mcpField
D
datastore

datastore

github.com/rcosteira79Development165138 installsUpdated 6w ago
Repository

About datastore

A Jetpack DataStore skill for Android and Kotlin Multiplatform developers choosing and implementing reactive storage. It compares Preferences and typed storage with Room, and highlights error recovery and platform-specific file-path pitfalls.

Key features

  1. 01Preferences and typed storage guidance
  2. 02DataStore versus Room selection
  3. 03IOException-specific Flow handling
  4. 04Typed-store corruption recovery
  5. 05Platform-specific path factory

Use cases

  1. 01Choosing storage for settings and structured data
  2. 02Handling DataStore read errors safely
  3. 03Recovering from corrupted typed-store files
  4. 04Creating stores across Android, iOS, and JVM

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/rcosteira79/android-skills --skill datastore

After installing, type/datastorein Claude Code to confirm it loaded.

Details

From the project README

Jetpack DataStore for Android and KMP

Reactive coroutine-based key-value / typed storage; the same androidx.datastore:datastore-preferences-core runs on Android, iOS, and JVM — only the file-path producer is platform-specific. (Web support is 1.3.0-alpha only, and it's sessionStorage/OPFS-backed there, not file-path-based.) This reference covers the storage-choice call plus the two error/path traps, not the basics of Preferences keys, edit { }, or Serializer<T>. Adapted from Meet-Miyani/compose-skill; MIT. Related: android-skills:android-data-layer, android-skills:kmp-boundaries, android-skills:kotlin-flows.

Preferences vs Typed vs Room

NeedStorage
Key-value flags (theme, locale, onboarding done)Preferences DataStore
One typed object, many related fields, schema evolutionTyped DataStore + Serializer<T>
Relational data, indexes, WHERE / JOIN, >100 entries, pagingRoom
Payloads above ~50KB per writeRoom / filesystem — DataStore rewrites the whole file on every edit

Rule of thumb: if a WHERE clause would be useful, use Room.

The two traps

.catch must match IOException specifically — never a broad catch. A bare catch { emit(emptyPreferences()) } swallows CancellationException (breaking structured concurrency) and hides serializer/corruption errors behind a silent empty state.

val settings: Flow<UserSettings> = dataStore.data
    .catch { e -> if (e is IOException) emit(emptyPreferences()) else throw e }
    .map { p -> UserSettings(p[Keys.DARK_MODE] ?: false, p[Keys.LOCALE] ?: "en") }

Typed-DataStore corruption recovery triggers on CorruptionException, NOT IOException. The serializer's readFrom must wrap parse failures in CorruptionException, and the store needs a ReplaceFileCorruptionHandler — without it, one corrupt file makes every read fail permanently.

object AppSettingsSerializer : Serializer<AppSettings> {
    override val defaultValue = AppSettings()
    override suspend fun readFrom(input: InputStream): AppSettings =
        try { Json.decodeFromString(input.readBytes().decodeToString()) }
        catch (e: SerializationException) { throw CorruptionException("Cannot read AppSettings", e) }
    override suspend fun writeTo(t: AppSettings, output: OutputStream) =
        output.write(Json.encodeToString(t).encodeToByteArray())
}
val store = DataStoreFactory.create(
    serializer = AppSettingsSerializer,
    corruptionHandler = ReplaceFileCorruptionHandler { AppSettings() },
    produceFile = { File(context.filesDir, "app_settings.json") },
)

KMP factory — per-platform path (not the temp dir)

// commonMain
fun createPreferencesDataStore(producePath: () -> String): DataStore<Preferences> =
    PreferenceDataStoreFactory.createWithPath(produceFile = { producePath().toPath() })
// android: context.filesDir.resolve(PREFS_FILE).absolutePath
// ios:     NSDocumentDirectory via NSFileManager
// jvm:     File(System.getProperty("user.home"), ".myapp") — NOT java.io.tmpdir (the OS may wipe it on reboot)

(In a composable, never runBlocking on DataStore — it parks the main thread on disk I/O, risking an ANR, and re-runs every recomposition. Expose a StateFlow and collect with collectAsStateWithLifecycle.)

Comments

Markdown supported