This matches the multi instance writer ID generator class which can
both handle advancing the current token over replication and by calling
the database.
Fix MSC3030 `/timestamp_to_event` endpoint returning `outliers` that it has no idea whether are near a gap or not (and therefore unable to determine whether it's actually the closest event). The reason Synapse doesn't know whether an `outlier` is next to a gap is because our gap checks rely on entries in the `event_edges`, `event_forward_extremeties`, and `event_backward_extremities` tables which is [not the case for `outliers`](2c63cdcc3f/docs/development/room-dag-concepts.md (outliers)).
Also fixes MSC3030 Complement `can_paginate_after_getting_remote_event_from_timestamp_to_event_endpoint` test flake. Although this acted flakey in Complement, if `sync_partial_state` raced and beat us before `/timestamp_to_event`, then even if we retried the failing `/context` request it wouldn't work until we made this Synapse change. With this PR, Synapse will never return an `outlier` event so that test will always go and ask over federation.
Fix https://github.com/matrix-org/synapse/issues/13944
### Why did this fail before? Why was it flakey?
Sleuthing the server logs on the [CI failure](https://github.com/matrix-org/synapse/actions/runs/3149623842/jobs/5121449357#step:5:5805), it looks like `hs2:/timestamp_to_event` found `$NP6-oU7mIFVyhtKfGvfrEQX949hQX-T-gvuauG6eurU` as an `outlier` event locally. Then when we went and asked for it via `/context`, since it's an `outlier`, it was filtered out of the results -> `You don't have permission to access that event.`
This is reproducible when `sync_partial_state` races and persists `$NP6-oU7mIFVyhtKfGvfrEQX949hQX-T-gvuauG6eurU` as an `outlier` before we evaluate `get_event_for_timestamp(...)`. To consistently reproduce locally, just add a delay at the [start of `get_event_for_timestamp(...)`](cb20b885cb/synapse/handlers/room.py (L1470-L1496)) so it always runs after `sync_partial_state` completes.
```py
from twisted.internet import task as twisted_task
d = twisted_task.deferLater(self.hs.get_reactor(), 3.5)
await d
```
In a run where it passes, on `hs2`, `get_event_for_timestamp(...)` finds a different event locally which is next to a gap and we request from a closer one from `hs1` which gets backfilled. And since the backfilled event is not an `outlier`, it's returned as expected during `/context`.
With this PR, Synapse will never return an `outlier` event so that test will always go and ask over federation.
Fix https://github.com/matrix-org/synapse/issues/13856
Fix https://github.com/matrix-org/synapse/issues/13865
> Discovered while trying to make Synapse fast enough for [this MSC2716 test for importing many batches](https://github.com/matrix-org/complement/pull/214#discussion_r741678240). As an example, disabling the `have_seen_event` cache saves 10 seconds for each `/messages` request in that MSC2716 Complement test because we're not making as many federation requests for `/state` (speeding up `have_seen_event` itself is related to https://github.com/matrix-org/synapse/issues/13625)
>
> But this will also make `/messages` faster in general so we can include it in the [faster `/messages` milestone](https://github.com/matrix-org/synapse/milestone/11).
>
> *-- https://github.com/matrix-org/synapse/issues/13856*
### The problem
`_invalidate_caches_for_event` doesn't run in monolith mode which means we never even tried to clear the `have_seen_event` and other caches. And even in worker mode, it only runs on the workers, not the master (AFAICT).
Additionally there was bug with the key being wrong so `_invalidate_caches_for_event` never invalidates the `have_seen_event` cache even when it does run.
Because we were using the `@cachedList` wrong, it was putting items in the cache under keys like `((room_id, event_id),)` with a `set` in a `set` (ex. `(('!TnCIJPKzdQdUlIyXdQ:test', '$Iu0eqEBN7qcyF1S9B3oNB3I91v2o5YOgRNPwi_78s-k'),)`) and we we're trying to invalidate with just `(room_id, event_id)` which did nothing.
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.
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)
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.
I've seen a few errors which can only plausibly be explained by the calculated
event id for an event being different from the ID of the event in the
database. It should be cheap to check this, so let's do so and raise an
exception.
