Files
Psychotoxical-psysonic/src/store/hotCacheStore.ts
T
Frank Stellmacher 45b9229ceb refactor(queue): thin-state refs as canonical, full Track via resolver (#872)
* refactor(queue): wire queue UI to the track resolver (thin-state phase 3)

cucadmuh's phase-3 steps:
- Selectors (useQueueTracks) read resolver-first: getCachedTrack → queue: Track[]
  fallback (until phase 4), F4 star/rating overrides merged on read.
- QueueList rows source their track from the resolver (queue fallback); rows show
  title/artist/duration only, so no override merge there.
- pendingStarSync star/rating success → invalidateQueueResolver so the cache
  reflects the synced value.
- queueResolverBridge re-seeds on queueIndex change too — the prefetch window
  travels with the playing track.

Additive: queue: Track[] stays canonical and behaviour is unchanged (rows
resolve to the same data). Phase 4 drops queue: Track[] and the fallbacks.

* docs(changelog): queue panel reads through track cache (#860)

* fix(queue): stop a render loop that froze the UI on long queues

A long virtualized queue + a track change could lock the WebView for ~2 min:
- useVirtualizer was handed a fresh `initialRect` object literal every render, so
  it kept re-initializing in a loop. Hoisted it to a stable module constant.
- getCachedTrack did an LRU bump (Map delete+set) during render — a render-time
  side effect. Made it a pure read; recency is set at write time in cacheSet.

* perf(mobile): virtualize the mobile player queue drawer

The mobile now-playing queue drawer rendered the full queue with .map; a
multi-thousand-track queue meant thousands of DOM nodes. Virtualize it with
@tanstack/react-virtual (uniform rows, stable initialRect) so the DOM stays at
O(visible rows), matching the desktop QueuePanel. Active track is centred on open
via scrollToIndex.

* perf(mini): virtualize the mini-player queue list

The mini-player queue rendered the full MiniSyncPayload queue with .map.
Virtualize it against the OverlayScrollArea viewport (stable initialRect) so the
mini window's DOM stays at O(visible rows). Drag-reorder is preserved: rows keep
data-mq-idx alongside the virtualizer's measureElement.

* refactor(queue): add resolveQueueTrack/getQueueTracksView helper (thin-state phase 4)

Render-safe ref→Track view for the phase-4 consumer migration off queue: Track[].
Resolver cache → caller fallback (legacy queue[idx] during dual-write) →
placeholder; ref queue-only flags carried, F4 overrides merged. Pure synchronous
read, no cache mutation (the freeze landmine), so it is safe in render.

* refactor(queue): keep queueItems as the canonical in-memory mirror (thin-state phase 4)

Step 1b: dual-write the thin queueItems ref list at every queue write site
(the 11 mutations, next/radio top-up, playTrack, undo/redo restore, instant-mix,
radio, server-queue init, lucky-mix rollback, and hydrate) so it tracks
queue: Track[] in memory, not only at persist time. Identity-preserving maps
(star/rating overrides) keep the same refs and are intentionally left untouched.

Resolves the restore double-role flagged for 1b: queueItemsIndex is now the
restore-pending sentinel that gates hydrateQueueFromIndex, while queueItems
stays canonical -- rebuilt from the whole queue after a full hydrate instead of
cleared. Normal mutations never set the sentinel, so it only fires on a fresh
cold-start restore, not on later server switches.

No behaviour change; queue: Track[] stays the source consumers read until
phase 3. tsc + full vitest suite (1119 tests) green.

* refactor(queue): mobile queue drawer reads through the track resolver (thin-state phase 4)

Step 2: the mobile now-playing queue drawer resolves each row's track from the
resolver cache (→ queue: Track[] fallback until phase 4), matching the desktop
QueueList wired in the phase-3 commit. Subscribes to the resolver version so
rows re-render as the cache fills. Structure (count, order, keys, the playTrack
arg) still comes from queue: Track[] until it is dropped in the final step.

The mobile drawer was the last queue display surface still reading track
metadata straight off the fat queue. tsc + full vitest suite green.

