Recently I have observed that dendrite spends a lot of time (~390s) in
`selectRoomIDsWithAnyMembershipSQL` query
```
dendrite_syncapi=# select total_exec_time, left(query,100) from pg_stat_statements order by total_exec_time desc limit 5 ;
total_exec_time | left
--------------------+------------------------------------------------------------------------------------------------------
747826.5800519128 | SELECT event_id, id, headered_event_json, session_id, exclude_from_sync, transaction_id, history_vis
389130.5490339942 | SELECT DISTINCT room_id, membership FROM syncapi_current_room_state WHERE type = $2 AND state_key =
376104.17514700035 | SELECT psd.datname, xact_commit, xact_rollback, blks_read, blks_hit, tup_returned, tup_fetched, tup_
363644.164092031 | SELECT event_type_nid, event_state_key_nid, event_nid FROM roomserver_events WHERE event_nid = ANY($
58570.48104699995 | SELECT event_id, headered_event_json FROM syncapi_current_room_state WHERE room_id = $1 AND ( $2::te
(5 rows)
```
Explain analyze showed correct usage of `syncapi_room_state_unique`
index:
```
dendrite_syncapi=#
explain analyze SELECT distinct room_id, membership FROM syncapi_current_room_state WHERE type = 'm.room.member' AND state_key = '@qjfl:dendrite.stg.globekeeper.com';
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Unique (cost=2749.38..2749.56 rows=24 width=52) (actual time=2.933..2.956 rows=65 loops=1)
-> Sort (cost=2749.38..2749.44 rows=24 width=52) (actual time=2.932..2.937 rows=65 loops=1)
Sort Key: room_id, membership
Sort Method: quicksort Memory: 34kB
-> Index Scan using syncapi_room_state_unique on syncapi_current_room_state (cost=0.41..2748.83 rows=24 width=52) (actual time=0.030..2.890 rows=65 loops=1)
Index Cond: ((type = 'm.room.member'::text) AND (state_key = '@qjfl:dendrite.stg.globekeeper.com'::text))
Planning Time: 0.140 ms
Execution Time: 2.990 ms
(8 rows)
```
Multi-column indexes in Postgres shall perform well for leftmost
columns, but I gave it a try and created
`syncapi_current_room_state_type_state_key_idx` index. I could observe
significant performance improvement. Execution time dropped from 2.9 ms
to 0.24 ms:
```
explain analyze SELECT distinct room_id, membership FROM syncapi_current_room_state WHERE type = 'm.room.member' AND state_key = '@qjfl:dendrite.stg.globekeeper.com';
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------
Unique (cost=96.46..96.64 rows=24 width=52) (actual time=0.199..0.218 rows=65 loops=1)
-> Sort (cost=96.46..96.52 rows=24 width=52) (actual time=0.199..0.202 rows=65 loops=1)
Sort Key: room_id, membership
Sort Method: quicksort Memory: 34kB
-> Bitmap Heap Scan on syncapi_current_room_state (cost=4.53..95.91 rows=24 width=52) (actual time=0.048..0.139 rows=65 loops=1)
Recheck Cond: ((type = 'm.room.member'::text) AND (state_key = '@qjfl:dendrite.stg.globekeeper.com'::text))
Heap Blocks: exact=59
-> Bitmap Index Scan on syncapi_current_room_state_type_state_key_idx (cost=0.00..4.53 rows=24 width=0) (actual time=0.037..0.037 rows=65 loops=1)
Index Cond: ((type = 'm.room.member'::text) AND (state_key = '@qjfl:dendrite.stg.globekeeper.com'::text))
Planning Time: 0.236 ms
Execution Time: 0.242 ms
(11 rows)
```
Next improvement is skipping DISTINCT and rely on map assignment in
`SelectRoomIDsWithAnyMembership`. Execution time drops by almost half:
```
explain analyze SELECT room_id, membership FROM syncapi_current_room_state WHERE type = 'm.room.member' AND state_key = '@qjfl:dendrite.stg.globekeeper.com';
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on syncapi_current_room_state (cost=4.53..95.91 rows=24 width=52) (actual time=0.032..0.113 rows=65 loops=1)
Recheck Cond: ((type = 'm.room.member'::text) AND (state_key = '@qjfl:dendrite.stg.globekeeper.com'::text))
Heap Blocks: exact=59
-> Bitmap Index Scan on syncapi_current_room_state_type_state_key_idx (cost=0.00..4.53 rows=24 width=0) (actual time=0.021..0.021 rows=65 loops=1)
Index Cond: ((type = 'm.room.member'::text) AND (state_key = '@qjfl:dendrite.stg.globekeeper.com'::text))
Planning Time: 0.087 ms
Execution Time: 0.136 ms
(7 rows)
```
In our env we spend only 1s on inserting to table, so the write penalty
of creating an index should be small.
```
dendrite_syncapi=# select total_exec_time, left(query,100) from pg_stat_statements where query like '%INSERT%syncapi_current_room_state%' order by total_exec_time desc;
total_exec_time | left
--------------------+------------------------------------------------------------------------------------------------------
1139.9057619999971 | INSERT INTO syncapi_current_room_state (room_id, event_id, type, sender, contains_url, state_key, he
(1 row)
```
This PR does not require test modifications.
