Files
subtracks/docs/implementation-plan.md
T

7.1 KiB

Implementation plan

The rewrite is delivered in vertical slices. Each phase leaves the app in a working, testable state.

Phase 0 - Foundation (done)

Nix devshell (JDK, Gradle, Android SDK, nushell, navidrome, gonic), Gradle Kotlin DSL at the repo root, a Compose skeleton, Roborazzi/Robolectric screenshot testing, and the carried-over launcher and notification assets. The app ships as com.subtracks.next.

Phase 1 - Data foundation (done)

Everything needed to mirror a server library locally:

  • MusicSource abstraction with SubsonicClient, SubsonicXml and SubsonicSource.
  • Room 3 schema on bundled SQLite, so the SQLite version is identical on every device.
  • SyncService: stream each entity in batches, upserting per batch and pruning the IDs the server no longer lists; the sync is not atomic.
  • Unit tests (client, XML, sync), native integration tests against navidrome and gonic, and CI (unit, lint, integration) with warm caches.

Phase 2 - UI vertical slice (done)

Goal: a usable app for browsing a server, end to end.

  • Koin DI and DataStore-backed preferences (album sort); the unused app_settings Room table was removed in favour of DataStore.
  • App shell: a root gate that shows the add-server flow until a server exists, then the tabbed library screen matching the Flutter app (icon-only tabs for albums, artists and playlists, with sync and settings actions), plus album and playlist detail screens and a monochrome theme.
  • Source setup: name/address/username/password with token-auth toggle, a connection test, save-and-sync, plus server switching and removal in settings.
  • Library browsing: albums (covers-only grid), artists and playlists paged from Room (PagingSource) and rendered with LazyPagingItems, with Coil cover art served from stable media URLs.
  • Artist detail: tapping an artist opens their albums, newest first, and an album can be opened from there.
  • Roborazzi screenshots for the new screens, rendered on demand for review (not committed).

Not done yet in this slice (deliberately): cover-art tonal colour extraction, long-press menus, editing an existing server, and per-section sort/filter controls (only album sort is wired).

Phase 3 - Playback (in progress)

  • (done) Media3/ExoPlayer playback service, queue and now-playing UI: a MediaSessionService plus a mini player and a Now Playing screen.
  • (done) Media notifications and headset/Bluetooth controls, from the media session.
  • (done) A SQL-backed queue of references (whole playlist / album / song, each with an optional ordinal range) resolved lazily, with only a bounded window in memory and in the player.
  • (done) Streaming bitrate and preferred format from settings (maxBitRate/format, so the server transcodes as configured).
  • (done) A queue view that lists the resolved queue, jumps to a track on tap, removes tracks and reorders them by dragging, with a one-step undo.
  • (done) Adding to the queue from the library (play next / add to queue), backed by a separate, editable up-next list that plays between the current track and the rest of the context.
  • Gapless format preferences.
  • Scrobbling and "now playing" back to the server.
  • Optional download/caching.

Phase 4 - Search, playlists, offline

  • (done) Search: a substring filter on the library queries, scoped to the active tab, with a docked search field and per-tab sort/filter controls. Queries of three or more characters run through per-entity FTS5 trigram tables (external content, trigger-maintained), with the instr scan as the below-three-character fallback.
  • Playlist browsing and editing (create/add/remove), syncing changes back to the server.
  • Offline mode and downloads.

Phase 5 - Release

  • Change applicationId from com.subtracks.next to com.subtracks.
  • Signing, versioning and Fastlane metadata.
  • F-Droid submission. Their buildserver must have Gradle >= 9.6 and SDK 37, which this project targets; do not rely on the Gradle wrapper there (fdroidserver deletes gradlew and uses its own Gradle).

Scale hardening (done)

Follow-up work for very large libraries, playlists and long sessions, grouped so each block could be one PR. All blocks have landed; the two residuals are noted at the end.

  • Library read path. Landed: the sort keys used by album/artist/playlist browsing are indexed (a denormalized albumArtist on albums, NOCASE text, one index per order). Row resolution seeks through a cached keyset cursor (QueueRepository.keyedRows over QueueDao's ...From/...Before/...After queries); the old offset query is kept only as the cold and far-jump fallback.
  • Shuffle. Landed: the permutation is derived from (seed, size) (a seekable Feistel in Shuffle.kt) and positions resolve through the same positional lookup, so shuffle_order and the retained whole-order/whole-id arrays are gone. The context is read-only while shuffled; manual ordering goes through the up-next block.
  • Sync pipeline. Landed: the prune diffs against a memory-bounded set of primitive long id hashes and walks the local ids in keyset pages; fetchRanks and the write-only rank columns are gone; playlists and the album-song fallback are fetched with a bounded, user-configurable concurrency; the page cap is raised to 20k pages and overrun fails with a "Library is too large to sync" error; unchanged rows are not rewritten (upsertChanged's changed-only WHERE).
  • Queue view and edits. Landed: the queue view loads a bounded window (QUEUE_CHUNK * 3) and extends and trims it as you scroll, its ViewModel is scoped to the overlay, adjacent same-ref ranges are compacted after an edit, and resolved entry lengths are cached.
  • Artwork. Landed: artwork_seeds carries a sourceId with ON DELETE CASCADE and an index, is pruned as an LRU by storedAt, and the seed prefetch requests the thumbnail.
  • FTS5 search. Landed: per-entity external-content trigram tables (album_search, artist_search, playlist_search) with Room-generated content-sync triggers, a three-character floor and an instr fallback below it.
  • UI/flow overhead. Landed: only the active tab's paging flow is collected, and the position ticker updates its own positionMs flow instead of the shared PlaybackState.

Two residuals remain, neither blocking: the sync does not skip unchanged albums or playlists on the fetch side (only unchanged writes are skipped), and the now-playing screen still prefetches the original artwork for the next track alongside the thumbnail.

A syncGen column was considered for the sync prune and rejected: it is part of the upsert's update set, so its always-changing value defeats upsertChanged's changed-only WHERE predicate and rewrites every row on every full sync. The prune should instead keep the two-phase diff with a memory-bounded id set (primitive long hashes) or diff a staging seen table in SQL.

Deferred decisions

  • Whether to drop the plaintext-auth fallback or keep it behind an "insecure server" toggle.
  • Robolectric runs at SDK 35 while targetSdk is 37; aligning them changes the on-demand Roborazzi renders (nothing committed to re-record).
  • Whether to extract :core:data and friends once the UI has grown.