* refactor(queue): ref-native queue mutations + dual-write bridge (thin-state phase 4)

Step 3a: the 11 queueMutationActions now splice/filter/reorder QueueItemRef[]
(matching by trackId + the ref's queue-only flags) instead of Track[].
`bridgeQueueFromItems` rebuilds the dual-written queue: Track[] from the new
refs by id — purely structural (no resolver/override merge), so behaviour is
byte-identical and playerStore.queue.test.ts stays unchanged green. The working
ref list comes from `itemsOf(state)` (derived from queue: Track[] for now); the
final step swaps that one line to state.queueItems once the fat queue is gone.

enqueue / enqueueAt / enqueueRadio seed the resolver cache with incoming tracks
(seed-before-splice) so they resolve without a network round-trip after the fat
queue is dropped. Adds a DEV-only id-parity guardrail (queue vs queueItems);
dev-runtime only, silent in vitest and prod.

tsc + full vitest suite (1119) green; contract test unchanged.

* refactor(queue): ref-native radio/infinite top-ups (thin-state phase 4)

Step 3b: nextAction's proactive infinite-queue and radio top-ups build the new
queue as QueueItemRef[] and bridge back to queue: Track[] (same as the queue
mutations), and seed the resolver cache with the freshly fetched tracks so they
resolve without a network round-trip after the fat queue is dropped. The radio
top-up keeps its HISTORY_KEEP front-trim, now expressed on refs.

The exhausted-queue refill paths hand their new queue to playTrack, which keeps
its fat-queue handling until the final step (its no-arg case needs the resolver-
derived queue that lands with the queue: Track[] removal). tsc + full vitest
(1119) green; contract test unchanged.

* refactor(queue): undo snapshots store thin refs, not Track[] (thin-state phase 4)

Step 4: QueueUndoSnapshot.queue: Track[] becomes queueItems: QueueItemRef[],
killing the undo "hidden multiplier" — 32 snapshots of a 50k queue now cost
refs, not 32×50k full tracks. applyQueueHistorySnapshot rebuilds the display
queue from the refs via resolveQueueTrack: resolver cache → the live queue by id
(covers tracks the edit didn't remove) → placeholder. currentTrack stays a full
track in the snapshot and is restored to the engine unchanged.

The snapshot refs derive from queue: Track[] for now (so the undo/redo contract
cases, which seed only `queue`, stay green); the final step swaps that to
[...s.queueItems]. tsc + full vitest suite (1119) green.

* perf(mini): cap the mini-player queue snapshot to ±100 around the current track (thin-state phase 4)

Step 5: the mini bridge no longer serializes the full queue over IPC on every
push — a 50k Artist-Radio queue would otherwise re-encode in full on every track
advance. snapshot() sends a window of 100 tracks before/after the playing song;
queueIndex is made slice-relative. The mini component stays unchanged (slice-
relative); jump/reorder/remove control events are translated back to absolute
queue indices via the window offset captured on the last push.

tsc + full vitest suite (1119) green. Mini bridge has no unit tests — needs a
quick mini-player smoke (queue shows ±100, jump/reorder/remove land correctly).

* refactor(queue): make queueItems a required PlayerState field (thin-state phase 4)

Foundation for the final consumer migration off queue: Track[]: queueItems has
been written at every queue write site since phase 1b, so promoting it from
optional to required is a no-op at runtime (tsc confirms zero new errors) and
lets the upcoming reader migrations read state.queueItems without `?? []` noise.

* refactor(queue): migrate structural queue readers off queue: Track[] (thin-state phase 4)

First reader batch toward dropping queue: Track[]: the queue-length selectors
(usePlaybackServerId, usePlaybackCoverArt, useQueuePanelDrag, useMiniQueueDrag)
now read state.queueItems.length, and FullscreenPlayer's next-track cover prefetch
resolves through useQueueTrackAt instead of indexing the fat queue. All behaviour-
identical during dual-write (queueItems is in lockstep with queue). tsc + full
vitest suite (1119) green.

Note: getPlaybackServerId() (playbackServer.ts) deliberately stays on queue for
now — it is called from many partially-mocked test stores, so it migrates with
the final field removal where the seedQueue helper covers those tests.