### Pull Request Checklist
<!-- Please read docs/CONTRIBUTING.md before submitting your pull
request -->
* [x] I have added added tests for PR _or_ I have justified why this PR
doesn't need tests.
* [x] Pull request includes a [sign
off](https://github.com/matrix-org/dendrite/blob/main/docs/CONTRIBUTING.md#sign-off)
Signed-off-by: `Piotr Kozimor <p1996k@gmail.com>`
Previously `LoadMembershipAtEvent` would fail if the state before one of
the events was not known, i.e. because it was an outlier. This modifies
it so that it gracefully handles not knowing the state and returns no
memberships instead, so that history visibility doesn't freak out and
kill `/sync` requests dead.
Some tweaks for the send-to-device consumers/producers:
- use `json.RawMessage` without marshalling it first
- try further devices (if available) if we failed to `PublishMsg` in the
producers
- some logging changes (to better debug E2EE issues)
Introduced index improves select query performance. Example execution time of `selectSendToDeviceMessagesSQL` query dropped from 80 ms to 15 ms. No sytest modifications are required.
### Pull Request Checklist
* [x] I have added added tests for PR _or_ I have justified why this PR doesn't need tests.
* [x] Pull request includes a [sign off](https://github.com/matrix-org/dendrite/blob/main/docs/CONTRIBUTING.md#sign-off)
Signed-off-by: `Piotr Kozimor <p1996k@gmail.com>`
This should fix an issue where we return less than the expected membership events, when doing an initial sync.
When doing an initial sync, the state limit is set to `math.MaxInt32`, while the default filter is set to 20.
- Reverts 9dc57122d9 as it was causing issues https://github.com/matrix-org/dendrite/issues/2660
- Updates the GMSL `DefaultStateFilter` to use a limit of 20 events
- Uses the timeline events to determine the new position instead of the state events
This should hopefully deflake Backfill works correctly with history visibility set to joined as we were using the default shared visibility, even if the events are set to joined (or something else)
* Use existing current room state if we have it
* Don't dedupe before applying the history vis filter
* Revert "Don't dedupe before applying the history vis filter"
This reverts commit d27c4a0874.
* Revert "Use existing current room state if we have it"
This reverts commit 5819b4a7ce.
* Tweaks
* Only return non-retired invites
* Revert "Only return non-retired invites"
This reverts commit 1150aa7f38.
* Check if we're doing an initial sync in the stream
* Add possibility to set history_visibility and user AccountType
* Add new DB queries
* Add actual history_visibility changes for /messages
* Add passing tests
* Extract check function
* Cleanup
* Cleanup
* Fix build on 386
* Move ApplyHistoryVisibilityFilter to internal
* Move queries to topology table
* Add filtering to /sync and /context
Some cleanup
* Add passing tests; Remove failing tests :(
* Re-add passing tests
* Move filtering to own function to avoid duplication
* Re-add passing test
* Use newly added GMSL HistoryVisibility
* Update gomatrixserverlib
* Set the visibility when creating events
* Default to shared history visibility
* Remove unused query
* Update history visibility checks to use gmsl
Update tests
* Remove unused statement
* Update migrations to set "correct" history visibility
* Add method to fetch the membership at a given event
* Tweaks and logging
* Use actual internal rsAPI, default to shared visibility in tests
* Revert "Move queries to topology table"
This reverts commit 4f0d41be9c.
* Remove noise/unneeded code
* More cleanup
* Try to optimize database requests
* Fix imports
* PR peview fixes/changes
* Move setting history visibility to own migration, be more restrictive
* Fix unit tests
* Lint
* Fix missing entries
* Tweaks for incremental syncs
* Adapt generic changes
Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
Co-authored-by: kegsay <kegan@matrix.org>
* Generic-based internal HTTP API (tested out on a few endpoints in the federation API)
* Add `PerformInvite`
* More tweaks
* Fix metric name
* Fix LookupStateIDs
* Lots of changes to clients
* Some serverside stuff
* Some error handling
* Use paths as metric names
* Revert "Use paths as metric names"
This reverts commit a9323a6a34.
* Namespace metric names
* Remove duplicate entry
* Remove another duplicate entry
* Tweak error handling
* Some more tweaks
* Update error behaviour
* Some more error tweaking
* Fix API path for `PerformDeleteKeys`
* Fix another path
* Tweak federation client proxying
* Fix another path
* Don't return typed nils
* Some more tweaks, not that it makes any difference
* Tweak federation client proxying
* Maybe fix the key backup test
* Bypass lazyLoadCache if we're doing an initial sync
* Make the linter happy again?
* Revert "Make the linter happy again?"
This reverts commit 52a5691ba3.
