Without setting JEMALLOC_OVERRIDE, we end up linking to two different
jemalloc builds. Once dynamically, as a transitive dependency through
rocksdb, and a second time to the static jemalloc that tikv-jemalloc-sys
builds.
Previously, we were returning redundant member count updates or encrypted
device updates from the /sync endpoint in some cases. The extra member
count updates are spec-compliant, but unnecessary, while the extra
encrypted device updates violate the spec.
The refactor necessary to fix this bug is also necessary to support
filtering on state events in sync.
Details:
Joined room incremental sync needs to examine state events for four
purposes:
1. determining whether we need to return an update to room member counts
2. determining the set of left/joined devices for encrypted rooms
(returned in `device_lists`)
3. returning state events to the client (in `rooms.joined.*.state`)
4. tracking which member events we have sent to the client, so they can
be omitted on future requests when lazy-loading is enabled.
The state events that we need to examine for the first two cases is member
events in the delta between `since` and the end of `timeline`. For the
second two cases, we need the delta between `since` and the start of
`timeline`, plus contextual member events for any senders that occur in
`timeline`. The second list is subject to filtering, while the first is
not.
Before this change, we were using the same set of state events that we are
returning to the client (cases 3/4) to do the analysis for cases 1/2.
In a compliant implementation, this would result in us missing some
relevant member events in 1/2 in addition to seeing redundant member
events. In current conduwuit this is not the case because the set of
events that we return to the client is always a superset of the set that
is needed for cases 1/2. This is because we don't support filtering, and
we have an existing bug[1] where we are returning the delta between
`since` and the end of `timeline` rather than the start.
[1]: https://github.com/girlbossceo/conduwuit/issues/361
Fixing this is necessary to implement filtering because otherwise
we would start missing some member events for member count or encrypted
device updates if the relevant member events are rejected by the filter.
This would be much worse than our current behavior.
This cache can serve invalid responses, and has an extremely low hit
rate.
It serves invalid responses because because it's only keyed off
the `since` parameter, but many of the other request parameters also
affect the response or it's side effects. This will become worse once we
implement filtering, because there will be a wider space of parameters
with different responses. This problem is fixable, but not worth it
because of the low hit rate.
The low hit rate is because normal clients will always issue the next
sync request with `since` set to the `prev_batch` value of the previous
response. The only time we expect to see multiple requests with the same
`since` is when the response is empty, but we don't cache empty
responses.
This was confirmed experimentally by logging cache hits and misses over
15 minutes with a wide variety of clients. This test was run on
matrix.computer.surgery, which has only a few active users, but a
large volume of sync traffic from many rooms. Over the test period, we
had 3 hits and 5309 misses. All hits occurred in the first minute, so I
suspect that they had something to do with client recovery from an
offline state. The clients that were connected during the test are:
- element web
- schildichat web
- iamb
- gomuks
- nheko
- fractal
- fluffychat web
- fluffychat android
- cinny web
- element android
- element X android
Fixes: #336
Hardcoding the output path to something in CWD is a pain if you're running
conduwuit through systemd or similar. Also made the error message when
it's unable to create the output file a little more friendly.
Previously, enabling the `tracing_flame` or `allow_jaeger` options would
prevent any logs from being written to stdout. In addition, enabling the
`allow_jaeger` option would inhibit the `tracing_flame` option.
Now that we have a way to use separate tracing filters with different
layers, we can enable all three at the same time without issues.
This commit also prevents the `debug log_level` command from modifying
the `tracing-flame` filter. This was supported previously, but I don't
think it's something that you would ever want to do intentionally. Now
that we have both the normal log filter and the `tracing-flame` filter
enabled at the same time, we want to `debug log_level` to only modify the
normal filter.
This turned out to be quite hairy, mostly because we need to apply the
config's log level filter to the actual logs (stdout and, optionally
sentry), but do not want to filter out the tokio tracing events needed by
the console_subscriber. I hit several edge cases in tracing getting
this to work, and we now depend on a git version of tracing with a
backported patch :(
Benjamin Lee