ForeignKey constraints can be invalidated in the middle of a transaction
even if a later statement in the same transaction will make them valid
again. This seems to be causing production crashes.
Defer foreign key constraint checks until the end of the transaction to
prevent this.
Clean up the notification handling code and fix a lot of bugs, hopefully
without introducing new ones in the process.
Specific bugs discovered and fixed:
- The code that tried to sync notification filtering state between the
server and Pachli could fail, leaving things in an inconsistent state,
resulting in dropped notifications. Remove that code, do filtering
client-side.
- Logging out of an account would disable push notifications for all
accounts.
- If any account did not support push notifications then push
notifications were disabled for all accounts.
- If any account did not support push notifications the user was
prompted to log out of all accounts. Drop that entirely.
- The UnifiedPush library could get to a state where configuring the
notification mechanism would silently fail,
The preferences UI now has a section for notifications, showing:
- The Unified Push distributor in use (if any)
- A mechanism to change the distributor
- Per-account configuration and notification fetch details
- Battery optimisation state
General changes:
- Update to UnifiedPush library 2.4.0.
- NotificationFetcher.fetchAndShow() can now fetch a single account's
notifications, or all accounts, depending on data passed to the worker.
- Use ApiResult for `push/subscription` responses.
- Drop the "needs migration" terminology for the more specific "has push
scope", to make it clear what the issue with the account is.
`TabData` recorded the type of the timeline the user had added to a tab.
`TimelineKind` is another type that records general information about
configured timelines, with identical properties.
There's no need for both, so remove `TabData` and use `TimelineKind` in
its place.
`TimelineKind` is itself mis-named; it's not just the timeline's kind
but also holds data necessary to display that timeline (e.g., the list
ID if it's a `.UserList`, or the hashtags if it's a `.Hashtags`) so
rename to `Timeline` to better reflect its usage. Move it to a new
`core.model` module.
The previous code forgot to close the DB after TimelineDaoTest was run,
so a warning message was displayed when the test was run locally.
Close the database using the `@After` annotation.
Fixes#511
Once desugaring is enabled it needs to be enabled for up/down the
dependency chain, so enable it in the shared configuration defined by
the build convention code.
Highlighted a failing test that wasn't being run, so fix that too.
Many servers that claim to be Mastodon-API compatible are not, as
evidenced by the content they include in the responses to
`/api/v1/instance` and `/api/v2/instance` requests.
Work around the worst of the breakage by providing defaults or marking
some fields as nullable (with a default null).
Bugs have been reported against the relevant projects.
Previous code expected all incoming enums values to map directly to
Kotlin enum constants.
This is a problem for servers with additional features -- e.g.,
"reaction" as a notification type.
Fix this with a new Moshi adapter that will set the incoming value to a
given constant if it's not recognised.
Apply this to the enum constants in core.network to ensure they are
handled.
Clean up enum handling in Converters.kt, ComposeViewModel.kt, and
Status.kt by using the existing `.ordinal` property and some extension
functions for idiomatic code.
Fixes#461
Some users report that Pachli is not retrieving/displaying notifications
in a timely fashion.
To assist in diagnosing these errors, provide an additional set of tabs
on the "About" screen that contain information about how Pachli is
fetching notifications, and if not, why not.
Allow the user to save notification related logs and other details to a
file that can be attached to an e-mail or bug report.
Recording data:
- Provide a `NotificationConfig` singleton with properties to record
different aspects of the notification configuration. Update these
properties as different notification actions occur.
- Store logs in a `LogEntryEntity` table. Log events of interest with a
new `Timber` `LogEntryTree` that is planted in all cases.
- Add `PruneLogEntryEntityWorker` to trim saved logs to the last 48
hours.
Display:
- Add a `NotificationFragment` to `AboutActivity`. It hosts two other
fragments in tabs to show details from `NotificationConfig` and the
relevant logs, as well as controls for interacting with them.
Bug fixes:
- Filter out notifications with a null tag when processing active
notifications, prevents an NPE crash
Other changes:
- Log more details when errors occur so the bug reports are more helpful
Moshi is faster to decode JSON at runtime, is actively maintained, has a
smaller memory and method footprint, and a slightly smaller APK size.
Moshi also correctly creates default constructor arguments instead of
leaving them null, which was a source of `NullPointerExceptions` when
using Gson.
The conversion broadly consisted of:
- Adding `@JsonClass(generateAdapter = true)` to data classes that
marshall to/from JSON.
- Replacing `@SerializedName(value = ...)` with `@Json(name = ...)`.
- Replacing Gson instances with Moshi in Retrofit, Hilt, and tests.
- Using Moshi adapters to marshall to/from JSON instead of Gson `toJson`
/ `fromJson`.
- Deleting `Rfc3339DateJsonAdapter` and related code, and using the
equivalent adapter bundled with Moshi.
- Rewriting `GuardedBooleanAdapter` as a more generic `GuardedAdapter`.
- Deleting unused ProGuard rules; Moshi generates adapters using code
generation, not runtime reflection.
The conversion surfaced some bugs which have been fixed.
- Not all audio attachments have attachment size metadata. Don't show
the attachment preview if the metadata is missing.
- Some `throwable` were not being logged correctly.
- The wrong type was being used when parsing the response when sending a
scheduled status.
