First bug: if an error occured during pref loading code, it would fail. If the person then readied up, it would likely cause the round to fail to start.
Why could they ready up? The code only checks that the prefs finished loading, not that they finished loading *successfully*. Whoops.
Anyways, now people get kicked if their prefs fail to load. And I improved the error handling.
Second bug: if a user disconnected while their prefs were loading, it would cause an exception. This exception would go unobserved on lobby servers or raise through gameticker on non-lobby servers.
This happened even on a live server once and then triggered the first bug, but idk how.
Fixed this by properly plumbing through cancellation into the preferences loading code. The stuff is now cancelled properly.
Third bug: if somebody has a loadout item with a playtime requirement active, load-time sanitization of player prefs could run into a race condition because the sanitization can happen *before* play time was loaded.
Fixed by moving pref sanitizations to a later stage in the load process.
How can ONE DATABASE COLUMN have so many cursed issues I don't know, but it certainly pissed off the devil in its previous life.
The start_date column on round entities in the database was added by https://github.com/space-wizards/space-station-14/pull/21153. For some reason, this PR gave the column a nonsensical default value instead of making it nullable. This default value causes the code from #25280 to break. It actually trips an assert though that's not what the original issue report ran into.
This didn't get noticed on wizden servers because we at some point backfilled the start_date column based on the stored admin logs.
So I change the database model to make this column nullable, updated the C# code to match, and made the existing migration set the invalid values to be NULL instead. Cool.
Wait how's SQLite handle in this scenario anyways? Well actually turns out the column was *completely broken* in the first place!
The code for inserting into the round table was copy pasted between SQLite and PostgreSQL, with the only difference being that the SQLite key manually assigned the primary key instead of letting SQLite AUTOINCREMENT it. And then the code to give a start_date value was only added to the PostgreSQL version (which is actually in the base class already). So for SQLite that column's been filled up with the same invalid default the whole time.
Why was the code manually assigning a PK? I checked the SQLite docs for AUTOINCREMENT[1], and the behavior seems appropriate.
I removed the SQLite-specific code path and it just seems to work regardless. The migration just sets the old values to NULL too.
BUT WAIT, THERE'S MORE!
Turns out just doing the migration on SQLite is a pain in the ass! EF Core has to create a new table to apply the nullability change, because SQLite doesn't support proper ALTER COLUMN. This causes the generated SQL commands to be weird and the UPDATE for the migration goes BEFORE the nullability change... I ended up having to make TWO migrations for SQLite. Yay.
Fixes#26800
[1]: https://www.sqlite.org/autoinc.html
God bloody christ. There's like three layers of shit here.
So firstly, apparently we were still using Npgsql.EnableLegacyTimestampBehavior. This means that time values (which are stored UTC in the database) were converted to local time when read out. This meant they were passed around as kind Local to clients (instead of UTC in the case of SQLite). That's easy enough to fix just turn off the flag and fix the couple spots we're passing a local DateTime ez.
Oh but it turns out there's a DIFFERENT problem with SQLite: See SQLite we definitely store the DateTimes as UTC, but when Microsoft.Data.Sqlite reads them it reads them as Kind Unspecified instead of Utc.
Why are these so bad? Because the admin notes system passes DateTime instances from EF Core straight to the rest of the game code. And that means it's a PAIN IN THE ASS to run the necessary conversions to fix the DateTime instances. GOD DAMNIT now I have to make a whole new set of "Record" entities so we avoid leaking the EF Core model entities. WAAAAAAA.
Fixes#19897
* Give .props files 2-space indents.
* Move to Central Package Management.
Allows us to store NuGet package versions all in one place. Yay!
* Update NuGet packages and fix code for changes.
Notable:
Changes to ILVerify.
Npgsql doesn't need hacks for inet anymore, now we need hacks to make the old code work with this new reality.
NUnit's analyzers are already complaining and I didn't even update it to 4.x yet.
TerraFX changed to GetLastSystemError so error handling had to be changed.
Buncha APIs have more NRT annotations.
* Remove dotnet-eng NuGet package source.
I genuinely don't know what this was for, and Central Package Management starts throwing warnings about it, so YEET.
* Remove Robust.Physics project.
Never used.
* Remove erroneous NVorbis reference.
Should be VorbisPizza and otherwise wasn't used.
* Sandbox fixes
* Remove unused unit test package references.
Castle.Core and NUnit.ConsoleRunner.
* Update NUnit to 4.0.1
This requires replacing all the old assertion methods because they removed them 🥲
* Oh so that's what dotnet-eng was used for. Yeah ok that makes sense.
* Add Robust.Analyzers.Test
* Update submodule
* commit to re-run CI
This makes SQLite DB ops properly asynchronous (instead of synchronously completing tasks). This makes them more consistent with postgres and means that any deadlock bugs introduced the next time somebody does .Result will be caught on SQLite too.
Thanks to julian figuring out IDesignTimeDbContextFactory exists in #6327.
All this DbContext configuration and options setup stuff is insane. Microsoft should be absolutely ashamed for coming up with this load of garbage.
* Add admin logging, models, migrations
* Add logging damage changes
* Add Log admin flag, LogFilter, Logs admin menu tab, message
Refactor admin logging API
* Change admin log get method names
* Fix the name again
* Minute amount of reorganization
* Reset Postgres db snapshot
* Reset Sqlite db snapshot
* Make AdminLog have a composite primary key of round, id
* Minute cleanup
* Change admin system to do a type check instead of index check
* Make admin logs use C# 10 interpolated string handlers
* Implement UI on its own window
Custom controls
Searching
Add admin log converters
* Implement limits into the query
* Change logs to be put into an OutputPanel instead for text wrapping
* Add log <-> player m2m relationship back
* UI improvements, make text wrap, add separators
* Remove entity prefix from damaged log
* Add explicit m2m model, fix any players filter
* Add debug command to test bulk adding logs
* Admin logs now just kinda go
* Add histogram for database update time
* Make admin log system update run every 5 seconds
* Add a cap to the log queue and a metric for how many times it has been reached
* Add metric for logs sent in a round
* Make cvars out of admin logs queue send delay and cap
* Merge fixes
* Reset some changes
* Add test for adding and getting a single log
* Add tests for bulk adding logs
* Add test for querying logs
* Add CallerArgumentExpression to LogStringHandler methods and test
* Improve UI, fix SQLite, add searching by round
* Add entities to admin logs
* Move distinct after orderby
* Add migrations
* ef core eat my ass
* Add cvar for client logs batch size
* Sort logs from newest to oldest by default
* Merge fixes
* Reorganize tests and add one for date ordering
* Add note to log types to not change their numeric values
* Add impacts to logs, better UI filtering
* Make log add callable from shared for convenience
* Get current round id directly from game ticker
* Revert namespace change for DamageableSystem
* Engine namespace changes.
* Automated remove redundant using statements.
* Simplified Graphics namespace.
* Apparently the container system stores full type names in the map file.😞 This updates those names.
* API Changes to LocalizationManager.LoadCulture.
* Update submodule to v0.3.2