* refactor(queue): QueuePanel save/share/playlist read queueItems (thin-state phase 4)

The id/length reads (save to playlist, share link, create playlist, empty-queue
guards, next-tracks divider) now read state.queueItems instead of the fat queue.
Behaviour-identical during dual-write; queue: Track[] stays for the rendered
QueueList + auto-scroll until the field is dropped. tsc + full suite (1119) green.

* refactor(queue): drop queue: Track[] — thin queueItems is the only queue (thin-state phase 4)

The store no longer holds the fat queue. `queueItems: QueueItemRef[]` is the sole
canonical queue; full `Track`s resolve on demand via the resolver (index batch →
getSong fallback, bounded LRU cache); only `currentTrack` stays a full Track. At
50k tracks the store holds ~hundreds of resolved tracks + the refs, not 50k Track
objects.

- **Persist:** partialize is refs-only (no windowed slice / PERSIST_QUEUE_HALF).
  A `merge` migrates every historical blob shape → `queueItems` (existing
  `queueItems` → legacy `queueRefs` → pre-ref windowed `queue: Track[]`) and drops
  the obsolete `queue` key, so saved queues survive the upgrade.
- **Restore (decision B):** `hydrateQueueFromIndex` eager-resolves the whole
  ref list into the cache on cold start (index → getSong, so an index-off queue
  still plays), clears the restore sentinel.
- **Resolver bridge:** keeps `[idx-50, idx+200]` warm via `resolveVisibleRange`.
- **Mutations / actions / playback:** operate on refs; the playing track is
  `currentTrack`, the next/neighbour tracks resolve from the cache. Navigation
  (next/previous/row-jump) keeps `queueItems` and only moves the index — no full
  resolve or queue rebuild per track change.
- **Persist tests** cover the three old-blob migrations; `seedQueue` test helper
  replaces the `setState({ queue })` seeds.

tsc + full vitest suite (1115) green. Behaviour-preserving by the test contract;
the gapless track change + cold-start restore + mini cap still want a live smoke
before merge.

* fix(queue): star/rating keeps the queue row resolved instead of blanking to "…" (thin-state)

Rating/starring a queue song flashed the row's title to the "…" placeholder
until the next track change. Root cause: on sync success pendingStarSync called
invalidateQueueResolver, which DROPPED the cached track — and with queue: Track[]
gone there's no fat fallback, so the row resolved to a placeholder until the
resolver bridge re-fetched the window.

Fix: add patchCachedTrack(trackId, patch) and use it on star/rating success to
update the cached entry in place (title kept, synced starred/userRating applied)
instead of dropping it. No placeholder flash, no re-fetch.

tsc + full vitest suite (1115) green.

* fix(player): quota-safe persist so a full localStorage can't kill playback

A very large queue (~50k refs) overflows the localStorage quota; the persist
write then threw QuotaExceededError from inside set(), which aborted playTrack
before audio_play — no audio output at all. Back the player persist with a
quota-safe storage wrapper so a failed write degrades to a no-op instead of
throwing. Restoring the full ref list at that ceiling (vs a windowed cap) is
left as a follow-up.

* polish(player): throttle the quota-skip persist warning to once per key

The quota-safe persist logs a skip on every failed write; on a huge queue that
floods the dev console once per mutation. Warn once per key per quota-exceeded
streak, re-armed when a write to that key next succeeds.

* fix(queue): port new cover-pipeline readers to thin-state

Main's cover pipeline (#870) reads s.queue.length and seeds the player
store with queue: [track] in its tests. Under thin-state, queue: Track[]
no longer exists — the canonical queue is queueItems: QueueItemRef[].
These four files were brought across in the merge but still spoke the
old shape; this commit aligns them with the thin-state contract.

- src/cover/usePlaybackCoverArt: queueLength = queueItems.length
- src/cover/usePlaybackCoverArt.test: seed via toQueueItemRefs
- src/api/coverCache.test: same
- src/hooks/useNowPlayingPrewarm.test: same (two test cases)

* fix(queue): canonicalize thin-state server identity for mixed-server queues

`QueueItemRef.serverId` and `PlayerState.queueServerId` are now written as
the URL-derived index key on every writer path, matching the library index
direction. Mixed-server queues with duplicate `trackId` across servers stay
unambiguous because the resolver cache, persistence, and playback bindings
all share one key shape.

