Kha-kis/arr-dashboard v2.18.4

v2.18.4 — Seerr profile override on approval + backup OOM fix

on GitHub

[2.18.4] - 2026-05-07

Added

• Admin quality-profile override on Seerr request approval (closes #434). A new "Options" button next to each pending request opens a profile-picker dialog where admins can pick a non-default quality profile, root folder, or server before approval — useful when a specific requester's content should land in a different profile than the server default (e.g. a "trash shows" profile for a user who watches everything). Jellyseerr/Overseerr stamps the override onto the request before handing off to Sonarr/Radarr. The one-click Approve path is unchanged — overrides are opt-in. Audit log records { overridden: true, profileId, rootFolder, ... } for accountability. Live-tested end-to-end against a real Jellyseerr v3.2.0 + Sonarr instance, which surfaced three real-data bugs that mocked unit tests didn't catch: serverId: 0 is valid (Jellyseerr is 0-indexed), languageProfiles: null is valid (was strict .optional()), and TV PUT requires seasons (Jellyseerr 500s otherwise). Lazy-loaded modal, dropdowns reuse the existing /request-options endpoint (#436).

Fixed

• Out-of-memory crash during scheduled backup (closes #427 follow-up). v2.18.3 fixed Readarr/Lidarr library-sync OOM, but a separate report showed scheduled backups still hitting the 768 MB heap cap. Root cause: exportDatabase ran 21 findMany() calls in parallel and the encryption chain (JSON.stringify → buffer → buffer → base64 → stringify) held ~5–6× row data simultaneously. Fixes:
  • Non-manual backups (scheduled + auto-update) now skip operational history tables (huntLog, huntSearchHistory, trashSyncHistory, templateDeploymentHistory) by default — these grow unbounded over time and are not needed to restore working configuration. Manual UI-triggered backups preserve full history. An info log records when exclusion fires so operators can correlate "empty huntLog after restore" to backup type.
  • When operational history is included, each table is capped to the most recent 1000 rows (ordered by timestamp DESC). A count() pre-check logs a warn whenever rows are dropped so truncation is never silent.
  • Tables are fetched sequentially instead of in parallel, and the JSON plaintext is built inside a block scope so V8 can reclaim the row data + JSON string before the base64 ciphertext is allocated, halving peak heap during encryption.
  • The size estimator now samples up to three rows per table (first / middle / last) and uses the max stringified size — single-row sampling silently underestimated when the first row was sparse (e.g., a huntLog with details: null). Non-serializable rows (undefined) no longer crash the estimator.
  • Envelope JSON is no longer pretty-printed (saves ~33% of envelope-stringify peak).
• Heap pressure during library cleanup runs. prefetchPlexData, prefetchJellyfinData, and prefetchTautulliData were loading every cache row for the user with no column projection — for a 50k-item Plex library, ~200 MB of in-memory objects per cleanup run. All three prefetchers now cursor-paginate at 500 rows and project only the columns the watch-map builder actually reads (skipping ratingKey, thumb, title, etc.). Cross-batch Map merging is preserved (same tmdbId appearing in batch 1 and batch 2 still aggregates into one entry with summed watchCount and unioned collections/labels).
• Heap pressure during auto-tag rule execution. executeAutoTagRule was loading the entire libraryCache for the instance — including the per-row data JSON blob — at once. Under webhook concurrency (Connect events firing in rapid succession), this could stack heap pressure and trip the cap. Now cursor-paginates at 500 rows.

Notes

• This patch went through code review, silent-failure analysis, test-coverage analysis, and security review before release. Two take-cap "defensive" limits proposed during initial implementation (take: 10000 on queueCleanerStrike and take: 50000 on plexCache) were removed before merge — review found the strike cap could silently break strike-threshold semantics by truncating rows non-deterministically (no orderBy), and the plexCache cap targeted a non-issue given the small column projection.
• Test additions: 15 new tests covering the new behaviors — exportDatabase history-exclusion + retention-truncation contract, createBackup type-based defaulting (manual / scheduled / update / explicit override), estimateBackupBytes (sparse-row + multi-sample), auto-tag cursor pagination, and prefetchPlexData cross-batch Map merge.