daily-held: HF storage 16TB vs 10TB - purge or upgrade
kind: infra#daily-held#needs-human
track:Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Held item
The HF account is 6 TB over its soft ceiling and it is now biting operationally:
- #2091's Step 6a.6 preflight FAILed on 2026-08-06T01:55Z: "HF headroom insufficient … 10 GB projected vs -5997 GB remaining (16.00/10.0 TB, live-api)" — the session proceeded only under an explicit override with a recorded disposition.
- #1491's Path B flagged the same night: "Path B uploads regenerated tensors over LFS, which 403s at the hard quota. The write-gate probe passes right now… but freeing quota means deleting HF artifacts, which is user-only."
- 2026-08-05 evening, Thomas personally drove a 577 GB VM-side reclaim gated on verified HF-backedness — the HF-side ledger itself was untouched.
Which carve-out holds it: destructive/irreversible actions (deleting HF artifacts) + spends money (a quota upgrade) — both user-only decisions.
Suggested action
Decide one (or both):
- Purge candidates — the parked
adapters/issue_397review (~242 GB, memoryproject_pending_hf_quota_upgrade.md, parked since the June quota saga) is the standing first candidate; a fuller purge list can be generated on request (largest prefixes by size, cross-checked against clean-result Repro pins so nothing promoted-body-referenced is deleted). - Upgrade/raise the ceiling — if the 16 TB footprint is the intended steady state.
Until decided, every large-upload plan (regenerated tensor stores, new ladder captures) runs against a preflight FAIL + override friction, and a hard-quota 403 mid-upload is the #552/#541 recovery class.
Provenance
- origin: /daily 2026-08-05 problem sweep — miner 1 P22 + miner 6 P3 (preflight FAIL probed; headroom figures from the sessions' live-api reads).