Hilt is an annotation processor built on top of Dagger which allows to
remove all the Android dependency injection boilerplate code (currently
around 900 lines) by writing it for us.
Hilt can use KSP instead of Kapt so Kapt can be completely removed from
the project. Kapt is slow, deprecated and has a few compatibility
issues. Removing Kapt will improve build times since no Java stubs have
to be generated for Kotlin classes anymore (Note that KSP also processes
annotations in Java classes so it can completely replace Kapt).
- Remove all modules related to manual dependency injection
configuration.
- Rename `AppModule` to `StorageModule` since it now only contains
configuration to retrieve the DataBase and SharedPreferences.
- Annotate all entry points (Activities, Fragments, BroadcastReceivers
and Services) with `@AndroidEntryPoint`.
- Annotate all injected ViewModels with `@HiltViewModel` and replace the
custom ViewModel Factory with the default one (which integrates with the
one generated by Hilt).
- Add a public field to allow overriding the default
ViewModelProvider.Factory in `BaseActivity` in tests.
- Annotate tested Activities with `@OptionalInject` since Activity tests
currently rely on the Activities not being injected automatically.
- Annotate injected `Context` arguments with `@ApplicationContext`. Hilt
provides the `Context` binding automatically but requires to specify if
the Application or Activity Context is wanted.
- Add WorkManager Hilt integration so all Workers are injected by Hilt
automatically using `HiltWorkerFactory`.
- Lazily initialize WorkManager in `TuskyApplication`.
- Remove Kapt and Kapt workarounds.
- ~~Remove toolchain configuration for Java 21. Toolchains force the
Java bytecode to match the JDK version used to build the project, and
apparently Hilt doesn't run inside the toolchain so cannot process the
source code if the JDK version of the toolchain is higher than the JDK
used to run Gradle. [And configuring a toolchain for an older Java
version causes other
issues](https://jakewharton.com/gradle-toolchains-are-rarely-a-good-idea/).
**Removing toolchains configuration doesn't prevent the project from
being built using JDK 21** or more recent versions but allows to build
the project using older JDKs as well.~~
Added a fix to allow Hilt to properly use the JDK toolchain.
- ~~Set the Java and Kotlin bytecode target to Java 17. The standard
bytecode target for Android projects is usually Java 8 or 11 (any higher
version doesn't provide any benefit but may cause compatibility issues).
However, since the app currently uses a library built against Java 17
bytecode (`networkresult-calladapter`), it needs to target at least Java
17 bytecode as well.~~
- Update the Dagger 2 URL in the licenses screen. Hilt is part of Dagger
2 so the label wasn't changed.
**! ! Warning**: Do not merge before testing every API call and database
read involving JSON !
**Gson** is obsolete and has been superseded by **Moshi**. But more
importantly, parsing Kotlin objects using Gson is _dangerous_ because
Gson uses Java serialization and is **not Kotlin-aware**. This has two
main consequences:
- Fields of non-null types may end up null at runtime. Parsing will
succeed, but the code may crash later with a `NullPointerException` when
trying to access a field member;
- Default values of constructor parameters are always ignored. When
absent, reference types will be null, booleans will be false and
integers will be zero.
On the other hand, Kotlin-aware parsers like **Moshi** or **Kotlin
Serialization** will validate at parsing time that all received fields
comply with the Kotlin contract and avoid errors at runtime, making apps
more stable and schema mismatches easier to detect (as long as logs are
accessible):
- Receiving a null value for a non-null type will generate a parsing
error;
- Optional types are declared explicitly by adding a default value. **A
missing value with no default value declaration will generate a parsing
error.**
Migrating the entity declarations from Gson to Moshi will make the code
more robust but is not an easy task because of the semantic differences.
