in 2 ways:
- The
full_state
from index 0 :[A, B, C, D'']
(aka the state between 0-11 exclusive) - A partial state from index 9 :
[D'']
(aka the state between 9-11 exclusive)
Servers advance state events (e.g from D'
to D''
) based on the state conflict resolution algorithm.
You might think that you could advance the current state by just updating the entry for the (event type, state_key)
tuple
for each state event, but this state can diverge from the state calculated using the state conflict resolution algorithm.
For example, if there are two "simultaneous" updates to the same state key, that is two updates at the same depth in the
event graph, then the final result of the state conflict resolution algorithm might not match the order the events appear
in the timeline.
The correct advancement for state events is represented by the AddsStateEventIDs
and RemovesStateEventIDs
that
are in OutputRoomEvents
from the room server.
This version of the sync server uses very simple indexing to calculate room state at various points.
This is inefficient when a very old since
value is provided, or the full_state
is requested, as the state delta becomes
very large. This is mitigated slightly with indexes, but better data structures could be used in the future.
Known Issues
m.room.history_visibility
is not honoured: it is always treated as "shared".- All ephemeral events are not implemented (presence, typing, receipts).
- Account data (both user and room) is not implemented.
to_device
messages are not implemented.- Back-pagination via
prev_batch
is not implemented. - The
limited
flag can lie. - Filters are not honoured or implemented. The
limit
for each room is hard-coded to 20. - The
full_state
query parameter is not implemented. - The
set_presence
query parameter is not implemented. - "Ignored" users are not ignored.
- Redacted events are still sent to clients.
- Invites over federation (if it existed) won't work as they aren't "real" events and so won't be in the right tables.
invite_state
is not implemented (for similar reasons to the above point).- The current implementation scales badly when a very old
since
token is provided. - The entire current room state can be re-sent to the client if they send a duplicate "join" event which should be a no-op.