While https://github.com/matrix-org/synapse/pull/13635 stops us from doing the slow thing after we've already done it once, this PR stops us from doing one of the slow things in the first place.
Related to
- https://github.com/matrix-org/synapse/issues/13622
- https://github.com/matrix-org/synapse/pull/13635
- https://github.com/matrix-org/synapse/issues/13676
Part of https://github.com/matrix-org/synapse/issues/13356
Follow-up to https://github.com/matrix-org/synapse/pull/13815 which tracks event signature failures.
With this PR, we avoid the call to the costly `_get_state_ids_after_missing_prev_event` because the signature failure will count as an attempt before and we filter events based on the backoff before calling `_get_state_ids_after_missing_prev_event` now.
For example, this will save us 156s out of the 185s total that this `matrix.org` `/messages` request. If you want to see the full Jaeger trace of this, you can drag and drop this `trace.json` into your own Jaeger, https://gist.github.com/MadLittleMods/4b12d0d0afe88c2f65ffcc907306b761
To explain this exact scenario around `/messages` -> backfill, we call `/backfill` and first check the signatures of the 100 events. We see bad signature for `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` and `$zuOn2Rd2vsC7SUia3Hp3r6JSkSFKcc5j3QTTqW_0jDw` (both member events). Then we process the 98 events remaining that have valid signatures but one of the events references `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` as a `prev_event`. So we have to do the whole `_get_state_ids_after_missing_prev_event` rigmarole which pulls in those same events which fail again because the signatures are still invalid.
- `backfill`
- `outgoing-federation-request` `/backfill`
- `_check_sigs_and_hash_and_fetch`
- `_check_sigs_and_hash_and_fetch_one` for each event received over backfill
- ❗ `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` fails with `Signature on retrieved event was invalid.`: `unable to verify signature for sender domain xxx: 401: Failed to find any key to satisfy: _FetchKeyRequest(...)`
- ❗ `$zuOn2Rd2vsC7SUia3Hp3r6JSkSFKcc5j3QTTqW_0jDw` fails with `Signature on retrieved event was invalid.`: `unable to verify signature for sender domain xxx: 401: Failed to find any key to satisfy: _FetchKeyRequest(...)`
- `_process_pulled_events`
- `_process_pulled_event` for each validated event
- ❗ Event `$Q0iMdqtz3IJYfZQU2Xk2WjB5NDF8Gg8cFSYYyKQgKJ0` references `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` as a `prev_event` which is missing so we try to get it
- `_get_state_ids_after_missing_prev_event`
- `outgoing-federation-request` `/state_ids`
- ❗ `get_pdu` for `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` which fails the signature check again
- ❗ `get_pdu` for `$zuOn2Rd2vsC7SUia3Hp3r6JSkSFKcc5j3QTTqW_0jDw` which fails the signature check
Fixes#13942. Introduced in #13575.
Basically, let's only get the ordered set of hosts out of the DB if we need an ordered set of hosts. Since we split the function up the caching won't be as good, but I think it will still be fine as e.g. multiple backfill requests for the same room will hit the cache.
There is no need to grab thousands of backfill points when we only need 5 to make the `/backfill` request with. We need to grab a few extra in case the first few aren't visible in the history.
Previously, we grabbed thousands of backfill points from the database, then sorted and filtered them in the app. Fetching the 4.6k backfill points for `#matrix:matrix.org` from the database takes ~50ms - ~570ms so it's not like this saves a lot of time 🤷. But it might save us more time now that `get_backfill_points_in_room`/`get_insertion_event_backward_extremities_in_room` are more complicated after https://github.com/matrix-org/synapse/pull/13635
This PR moves the filtering and limiting to the SQL query so we just have less data to work with in the first place.
Part of https://github.com/matrix-org/synapse/issues/13356
c.f. #12993 (comment), point 3
This stores all device list updates that we receive while partial joins are ongoing, and processes them once we have the full state.
Note: We don't actually process the device lists in the same ways as if we weren't partially joined. Instead of updating the device list remote cache, we simply notify local users that a change in the remote user's devices has happened. I think this is safe as if the local user requests the keys for the remote user and we don't have them we'll simply fetch them as normal.
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/
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>
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>
Instead, use the `room_version` property of the event we're validating.
The `room_version` was originally added as a parameter somewhere around #4482,
but really it's been redundant since #6875 added a `room_version` field to `EventBase`.
Refactor how the `EventContext` class works, with the intention of reducing the amount of state we fetch from the DB during event processing.
The idea here is to get rid of the cached `current_state_ids` and `prev_state_ids` that live in the `EventContext`, and instead defer straight to the database (and its caching).
One change that may have a noticeable effect is that we now no longer prefill the `get_current_state_ids` cache on a state change. However, that query is relatively light, since its just a case of reading a table from the DB (unlike fetching state at an event which is more heavyweight). For deployments with workers this cache isn't even used.
Part of #12684
Try to avoid an OOM by checking fewer extremities.
Generally this is a big rewrite of _maybe_backfill, to try and fix some of the TODOs and other problems in it. It's best reviewed commit-by-commit.
