Many fields here are not argument of the `send` method, but are set
later with `Span::record`. Grepping all these fields reveal they are all
set except `request_body` apparently.
This patch is twofold. First off, it provides a new schema allowing to
improve the performance of `SqliteEventCacheStore` for 100_000 events
from 6.7k events/sec to 284k events/sec on my machine.
Second, it now assumes that `EventCacheStore` does NOT store invalid
events. It was already the case, but the SQLite schema was not rejecting
invalid event in case some were handled. It's now explicitely forbidden.
This patch removes the `EventsPostProcessing` type, it assumes
`with_events_muts` will always return events that will be post-process.
The case where `EventsPostProcessing::None` becomes a `vec![]`.
This patch updates `maybe_apply_new_redaction` to remove the first
`&RoomVersionId` argument. Indeed, due to the refactoring, it's now
possible for `maybe_apply_new_redaction` to read this value directly
from `Self::room_version`.
This patch updates `maybe_apply_new_redaction` to use `find_event`, so
that the target event is looked up in memory or in the store.
The case where it is in the store is a simple `todo!()` for the moment.
I wanted to separate the update of the `maybe_apply_new_redaction`
signature from the `InStore` implementation. The method is now async and
returns a `Result`.
This patch introduces `EventLocation` to know if an event has been found
in the memory (in `RoomEvents`) or in the store (in `EventCacheStore`).
This is used by the `RoomEventCacheState::find_event`.
This patch moves the `maybe_apply_new_redaction` method from
`RoomEvents` inside `RoomEventCacheState` so that it has an access
to the store (necessary for the next patch). This patch creates a new
`RoomEvents::replace_event_at` method, which is a thin wrapper around
`LinkedChunk::replace_item_at`.
This patch renames `sync_timeline_events` into `timeline_events`.
Moreover, this change has spotted a possible improvement
in `AllEventsCache` where it now receives events from
`collect_valid_and_duplicated_events`, which allows to only store valid
events in it.
This patch updates the callback passed to `with_events_mut`. It now
returns an `EventsPostProcessing` which can automatically run the, now
inlined, `on_new_events`.
This patch updates where the `RoomVersionId` is also stored. It's not
held by `RoomEventCacheState` instead of `RoomEventCacheInner`.
I have to disable LTO every time I'm building any final binary using the
SDK, because otherwise, the builds can take easily more than 10 minutes
to complete, killing iteration times, and making it almost impractical
to use the programs (benchmarks, or multiverse).
I think it should be the decision of the final embedder to enable or
disable LTO, and that for the purpose of our own binaries hosted in the
SDK repository, we don't need the absolute best performance (or, for the
sake of benchmarking, we can tweak the profiling profile).
The linked chunk always starts with an empty events chunk. If we receive
a gap from sync, then we will immediately push a gap chunk; in this
case, it might be better to replace the events chunk with a gap chunk.
This is equivalent to removing the empty events chunk, after pushing
back the first one (we can't do it before, otherwise we might get rid of
the only chunk in the linked chunk, which breaks the invariant that a
linked chunk is never empty).