- new `canonicalQueueServerKey()` helper (idempotent UUID-or-key normalizer)
- `toQueueItemRefs`, `bindQueueServerForPlayback`, `seedQueueResolver`, and
  `hydrateQueueFromIndex` emit canonical keys
- `getCachedTrack` falls back to the canonical lookup so refs persisted in
  the legacy UUID shape still resolve through the migration window
- persist `merge` rewrites `queueServerId` and every ref `serverId` on
  rehydrate, so the live store never holds mixed shapes
- `removeServer` compares against the resolved id so a profile delete still
  clears the matching queue binding
- the two `playbackServer.test.ts` asserts that hard-coded the UUID shape
  are updated to the canonical key (existing reader-tolerance is unchanged)

* fix(queue-undo): bind snapshot prepend to snapshot-canonical server identity

When `applyQueueHistorySnapshot` has to prepend the still-playing track
(the snapshot's queue does not contain it), the new ref must follow the
snapshot's playback server, not the live `queueServerId`. A server switch
racing the undo would otherwise stamp the prepended ref with the new
server, mis-resolving the playing track on the very next render.

- `QueueUndoSnapshot` now carries `queueServerId` (captured by
  `queueUndoSnapshotFromState`); older in-memory entries fall back through
  the snapshot's own refs and finally the live store value
- the prepend in `applyQueueHistorySnapshot` plus the post-restore
  `seedQueueResolver` both source the server identity from this snapshot
  context, run through `canonicalQueueServerKey` so cache bucket and ref
  shape stay in lockstep

* test(queue): regression cluster for mixed-server queues with duplicate trackId

Covers the four invariants the thin-state review called out:

- resolver correctness: same `trackId` on two servers maps to two distinct
  cache entries via canonical keys, and legacy UUID-shaped refs still read
  the same entries through the compat lookup path
- restore/hydrate: persist `merge` forward-migrates UUID-form blobs in
  three shapes (canonical `queueItems`, legacy `queueRefs`, mixed-server
  `queueItems`) to canonical keys
- undo snapshot application: prepended ref follows the snapshot's playback
  server even when the live queue has been rebound to a different one,
  with fallback to snapshot refs and live state for legacy entries
- queue sync id emission: `flushPlayQueuePosition` -> `savePlayQueue`
  passes plain track ids and the playback server out of band, no per-ref
  `serverId` ever leaks into the request body

Also asserts the write helpers (`toQueueItemRefs`,
`bindQueueServerForPlayback`) emit canonical keys directly.

* perf(queue-header): coalesce resolver burst updates and aggregate in one pass

`QueueHeader` recomputed total and remaining queue durations on every
resolver cache version bump via two separate full-queue reduces. A mass
resolve burst (queue restore, prefetch window slide) bumps the version
dozens of times in one frame, and very long queues turned that into
visible main-thread stutter.

- one pass: a single for-loop produces both total and future-tracks
  duration; a 50k-track queue costs one walk per recompute, not two
- `useDeferredValue(version)` coalesces the burst into a single
  low-priority commit so the cache version is only sampled once per
  React frame instead of once per cache write

* fix(queue): use stable artist seed for radio top-up

The proactive radio top-up in `runNext` seeded `getSimilarSongs2` and
`getTopSongs` from `resolveQueueTrack(nextRef)` metadata. When the next
ref is still cold in the resolver cache, the placeholder track has empty
artist fields, and the top-up would fire `getSimilarSongs2('')` -- silently
returning nothing and leaving the queue dry just before the radio rail
would have refilled.

