Skip to content

0048. A Ready clip Loomarr fetched keeps only its airing copy

  • Date: 2026-11-12
  • Status: accepted
  • Supersedes: 0028’s “an admitted ad keeps its master” (its 2026-10-27 amendment) and “masters are not garbage-collected”

Context

0028 kept every acquired original for good, so reprocessing and recovery never depended on a lossy rendition. On a household install that made filler a collection that only grows: on 2026-11-12 the maintainer’s install had a full 50 GB budget, every fetch refused, and a large share of the bytes in originals of clips already on air. The maintainer decided filler is a cache sized to the channels, not a collection to keep (#3838, decision 14, grilled 2026-11-12).

Decision

  • Once a clip Loomarr fetched is Ready, its original download is released. The file goes to the 24-hour quarantine, then away. The sidecar keeps the original’s SHA-256, length and sparse id with no path: what was acquired stays on record, and the playback copy is what airs.
  • A clip added by hand keeps its original. So does a clip the library re-encode still has to move onto the current profile; its original is released after.
  • An original another clip still names stays until the last clip naming it is released.
  • An original stays while it is the only way back to the clip: while the file the clip airs is missing or not the bytes its sidecar records, or its acquisition is not consumed (in repair).
  • A released clip stays on the encode it has (maintainer, 2026-11-14). A change of encoder or playback recipe never retires it; ads turn over within weeks, so the older encode leaves on its own. Loomarr never re-runs the transcode, split or screening of a clip whose original is released. A person who asks to re-run one of those rungs retires the clip instead: it comes off air as a retired reject, its files go when a reject’s rest is over, and the fetcher’s top-up replaces it. Later rungs read the playback copy and still run.
  • No original is released while the GPU may only be late. A boot whose encoder detection fell back to software, on a host that measured a hardware encoder within the week, releases nothing until hardware is back or the week is over.

Consequences

  • Disk per airable minute falls to about the size of the playback copy.
  • A playback-recipe change after release reaches only clips fetched afterwards; the library re-encode skips a clip with no original.
  • The evidence copy was already dropped at admission (0028’s amendment), so no rung after Ready read the original; nothing that runs on its own loses an input.
  • The sidecar’s media manifest accepts a source master with no path; its authority digest ignores paths, so no screening identity moves when an original is released.