Original patch by Matthias. Benjamin modified the logic to include
logging when an event was rejected, for consistency with the existing
check on device key updates.
Co-authored-by: Benjamin Lee <benjamin@computer.surgery>
This ensures that the tokenization algorithm will remain in sync between
querying, indexing, and deindexing. The existing code had slightly
different behavior for querying, because it did not discard words with
>50 bytes. This was inconsequential, because >50 byte tokens are never
present in the index.
The previous code would drop some events entirely if any events between
`skip` and `skip + limit` were not visible to the user. This would cause
the set of events skipped by the `skip(skip)` method to extend past
`skip` in the raw result set, because `skip(skip)` was being called
*after* filtering out invisible events.
This bug will become much more severe with a full filtering
implementation, because it will be more likely for events to be filtered
out. Currently, it is only possible to trigger with rooms that have
history visibility set to "invited" or "joined".
The previous code would fail to return next_batch if any of the events
in the window were not visible to the user. It would also return an
unnecessary next_batch when no more results are available if the total
number of results is exactly `skip + limit`.
This bug will become much more severe with a full filtering
implementation, because we will be more likely to trigger it by
filtering out events in a search call. Currently, it is only possible to
trigger with rooms that have history visibility set to "invited" or
"joined".
* Extract into reusable function (we'll need this later)
* Merge the values we want to override over the defaults, instead of
dropping the defaults completely
* Don't unnecessarily allocate (the `vec![]` usage is gone)
Both the hand-rolled parser and serialization were wrong in countless
ways. The current Ruma parser is much better, and the Ruma serialization
will be fixed by https://github.com/ruma/ruma/pull/1830.
This is nice when messing around with the nix configuration, since
you'll often end up building rocksdb *many* times.
We shouldn't need to do anything to get the logs in CI because we're
calling `nix-build-and-cache ci` at the beginning
Without this, we're building a static rocksdb inside the rust-rocksdb
build script every time. As far as I know this doesn't change clippy's
behavior, but it does take a *long* time.
Because the nix rocksdb build depends on the jemalloc feature, you need
to use a different devshell when passing --all-features to cargo than
the default.
Some of the features affect nix dependencies, so we need to have a
full feature list available when constructing the nix derivation. This
incidentally fixes the bug where we weren't enabling jemalloc on rocksdb
in CI/devshells, because jemalloc is now a default feature. It does not
fix the more general class of that issue, where CI is performing an
`--all-features` build in a nix devshell built for default-features.
I am now passing `--no-default-features` to cargo, and having it use our
unified feature list rather than duplicating the unification inside cargo.
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.
This fixes dynamically-linked jemalloc builds, for the reasons described
in <https://github.com/girlbossceo/conduwuit/pull/400#issue-2316700200>.
We inherited the disabled-by-default setting from conduit. Conduwuit
change it to enabled-by-default in [1]. This can make things confusing
for users migrating from conduwuit to grapevine, especially since we
currently do not log a warning when federation is disabled.
[1]: 24605e151d