- prefer the just-played `currentTrack` (always fully resolved in the
  player store) and the stored radio seed artist id
- fall back to the next-track metadata only when those are missing
- skip the top-up entirely when no stable seed is available, instead of
  emitting a non-deterministic empty request

* docs(changelog): queue mixed-server routing and quota-safe persist (#872)
2026-05-27 00:10:34 +02:00

275 lines
8.8 KiB
TypeScript

import type { QueueItemRef } from './playerStoreTypes';
import { create } from 'zustand';
import { persist, createJSONStorage } from 'zustand/middleware';
import { invoke } from '@tauri-apps/api/core';
import { isHotCachePreviousTrackUnderGrace } from '../utils/cache/hotCacheGate';
import { emitAnalysisStorageChanged } from './analysisSync';
import { useAuthStore } from './authStore';
/** How many queue slots after the current index are eviction-protected (1 = current + next only). */
export const HOT_CACHE_PROTECT_AFTER_CURRENT = 1;
export interface HotCacheEntry {
localPath: string;
sizeBytes: number;
cachedAt: number;
/** Last time this file was started as the current track (eviction tie-break: newer = keep longer). */
lastPlayedAt?: number;
}
interface HotCacheState {
/** Persisted map `${serverId}:${trackId}` → file meta */
entries: Record<string, HotCacheEntry>;
getLocalUrl: (trackId: string, serverId: string) => string | null;
setEntry: (
trackId: string,
serverId: string,
localPath: string,
sizeBytes: number,
debugSource?: string,
) => void;
/** Bump LRU when the user actually plays this track (if it is in the hot cache). */
touchPlayed: (trackId: string, serverId: string) => void;
removeEntry: (trackId: string, serverId: string) => void;
totalBytes: () => number;
/** Evict until total size ≤ maxBytes. Protects current + next (+ grace for last «previous» track).
* Thin-state: only track ids / positions matter here, so it takes the canonical refs. */
evictToFit: (
queue: QueueItemRef[],
queueIndex: number,
maxBytes: number,
activeServerId: string,
hotCacheCustomDir: string | null,
) => Promise<void>;
clearAllDisk: (customDir: string | null) => Promise<void>;
}
function entryKey(serverId: string, trackId: string): string {
return `${serverId}:${trackId}`;
}
function parseKey(key: string): { serverId: string; trackId: string } | null {
const i = key.indexOf(':');
if (i <= 0) return null;
return { serverId: key.slice(0, i), trackId: key.slice(i + 1) };
}
function lruStamp(meta: HotCacheEntry | undefined): number {
if (!meta) return 0;
return meta.lastPlayedAt ?? meta.cachedAt ?? 0;
}
function evictionReasonForTier(tier: number): string {
const labels: Record<number, string> = {
0: 'inactive-server',
1: 'not-in-queue',
2: 'ahead-of-protected-window',
3: 'behind-current-in-queue',
};
return labels[tier] ?? `tier-${tier}`;
}
/** Settings → Logging → Debug, same as `emitNormalizationDebug` / lucky-mix. */
function hotCacheFrontendDebug(payload: Record<string, unknown>): void {
if (useAuthStore.getState().loggingMode !== 'debug') return;
void invoke('frontend_debug_log', {
scope: 'hot-cache',
message: JSON.stringify(payload),
}).catch(() => {});
}
export const useHotCacheStore = create<HotCacheState>()(
persist(
(set, get) => ({
entries: {},
getLocalUrl: (trackId, serverId) => {
const e = get().entries[entryKey(serverId, trackId)];
if (!e?.localPath) return null;
return `psysonic-local://${e.localPath}`;
},
setEntry: (trackId, serverId, localPath, sizeBytes, debugSource) => {
const now = Date.now();
set(s => ({
entries: {
...s.entries,
[entryKey(serverId, trackId)]: {
localPath,
sizeBytes,
cachedAt: now,
lastPlayedAt: now,
},
},
}));
hotCacheFrontendDebug({
event: 'index-add',
trackId,
serverId,
sizeBytes,
source: debugSource ?? 'unknown',
});
},
touchPlayed: (trackId, serverId) => {
const k = entryKey(serverId, trackId);
set(s => {
const e = s.entries[k];
