This is essentially an implementation of the proposal made at https://hackmd.io/@richvdh/BJYXQMQHO, though the details have ended up looking slightly different.
This implements similar behavior to sytest where a matching branch is used,
if one exists. This is useful when needing to modify both application code
and tests at the same time. The following rules are used to find a matching
complement branch:
1. Search for the branch name of the pull request. (E.g. feature/foo.)
2. Search for the base branch of the pull request. (E.g. develop or release-vX.Y.)
3. Search for the reference branch of the commit. (E.g. master or release-vX.Y.)
4. Fallback to 'master', the default complement branch name.
Spawned from missing messages we were seeing on `matrix.org` from a
federated Gtiter bridged room, https://gitlab.com/gitterHQ/webapp/-/issues/2770.
The underlying issue in Synapse is tracked by https://github.com/matrix-org/synapse/issues/10066
where the message and join event race and the message is `soft_failed` before the
`join` event reaches the remote federated server.
Less soft_failed events = better and usually this should only trigger for events
where people are doing bad things and trying to fuzz and fake everything.
This PR updates the build tags that we perform Complement runs with to match our [buildkite pipeline](618b3e90bc/synapse/pipeline.yml (L570)), as well as adding `msc2403` (as it will be required once #9359 is merged). Build tags are what we use to determine which tests to run in Complement (really it determines which test files are compiled into the final binary).
I haven't put in a comment about updating the buildkite side here, as we've decided to migrate fully to GitHub Actions anyhow.
With the prior format, 1.33.0 / 1.33.1 / 1.33.2 got separate branches:
release-v1.33.0
release-v1.33.1
release-v1.33.2
Under the new model, all three would share a common branch:
release-v1.33
As before, RCs and actual releases exist as tags on these branches.
This better reflects our support model, e.g., that the "1.33" series had
a formal release followed by two patches / updates.
Signed-off-by: Dan Callahan <danc@element.io>
Fixes#1834.
`get_new_events_for_appservice` internally calls `get_events_as_list`, which will filter out any rejected events. If all returned events are filtered out, `_notify_interested_services` will return without updating the last handled stream position. If there are 100 consecutive such events, processing will halt altogether.
Breaking the loop is now done by checking whether we're up-to-date with `current_max` in the loop condition, instead of relying on an empty `events` list.
Signed-off-by: Willem Mulder <14mRh4X0r@gmail.com>
If backfilling is slow then the client may time out and retry, causing
Synapse to start a new `/backfill` before the existing backfill has
finished, duplicating work.
This adds quite a lot of OpenTracing decoration for database activity. Specifically it adds tracing at four different levels:
* emit a span for each "interaction" - ie, the top level database function that we tend to call "transaction", but isn't really, because it can end up as multiple transactions.
* emit a span while we hold a database connection open
* emit a span for each database transaction - actual actual transaction.
* emit a span for each database query.
I'm aware this might be quite a lot of overhead, but even just running it on a local Synapse it looks really interesting, and I hope the overhead can be offset just by turning down the sampling frequency and finding other ways of tracing requests of interest (eg, the `force_tracing_for_users` setting).
The existing tracing reports an error each time there is a timeout, which isn't
really representative.
Additionally, we log things about the way `wait_for_events` works
(eg, the result of the callback) to the *parent* span, which is confusing.
So that they render nicely in mdbook (see #10086), and so that we no longer have a mix of structured text languages in our documentation (excluding files outside of `docs/`).
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.
* Make `invalidate` and `invalidate_many` do the same thing
... so that we can do either over the invalidation replication stream, and also
because they always confused me a bit.
* Kill off `invalidate_many`
* changelog
`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.
* Fix /upload 500'ing when presented a very large image
Catch DecompressionBombError and re-raise as ThumbnailErrors
* Set PIL's MAX_IMAGE_PIXELS to match homeserver.yaml
to get it to bomb out quicker, to load less into memory
in the case of super large images
* Add changelog entry for 10029
https://github.com/matrix-org/synapse/issues/9962 uncovered that we accidentally removed all but one of the presence updates that we store in the database when persisting multiple updates. This could cause users' presence state to be stale.
The bug was fixed in #10014, and this PR just adds a test that failed on the old code, and was used to initially verify the bug.
The test attempts to insert some presence into the database in a batch using `PresenceStore.update_presence`, and then simply pulls it out again.