Multiple calls to `EventsWorkerStore._get_events_from_cache_or_db` can
reuse the same database fetch, which is initiated by the first call.
Ensure that cancelling the first call doesn't cancel the other calls
sharing the same database fetch.
Signed-off-by: Sean Quah <seanq@element.io>
When we join a room via the faster-joins mechanism, we end up with "partial
state" at some points on the event DAG. Many parts of the codebase need to
wait for the full state to load. So, we implement a mechanism to keep track of
which events have partial state, and wait for them to be fully-populated.
We work through all the events with partial state, updating the state at each
of them. Once it's done, we recalculate the state for the whole room, and then
mark the room as having complete state.
If we're missing most of the events in the room state, then we may as well call the /state endpoint, instead of individually requesting each and every event.
* `@cached` can now take an `uncached_args` which is an iterable of names to not use in the cache key.
* Requires `@cached`, @cachedList` and `@lru_cache` to use keyword arguments for clarity.
* Asserts that keyword-only arguments in cached functions are not accepted. (I tested this briefly and I don't believe this works properly.)
When we get a partial_state response from send_join, store information in the
database about it:
* store a record about the room as a whole having partial state, and stash the
list of member servers too.
* flag the join event itself as having partial state
* also, for any new events whose prev-events are partial-stated, note that
they will *also* be partial-stated.
We don't yet make any attempt to interpret this data, so API calls (and a bunch
of other things) are just going to get incorrect data.
If the latest event in a thread was edited than the original
event content was included in bundled aggregation for
threads instead of the edited event content.
* Push `get_room_{min,max_stream_ordering}` into StreamStore
Both implementations of this are identical, so we may as well push it down and
get rid of the abstract base class nonsense.
* Remove redundant `StreamStore` class
This is empty now
* Remove redundant `get_current_events_token`
This was an exact duplicate of `get_room_max_stream_ordering`, so let's get rid
of it.
* newsfile
MSC3030: https://github.com/matrix-org/matrix-doc/pull/3030
Client API endpoint. This will also go and fetch from the federation API endpoint if unable to find an event locally or we found an extremity with possibly a closer event we don't know about.
```
GET /_matrix/client/unstable/org.matrix.msc3030/rooms/<roomID>/timestamp_to_event?ts=<timestamp>&dir=<direction>
{
"event_id": ...
"origin_server_ts": ...
}
```
Federation API endpoint:
```
GET /_matrix/federation/unstable/org.matrix.msc3030/timestamp_to_event/<roomID>?ts=<timestamp>&dir=<direction>
{
"event_id": ...
"origin_server_ts": ...
}
```
Co-authored-by: Erik Johnston <erik@matrix.org>
The previous fix for the ongoing event fetches counter
(8eec25a1d9) was both insufficient and
incorrect.
When the database is unreachable, `_do_fetch` never gets run and so
`_event_fetch_ongoing` is never decremented.
The previous fix also moved the `_event_fetch_ongoing` decrement outside
of the `_event_fetch_lock` which allowed race conditions to corrupt the
counter.
When an event fetcher aborts due to an exception, `_event_fetch_ongoing`
must be decremented, otherwise the event fetcher would never be
replaced. If enough event fetchers were to fail, no more events would be
fetched and requests would get stuck waiting for events.
The code to deduplicate repeated fetches of the same set of events was
N^2 (over the number of events requested), which could lead to a process
being completely wedged.
The main fix is to deduplicate the returned deferreds so we only await
on a deferred once rather than many times. Seperately, when handling the
returned events from the defrered we only add the events we care about
to the event map to be returned (so that we don't pay the price of
inserting extraneous events into the dict).
Empirically, this helped my server considerably when handling gaps in Matrix HQ. The problem was that we would repeatedly call have_seen_events for the same set of (50K or so) auth_events, each of which would take many minutes to complete, even though it's only an index scan.
`keylen` seems to be a thing that is frequently incorrectly set, and we don't really need it.
The only time it was used was to figure out if we had removed a subtree in `del_multi`, which we can do better by changing `TreeCache.pop` to return a different type (`TreeCacheNode`).
Commits should be independently reviewable.
