Optimize how we calculate `likely_domains` during backfill because I've seen this take 17s in production just to `get_current_state` which is used to `get_domains_from_state` (see case [*2. Loading tons of events* in the `/messages` investigation issue](https://github.com/matrix-org/synapse/issues/13356)).
There are 3 ways we currently calculate hosts that are in the room:
1. `get_current_state` -> `get_domains_from_state`
- Used in `backfill` to calculate `likely_domains` and `/timestamp_to_event` because it was cargo-culted from `backfill`
- This one is being eliminated in favor of `get_current_hosts_in_room` in this PR 🕳
1. `get_current_hosts_in_room`
- Used for other federation things like sending read receipts and typing indicators
1. `get_hosts_in_room_at_events`
- Used when pushing out events over federation to other servers in the `_process_event_queue_loop`
Fix https://github.com/matrix-org/synapse/issues/13626
Part of https://github.com/matrix-org/synapse/issues/13356
Mentioned in [internal doc](https://docs.google.com/document/d/1lvUoVfYUiy6UaHB6Rb4HicjaJAU40-APue9Q4vzuW3c/edit#bookmark=id.2tvwz3yhcafh)
### Query performance
#### Before
The query from `get_current_state` sucks just because we have to get all 80k events. And we see almost the exact same performance locally trying to get all of these events (16s vs 17s):
```
synapse=# SELECT type, state_key, event_id FROM current_state_events WHERE room_id = '!OGEhHVWSdvArJzumhm:matrix.org';
Time: 16035.612 ms (00:16.036)
synapse=# SELECT type, state_key, event_id FROM current_state_events WHERE room_id = '!OGEhHVWSdvArJzumhm:matrix.org';
Time: 4243.237 ms (00:04.243)
```
But what about `get_current_hosts_in_room`: When there is 8M rows in the `current_state_events` table, the previous query in `get_current_hosts_in_room` took 13s from complete freshness (when the events were first added). But takes 930ms after a Postgres restart or 390ms if running back to back to back.
```sh
$ psql synapse
synapse=# \timing on
synapse=# SELECT COUNT(DISTINCT substring(state_key FROM '@[^:]*:(.*)$'))
FROM current_state_events
WHERE
type = 'm.room.member'
AND membership = 'join'
AND room_id = '!OGEhHVWSdvArJzumhm:matrix.org';
count
-------
4130
(1 row)
Time: 13181.598 ms (00:13.182)
synapse=# SELECT COUNT(*) from current_state_events where room_id = '!OGEhHVWSdvArJzumhm:matrix.org';
count
-------
80814
synapse=# SELECT COUNT(*) from current_state_events;
count
---------
8162847
synapse=# SELECT pg_size_pretty( pg_total_relation_size('current_state_events') );
pg_size_pretty
----------------
4702 MB
```
#### After
I'm not sure how long it takes from complete freshness as I only really get that opportunity once (maybe restarting computer but that's cumbersome) and it's not really relevant to normal operating times. Maybe you get closer to the fresh times the more access variability there is so that Postgres caches aren't as exact. Update: The longest I've seen this run for is 6.4s and 4.5s after a computer restart.
After a Postgres restart, it takes 330ms and running back to back takes 260ms.
```sh
$ psql synapse
synapse=# \timing on
Timing is on.
synapse=# SELECT
substring(c.state_key FROM '@[^:]*:(.*)$') as host
FROM current_state_events c
/* Get the depth of the event from the events table */
INNER JOIN events AS e USING (event_id)
WHERE
c.type = 'm.room.member'
AND c.membership = 'join'
AND c.room_id = '!OGEhHVWSdvArJzumhm:matrix.org'
GROUP BY host
ORDER BY min(e.depth) ASC;
Time: 333.800 ms
```
#### Going further
To improve things further we could add a `limit` parameter to `get_current_hosts_in_room`. Realistically, we don't need 4k domains to choose from because there is no way we're going to query that many before we a) probably get an answer or b) we give up.
Another thing we can do is optimize the query to use a index skip scan:
- https://wiki.postgresql.org/wiki/Loose_indexscan
- Index Skip Scan, https://commitfest.postgresql.org/37/1741/
- https://www.timescale.com/blog/how-we-made-distinct-queries-up-to-8000x-faster-on-postgresql/
Part of #13019
This changes all the permission-related methods to rely on the Requester instead of the UserID. This is a first step towards enabling scoped access tokens at some point, since I expect the Requester to have scope-related informations in it.
It also changes methods which figure out the user/device/appservice out of the access token to return a Requester instead of something else. This avoids having store-related objects in the methods signatures.
This improves load times for push rules:
| Version | Time per user | Time for 1k users |
| -------------------- | ------------- | ----------------- |
| Before | 138 µs | 138ms |
| Now (with custom) | 2.11 µs | 2.11ms |
| Now (without custom) | 49.7 ns | 0.05 ms |
This therefore has a large impact on send times for rooms
with large numbers of local users in the room.