* Try that again
* Invalidate LazyLoadCache on initial syncs
* Remove unneeded check
* Add TODO
* Rename Invalite -> InvalidateLazyLoadedUser
* Thanks IDE
* Add race testing to tests, and fix a few small race conditions in the tests
* Enable run-sytest on MacOS
* Remove deadlock detecting mutex, per code review feedback
* Remove autoformatting related changes and a closure that is not needed
* Adjust to importing nats client as 'natsclient'
Signed-off-by: Brian Meek <brian@hntlabs.com>
* Clarify the use of gooseMutex to proect goose internal state
Signed-off-by: Brian Meek <brian@hntlabs.com>
* Remove no longer needed mutex for guarding goose
Signed-off-by: Brian Meek <brian@hntlabs.com>
* Fix query issue, only add "changed" users if we actually share a room
* Avoid log spam if context is done
* Undo changes to filterSharedUsers
* Add logging again..
* Fix SQLite shared users query
* Change query to include invited users
Issue: During conversation, under some conditions, sync cookie is not advanced, and, as a result, client loops on the same sync API call creating high traffic and CPU load.
Fix: pdu component of cookie was updated incorrectly.
* Add new db migration
* Update migrations
Remove goose
* Add possibility to test direct upgrades
* Try to fix WASM test
* Add checks for specific migrations
* Remove AddMigration
Use WithTransaction
Add Dendrite version to table
* Fix linter issues
* Update tests
* Update comments, outdent if
* Namespace migrations
* Add direct upgrade tests, skipping over one version
* Split migrations
* Update go version in CI
* Fix copy&paste mistake
* Use contexts in migrations
Co-authored-by: kegsay <kegan@matrix.org>
Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
* Membership updater refactoring
* Pass in membership state
* Use membership check rather than referring to state directly
* Delete irrelevant membership states
* We don't need the leave event after all
* Tweaks
* Put a log entry in that I might stand a chance of finding
* Be less panicky
* Tweak invite handling
* Don't freak if we can't find the event NID
* Use event NID from `types.Event`
* Clean up
* Better invite handling
* Placate the almighty linter
* Blacklist a Sytest which is otherwise fine under Complement for reasons I don't understand
* Fix the sytest after all (thanks @S7evinK for the spot)
* Add function to the sync API storage package for filtering shared users
* Use the database instead of asking the RS API
* Fix unit tests
* Fix map handling in `filterSharedUsers`
* Try Ristretto cache
* Tweak
* It's beautiful
* Update GMSL
* More strict keyable interface
* Fix that some more
* Make less panicky
* Don't enforce mutability checks for now
* Determine mutability using deep equality
* Tweaks
* Namespace keys
* Make federation caches mutable
* Update cost estimation, add metric
* Update GMSL
* Estimate cost for metrics better
* Reduce counters a bit
* Try caching events
* Some guards
* Try again
* Try this
* Use separate caches for hopefully better hash distribution
* Fix bug with admitting events into cache
* Try to fix bugs
* Check nil
* Try that again
* Preserve order jeezo this is messy
* thanks VS Code for doing exactly the wrong thing
* Try this again
* Be more specific
* aaaaargh
* One more time
* That might be better
* Stronger sorting
* Cache expiries, async publishing of EDUs
* Put it back
* Use a shared cache again
* Cost estimation fixes
* Update ristretto
* Reduce counters a bit
* Clean up a bit
* Update GMSL
* 1GB
* Configurable cache sizees
* Tweaks
* Add `config.DataUnit` for specifying friendly cache sizes
* Various tweaks
* Update GMSL
* Add back some lazy loading caching
* Include key in cost
* Include key in cost
* Tweak max age handling, config key name
* Only register prometheus metrics if requested
* Review comments @S7evinK
* Don't return errors when creating caches (it is better just to crash since otherwise we'll `nil`-pointer exception everywhere)
* Review comments
* Update sample configs
* Update GHA Workflow
* Update Complement images to Go 1.18
* Remove the cache test from the federation API as we no longer guarantee immediate cache admission
* Don't check the caches in the renewal test
* Possibly fix the upgrade tests
* Update to matrix-org/gomatrixserverlib#322
* Update documentation to refer to Go 1.18
This should avoid coercions between signed and unsigned ints which might fix problems like `sql: converting argument $5 type: uint64 values with high bit set are not supported`.
* Check state before event
* Tweaks
* Refactor a bit, include in output events
* Don't waste time if soft failed either
* Tweak control flow, comments, use GMSL history visibility type
* syncapi: don't return early for no-op incremental syncs
Comments explain why, but basically it's an inefficient use
of bandwidth and some sytests rely on /sync to block.
* Honour timeouts
* Actually return a response with timeout=0
* bugfix: fix race condition when updating presence via /sync
Previously when presence is updated via /sync, we would send the presence update
asyncly via NATS. This created a race condition:
- If the presence update is processed quickly, the /sync which triggered the presence
update would see an online presence.
- If the presence update was processed slowly, the /sync which triggered the presence
update would see an offline presence.
This is the root cause behind the flakey sytest: 'User sees their own presence in a sync'.
The fix is to ensure we update the database/advance the stream position synchronously
for local users.
* Bugfix for test