Fixes: https://github.com/matrix-org/synapse/issues/9962
This is a fix for above problem.
I fixed it by swaping the order of insertion of new records and deletion of old ones. This ensures that we don't delete fresh database records as we do deletes before inserts.
Signed-off-by: Marek Matys <themarcq@gmail.com>
Also add support for giving a callback to generate the JSON object to
verify. This should reduce memory usage, as we no longer have the event
in memory in dict form (which has a large memory footprint) for extend
periods of time.
Instead of parsing the full response to `/send_join` into Python objects (which can be huge for large rooms) and *then* parsing that into events, we instead use ijson to stream parse the response directly into `EventBase` objects.
To be more consistent with similar code. The check now automatically
raises an AuthError instead of passing back a boolean. It also absorbs
some shared logic between callers.
- use a tuple rather than a list for the iterable that is passed into the
wrapped function, for performance
- test that we can pass an iterable and that keys are correctly deduped.
It's not obvious that instances of SQLBaseStore each need their own
instances of random.SystemRandom(); let's just use random directly.
Introduced by 52839886d6
Signed-off-by: Dan Callahan <danc@element.io>
We can get away with just catching UnicodeError here.
⋮
+-- ValueError
| +-- UnicodeError
| +-- UnicodeDecodeError
| +-- UnicodeEncodeError
| +-- UnicodeTranslateError
⋮
https://docs.python.org/3/library/exceptions.html#exception-hierarchy
Signed-off-by: Dan Callahan <danc@element.io>
Functionally identical, but more obviously cryptographically secure.
...Explicit is better than implicit?
Avoids needing to know that SystemRandom() implies a CSPRNG, and
complies with the big scary red box on the documentation for random:
> Warning:
> The pseudo-random generators of this module should not be used for
> security purposes. For security or cryptographic uses, see the
> secrets module.
https://docs.python.org/3/library/random.html
Signed-off-by: Dan Callahan <danc@element.io>
This should help ensure that equivalent results are achieved between
homeservers querying for the summary of a space.
This implements modified MSC1772 rules, according to MSC2946.
The different is that the origin_server_ts of the m.room.create event
is not used as a tie-breaker since this might not be known if the
homeserver is not part of the room.
Per changes in MSC2946, the C-S and S-S APIs for spaces summary
should use GET requests.
Until this is stable, the POST endpoints still exist.
This does not switch federation requests to use the GET version yet
since it is newly added and already deployed servers might not support
it. When switching to the stable endpoint we should switch to GET
requests.
MSC1772 specifies the m.room.create event should be sent as part
of the invite_state. This was done optionally behind an experimental
flag, but is now done by default due to MSC1772 being approved.
Now that cross signing exists there is much less of a need for other people to look at devices and verify them individually. This PR adds a config option to allow you to prevent device display names from being shared with other servers.
Signed-off-by: Aaron Raimist <aaron@raim.ist>
We were pulling the full auth chain for the room out of the DB each time
we backfilled, which can be *huge* for large rooms and is totally
unnecessary.
The hope here is that by moving all the schema files into synapse/storage/schema, it gets a bit easier for newcomers to navigate.
It certainly got easier for me to write a helpful README. There's more to do on that front, but I'll follow up with other PRs for that.
This is an update based on changes to MSC2946. The origin_server_ts
of the m.room.create event is copied into the creation_ts field for each
room returned from the spaces summary.
Synapse can be quite memory intensive, and unless care is taken to tune
the GC thresholds it can end up thrashing, causing noticable performance
problems for large servers. We fix this by limiting how often we GC a
given generation, regardless of current counts/thresholds.
This does not help with the reverse problem where the thresholds are set
too high, but that should only happen in situations where they've been
manually configured.
Adds a `gc_min_seconds_between` config option to override the defaults.
Fixes#9890.
* Add healthcheck startup delay by 5secs and reduced interval check to 15s
to reduce waiting time for docker aware edge routers bringing an
instance online
This leaves out all optional keys from /sync. This should be fine for all clients tested against conduit already, but it may break some clients, as such we should check, that at least most of them don't break horribly and maybe back out some of the individual changes. (We can probably always leave out groups for example, while the others may cause more issues.)
Signed-off-by: Nicolas Werner <nicolas.werner@hotmail.de>