* Clarifies comments.
* Fixes an erroneous comment (about return type) added in #13455
(ec24813220).
* Clarifies the name of a variable.
* Simplifies logic of pulling out the latest join for the requesting user.
This adds support for the stable identifiers of MSC2285 while
continuing to support the unstable identifiers behind the configuration
flag. These will be removed in a future version.
* Adds docstrings and inline comments.
* Formats SQL queries using triple quoted strings.
* Minor formatting changes.
* Avoid fetching `event_push_summary_stream_ordering` multiple times
in the same transactions.
Still maintains local in memory lookup optimisation, but does any external
lookup as part of the deferred that prevents duplicate lookups for the same
event at once. This makes the assumption that fetching from an external
cache is a non-zero load operation.
See #10826 and #10786 for context as to why we had to disable pruning on
those caches.
Now that `get_users_who_share_room_with_user` is called frequently only
for presence, we just need to make calls to it less frequent and then we
can remove the various levels of caching that is going on.
Fix race conditions in the async cache invalidation logic, by separating
the async & local invalidation calls and ensuring any async call i
executed first.
Signed off by Nick @ Beeper (@Fizzadar).
More prep work for asyncronous caching, also makes all process_replication_rows methods consistent (presence handler already is so).
Signed off by Nick @ Beeper (@Fizzadar)
These columns were added back in Synapse 1.52, and have been populated for new
events since then. It's now (beyond) time to back-populate them for existing
events.
Some experimental prep work to enable external event caching based on #9379 & #12955. Doesn't actually move the cache at all, just lays the groundwork for async implemented caches.
Signed off by Nick @ Beeper (@Fizzadar)
* Replace `get_new_events_for_appservice` with `get_all_new_events_stream`
The functions were near identical and this brings the AS worker closer
to the way federation senders work which can allow for multiple workers
to handle AS traffic.
* Pull received TS alongside events when processing the stream
This avoids an extra query -per event- when both federation sender
and appservice pusher process events.
Bounce recalculation of current state to the correct event persister and
move recalculation of current state into the event persistence queue, to
avoid concurrent updates to a room's current state.
Also give recalculation of a room's current state a real stream
ordering.
Signed-off-by: Sean Quah <seanq@matrix.org>
This happened if we encountered a stream ordering in `event_push_actions` that had more rows than the batch size of the delete, as If we don't delete any rows in an iteration then the next time round we get the exact same stream ordering and get stuck.
Whenever we want to persist an event, we first compute an event context,
which includes the state at the event and a flag indicating whether the
state is partial. After a lot of processing, we finally try to store the
event in the database, which can fail for partial state events when the
containing room has been un-partial stated in the meantime.
We detect the race as a foreign key constraint failure in the data store
layer and turn it into a special `PartialStateConflictError` exception,
which makes its way up to the method in which we computed the event
context.
To make things difficult, the exception needs to cross a replication
request: `/fed_send_events` for events coming over federation and
`/send_event` for events from clients. We transport the
`PartialStateConflictError` as a `409 Conflict` over replication and
turn `409`s back into `PartialStateConflictError`s on the worker making
the request.
All client events go through
`EventCreationHandler.handle_new_client_event`, which is called in
*a lot* of places. Instead of trying to update all the code which
creates client events, we turn the `PartialStateConflictError` into a
`429 Too Many Requests` in
`EventCreationHandler.handle_new_client_event` and hope that clients
take it as a hint to retry their request.
On the federation event side, there are 7 places which compute event
contexts. 4 of them use outlier event contexts:
`FederationEventHandler._auth_and_persist_outliers_inner`,
`FederationHandler.do_knock`, `FederationHandler.on_invite_request` and
`FederationHandler.do_remotely_reject_invite`. These events won't have
the partial state flag, so we do not need to do anything for then.
The remaining 3 paths which create events are
`FederationEventHandler.process_remote_join`,
`FederationEventHandler.on_send_membership_event` and
`FederationEventHandler._process_received_pdu`.
We can't experience the race in `process_remote_join`, unless we're
handling an additional join into a partial state room, which currently
blocks, so we make no attempt to handle it correctly.
`on_send_membership_event` is only called by
`FederationServer._on_send_membership_event`, so we catch the
`PartialStateConflictError` there and retry just once.
`_process_received_pdu` is called by `on_receive_pdu` for incoming
events and `_process_pulled_event` for backfill. The latter should never
try to persist partial state events, so we ignore it. We catch the
`PartialStateConflictError` in `on_receive_pdu` and retry just once.
Refering to the graph of code paths in
https://github.com/matrix-org/synapse/issues/12988#issuecomment-1156857648
may make the above make more sense.
Signed-off-by: Sean Quah <seanq@matrix.org>