if (!e) return s;
return {
entries: {
...s.entries,
[k]: { ...e, lastPlayedAt: Date.now() },
},
};
});
},
removeEntry: (trackId, serverId) => {
set(s => {
const next = { ...s.entries };
delete next[entryKey(serverId, trackId)];
return { entries: next };
});
hotCacheFrontendDebug({
event: 'index-remove',
trackId,
serverId,
reason: 'explicit-removeEntry',
});
emitAnalysisStorageChanged({ trackId, reason: 'hotcache-delete' });
},
totalBytes: () =>
Object.values(get().entries).reduce((acc, e) => acc + (e.sizeBytes || 0), 0),
evictToFit: async (queue, queueIndex, maxBytes, activeServerId, hotCacheCustomDir) => {
if (maxBytes <= 0) return;
const protectLo = Math.max(0, queueIndex);
const protectHi = Math.min(queue.length - 1, queueIndex + HOT_CACHE_PROTECT_AFTER_CURRENT);
const protectedIds = new Set<string>();
for (let i = protectLo; i <= protectHi; i++) {
protectedIds.add(queue[i].trackId);
}
const indexOfInQueue = (trackId: string): number | null => {
const idx = queue.findIndex(r => r.trackId === trackId);
return idx >= 0 ? idx : null;
};
let entries = { ...get().entries };
let sum = Object.values(entries).reduce((a, e) => a + (e.sizeBytes || 0), 0);
if (sum <= maxBytes) return;
const keys = Object.keys(entries);
type Cand = { key: string; tier: number; primary: number; lru: number };
const cands: Cand[] = [];
for (const key of keys) {
const parsed = parseKey(key);
if (!parsed) continue;
const { serverId, trackId } = parsed;
if (protectedIds.has(trackId) && serverId === activeServerId) continue;
if (isHotCachePreviousTrackUnderGrace(trackId, serverId)) continue;
const meta = entries[key];
const lru = lruStamp(meta);
if (serverId !== activeServerId) {
cands.push({ key, tier: 0, primary: 0, lru });
continue;
}
const qIdx = indexOfInQueue(trackId);
if (qIdx === null) {
cands.push({ key, tier: 1, primary: 0, lru });
} else if (qIdx > protectHi) {
cands.push({ key, tier: 2, primary: -qIdx, lru });
} else if (qIdx < protectLo) {
cands.push({ key, tier: 3, primary: qIdx, lru });
}
}
cands.sort((a, b) => {
if (a.tier !== b.tier) return a.tier - b.tier;
if (a.primary !== b.primary) return a.primary - b.primary;
return a.lru - b.lru;
});
if (cands.length === 0) {
hotCacheFrontendDebug({
event: 'evict-no-candidates',
sumBytes: sum,
maxBytes,
queueIndex,
entryKeys: keys.length,
reason: 'all-protected-or-grace-or-parse-fail',
});
return;
}
hotCacheFrontendDebug({
event: 'evict-start',
sumBytes: sum,
maxBytes,
queueIndex,
protectLo,
protectHi,
candidateCount: cands.length,
});
for (const cand of cands) {
if (sum <= maxBytes) break;
const { key, tier } = cand;
const meta = entries[key];
if (!meta) continue;
const parsed = parseKey(key);
if (!parsed) continue;
await invoke('delete_hot_cache_track', {
localPath: meta.localPath,
customDir: hotCacheCustomDir || null,
}).catch((e: unknown) => {
hotCacheFrontendDebug({
event: 'evict-disk-delete-failed',
trackId: parsed.trackId,
serverId: parsed.serverId,
error: String(e),
});
});
hotCacheFrontendDebug({
event: 'evict-remove',
trackId: parsed.trackId,
serverId: parsed.serverId,
reason: `budget:${evictionReasonForTier(tier)}`,
tier,
bytes: meta.sizeBytes,
sumBytesAfter: sum - (meta.sizeBytes || 0),
maxBytes,
});
sum -= meta.sizeBytes || 0;
delete entries[key];
emitAnalysisStorageChanged({ trackId: parsed.trackId, reason: 'hotcache-delete' });
}
set({ entries });
},
clearAllDisk: async (customDir: string | null) => {
hotCacheFrontendDebug({
event: 'purge-all',
customDir: customDir && customDir.length > 0 ? '(custom)' : 'default',
});
await invoke('purge_hot_cache', { customDir: customDir || null }).catch(() => {});
set({ entries: {} });
emitAnalysisStorageChanged({ trackId: null, reason: 'hotcache-purge' });
},
}),
{
name: 'psysonic-hot-cache',
storage: createJSONStorage(() => localStorage),
partialize: s => ({ entries: s.entries }),
},
),
);