- Exceptions other than `HttpException` or `IoException` would also
cause a status to be resent. If there's a JSON error parsing a response
the status would be repeatedly sent.
- In tests strings containing error responses were not valid JSON.
- Workaround Mastodon a bug and ensure `filter.keywords` is populated,
https://github.com/mastodon/mastodon/issues/29142
Previously when the user interacted with a status the operation (reblog,
favourite, etc) travels through multiple layers of code, carrying with
it the position of the item in the list that the user operated on.
At some point the status is retrieved from the list using its position
so that the correct status ID can be used in the network operation.
If this happens while the list is also refreshing there's a possible
race condition, and the original status' position may have changed in
the list. Looking up the status by position to determine which status to
perform the action on may cause the action to happen on the wrong
status.
Fix this by passing the status' viewdata to any actions instead of its
position. This includes all the information necessary to make the API
call, so there is no chance of a race.
This is quite an involved change because there are three types of
viewdata:
- `StatusViewData`, used for regular timelines
- `NotificationViewData`, used for notifications, may wrap a status that
can be operated on
- `ConversationViewData`, used for conversations, does wrap a status
The previous code treated them all differently, which is probably why it
operated by position instead of type.
The high level fix is to:
1. Create an interface, `IStatusViewData`, that contains the data
exposed by any viewdata that contains a status.
2. Implement the interface in `StatusViewData`, `NotificationViewData`,
and `ConversationViewData`.
3. Change the code that operates on viewdata (`SFragment`,
`StatusActionListener`, etc) to be generic over anything that implements
`IStatusViewData`.
4. Change the code that handles actions to pass the viewdata instead of
the position.
Fixes#370
The previous code was serialising the `TabKind` enum without the `_`, so
when it was converted back to the enum name (which has a `_`) it failed
and immediately crashed.
Fixes#306
If a status was part of a thread, and it was not the "detailed" status,
and it had been translated, then the view data was marked as "show the
translation". But the translation was not loaded, so the status content
appeared as empty.
Fix that by loading the translated content of all statuses in the thead
and ensure that the translated content is rendered.
Throw an `IllegalStateException` in debug builds to catch any future
occurrences of this.
Fixes#281
The previous code generally started an activity by having the activity
provide a method in a companion object that returns the relevant intent,
possibly taking additional parameters that will be included in the
intent as extras.
E.g., if A wants to start B, B provides the method that returns the
intent that starts B.
This introduces a dependency between A and B.
This is worse if B also wants to start A.
For example, if A is `StatusListActivity` and B is`ViewThreadActivity`.
The user might click a status in `StatusListActivity` to view the
thread, starting `ViewThreadActivity`. But from the thread they might
click a hashtag to view the list of statuses with that hashtag. Now
`StatusListActivity` and `ViewThreadActivity` have a circular
dependency.
Even if that doesn't happen the dependency means that any changes to B
will trigger a rebuild of A, even if the changes to B are not relevant.
Break this dependency by adding a `:core:navigation` module with an
`app.pachli.core.navigation` package that contains `Intent` subclasses
that should be used instead. The `quadrant` plugin is used to generate
constants that can be used to launch activities by name instead of by
class, breaking the dependency chain.
The plugin uses the `Activity` names from the manifest, so when an
activity is moved in the future the constant will automatically update
to reflect the new package name.
If the activity's intent requires specific extras those are passed via
the constructor, with companion object methods to extract them from the
intent.
Using the intent classes from this package is enforced by a lint
`IntentDetector` which will warn if any intents are created using a
class literal.
See #291
The existing code base is a single monolithic module. This is relatively
simple to configure, but many of the tasks to compile the module and
produce the final app have to run in series.
This is unnecessarily slow.
This change starts to split the code in to multiple modules, which are:
- :core:account - AccountManager, to break a dependency cycle
- :core:common - low level types or utilities used in many other modules
- :core:database - database types, DAOs, and DI infrastructure
- :core:network - network types, API definitions, and DI infrastructure
- :core:preferences - shared preferences definitions and DI
infrastructure
- :core:testing - fakes and rules used across different modules
Benchmarking with gradle-profiler shows a ~ 17% reduction in incremental
build times after an ABI change. That will improve further as more code
is moved to modules.
The rough mechanics of the changes are:
- Create the modules, and move existing files in to them. This causes a
lot of churn in import arguments.
- Convert build.gradle files to build.gradle.kts
- Separate out the data required to display a tab (`TabViewData`) from
the data required to configure a tab (`TabData`) to avoid circular
dependencies.
- Abstract the repeated build logic shared between the modules in to
a set of plugins under `build-logic/`, to simplify configuration of
the application and library builds.
- Be explicit that some nullable types are non-null at time of use.
Nullable properties in types imported from modules generally can't be
smart cast to non-null. There's a detailed discussion of why this
restriction exists at
https://discuss.kotlinlang.org/t/what-is-the-reason-behind-smart-cast-being-impossible-to-perform-when-referenced-class-is-in-another-module/2201.
The changes highlight design problems with the current code, including:
- The main application code is too tightly coupled to the network types
- Too many values are declared unnecessarily nullable
- Dependency cycles between code that make modularisation difficult
Future changes will add more modules.
See #291.
Nik Clayton
sanao