This is a refactor of the federation destination queues.
It fixes a few things, namely:
- actually retry outgoing events with backoff behaviour
- obtain enough events from the database to fill messages as much as
possible
- minimize the amount of running goroutines
- use pure timers for backoff
- don't restart queue unless necessary
- close the background task when backing off
- increase max edus in a transaction to match the spec
- cleanup timers more aggresively to reduce memory usage
- add jitter to backoff timers to reduce resource spikes
- add a bunch of tests (with real and fake databases) to ensure
everything is working
To avoid returning results from e.g. `/redact` on `/sendToDevice`
requests.
Takes the raw URL path and uses `filepath.Dir` to remove the `txnID`
(file) from it.
Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
commit c6e18c18e93b54c006c6b4d0044aa53a0735a906
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date: Fri Oct 14 15:42:58 2022 +0100
Changelog and version bump
This prevents us from holding onto durable consumers indefinitely for
rooms that have long since turned inactive, since they do have a bit of
a processing overhead in the NATS Server. If we clear up a consumer and
then a room becomes active again, the consumer gets recreated as needed.
The threshold is set to 24 hours for now, we can tweak it later if needs
be.
Fixes `outliers whose auth_events are in a different room are correctly
rejected`, by validating that auth events are all from the same room and
not using rejected events for event auth.
This fixes some edge cases where federation queue backoffs and
blacklisting weren't behaving as expected.
It also adds new tests for the federation queues to ensure their
behaviour continues to work correctly.
Sytest was using a wrong `history_visibility` for `invited`
(https://github.com/matrix-org/sytest/pull/1303), so `invited` was
passing for the wrong reason (-> defaulted to `shared`, as `invite`
wasn't understood).
This change now handles missing events like Synapse, if a server isn't
allowed to see the event, it gets a redacted version of it, making the
`get_missing_events` tests pass.
Improves the control flow of `GetStateDeltas` for clarity and possibly
also fixes a bug where duplicate state delta entries could be inserted
with different memberships instead of being correctly overridden by
`join`.
This ensures that the joined hosts in the federation API are correct
after the state is rewritten. This might fix some races around the time
of joining federated rooms.
Use the stream positions of the notifier, which might have advanced
since setting it at the beginning of the loop. This possibly helps in
reducing roundtrips to the SyncAPI, just because we didn't fetch the
latest data.
Also fixes a minor oversight in the receipts stream.
If the private key file is lost, it's often possible to retrieve the
public key from another server elsewhere, so we should make it possible
to configure it in that way.
First attempt at removing empty fields from `/sync` responses. Needs
https://github.com/matrix-org/sytest/pull/1298 to keep Sytest happy.
Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
This PR queries PostgreSQL for the `max_connections` and
`superuser_reserved_connections` settings and then ensures that
Dendrite's `max_open_conns` doesn't exceed the allowed value.
This is a really common source of configuration problems and can either
result in blocking queries or deadlocks, so it seems reasonable that we
complain as loudly as possible when it happens.