With Gson, both nullable and optional fields are represented with a null
value. After converting to Moshi, some nullable entities can become
non-null with a default value (if they are optional and not nullable),
others can stay nullable with no default value (if they are mandatory
and nullable), and others can become **nullable with a default value of
null** (if they are optional _or_ nullable _or_ both). That third option
is the safest bet when it's not clear if a field is optional or not,
except for lists which can usually be declared as non-null with a
default value of an empty list (I have yet to see a nullable array type
in the Mastodon API).
Fields that are currently declared as non-null present another
challenge. In theory, they should remain as-is and everything will work
fine. In practice, **because Gson is not aware of nullable types at
all**, it's possible that some non-null fields currently hold a null
value in some cases but the app does not report any error because the
field is not accessed by Kotlin code in that scenario. After migrating
to Moshi however, parsing such a field will now fail early if a null
value or no value is received.
These fields will have to be identified by heavily testing the app and
looking for parsing errors (`JsonDataException`) and/or by going through
the Mastodon documentation. A default value needs to be added for
missing optional fields, and their type could optionally be changed to
nullable, depending on the case.
Gson is also currently used to serialize and deserialize objects to and
from the local database, which is also challenging because backwards
compatibility needs to be preserved. Fortunately, by default Gson omits
writing null fields, so a field of type `List<T>?` could be replaced
with a field of type `List<T>` with a default value of `emptyList()` and
reading back the old data should still work. However, nullable lists
that are written directly (not as a field of another object) will still
be serialized to JSON as `"null"` so the deserializing code must still
be handling null properly.
Finally, changing the database schema is out of scope for this pull
request, so database entities that also happen to be serialized with
Gson will keep their original types even if they could be made non-null
as an improvement.
In the end this is all for the best, because the app will be more
reliable and errors will be easier to detect by showing up earlier with
a clear error message. Not to mention the performance benefits of using
Moshi compared to Gson.
- Replace Gson reflection with Moshi Kotlin codegen to generate all
parsers at compile time.
- Replace custom `Rfc3339DateJsonAdapter` with the one provided by
moshi-adapters.
- Replace custom `JsonDeserializer` classes for Enum types with
`EnumJsonAdapter.create(T).withUnknownFallback()` from moshi-adapters to
support fallback values.
- Replace `GuardedBooleanAdapter` with the more generic `GuardedAdapter`
which works with any type. Any nullable field may now be annotated with
`@Guarded`.
- Remove Proguard rules related to Json entities. Each Json entity needs
to be annotated with `@JsonClass` with no exception, and adding this
annotation will ensure that R8/Proguard will handle the entities
properly.
- Replace some nullable Boolean fields with non-null Boolean fields with
a default value where possible.
- Replace some nullable list fields with non-null list fields with a
default value of `emptyList()` where possible.
- Update `TimelineDao` to perform all Json conversions internally using
`Converters` so no Gson or Moshi instance has to be passed to its
methods.
- ~~Create a custom `DraftAttachmentJsonAdapter` to serialize and
deserialize `DraftAttachment` which is a special entity that supports
more than one json name per field. A custom adapter is necessary because
there is not direct equivalent of `@SerializedName(alternate = [...])`
in Moshi.~~ Remove alternate names for some `DraftAttachment` fields
which were used as a workaround to deserialize local data in 2-years old
builds of Tusky.
- Update tests to make them work with Moshi.
- Simplify a few `equals()` implementations.
- Change a few functions to `val`s
- Turn `NetworkModule` into an `object` (since it contains no abstract
methods).
Please test the app thoroughly before merging. There may be some fields
currently declared as mandatory that are actually optional.
There are some new rules, I think they mostly make sense, except for the
max line length which I had to disable because we are over it in a lot
of places.
---------
Co-authored-by: Goooler <wangzongler@gmail.com>
* update gradle, kotlin and other dependencies
* fix new warnings
* remove unused import
* update Proguard rules
* add explicit dependency on Gson to get the newest version
* remove debug flag from proguard rules again
* fix typo
Christophe Beyls
Konrad Pozniak
Ivan Kupalov
Mélanie Chauvel (ariasuni)