Summary
hyperdb-bootstrap never removes a superseded hyperd release, so every pin bump leaves the previous one on disk permanently. On this workstation that is already 793 MB for a single pinned version:
263M .hyperd/0.0.26359.r07abb490 <- superseded pin, nothing references it
265M .hyperd/0.0.26479.r96880f6a <- the current pin
265M .hyperd/current <- a full copy of the current pin
793M .hyperd
Two things compound:
- No pruning on bump.
download_and_extract clears only the directory for the release it is installing, and refresh_current clears only current/. Neither considers older versioned directories, so each bump adds ~263 MB that is never reclaimed.
current/ is a full file copy rather than a symlink. That is deliberate — it avoids requiring symlink privileges on Windows — but it means the active pin is always stored twice.
Nothing in the repository references the superseded directory by path; the only 26359 matches are version strings in changelogs, benchmark history, and a --version example.
Why the existing cleanup does not help
make clean does rm -rf .hyperd, which removes current/ along with everything else and forces a full re-download of the pinned release. So the only available cleanup is all-or-nothing: either carry the superseded copies indefinitely, or discard the working engine too and re-fetch ~265 MB.
There is no targeted way to say "keep the current pin, drop the rest."
Fix direction
Either would resolve it, and they compose:
- Prune on install. After
refresh_current succeeds, remove versioned directories under .hyperd/ other than the one just installed. Doing it after the new pin is verified and current/ is refreshed keeps the operation safe — a failed download must not leave the tree with no usable engine.
- Add a targeted cleanup target. A
make clean-hyperd-stale (plus the build.ps1 equivalent) that removes superseded versioned directories while leaving current/ and the active pin intact. Useful independently of (1), since it also cleans up trees that predate the change.
Worth considering whether to keep exactly one previous pin, which would make an A/B benchmark against the prior release — the procedure the update-hyperd-release skill documents — possible without re-downloading. That argues for pruning to N most recent rather than to one.
Provenance
Noticed while correcting the HYPERD_PATH guidance (#298). Verified against main at 05e993f.
Summary
hyperdb-bootstrapnever removes a supersededhyperdrelease, so every pin bump leaves the previous one on disk permanently. On this workstation that is already 793 MB for a single pinned version:Two things compound:
download_and_extractclears only the directory for the release it is installing, andrefresh_currentclears onlycurrent/. Neither considers older versioned directories, so each bump adds ~263 MB that is never reclaimed.current/is a full file copy rather than a symlink. That is deliberate — it avoids requiring symlink privileges on Windows — but it means the active pin is always stored twice.Nothing in the repository references the superseded directory by path; the only
26359matches are version strings in changelogs, benchmark history, and a--versionexample.Why the existing cleanup does not help
make cleandoesrm -rf .hyperd, which removescurrent/along with everything else and forces a full re-download of the pinned release. So the only available cleanup is all-or-nothing: either carry the superseded copies indefinitely, or discard the working engine too and re-fetch ~265 MB.There is no targeted way to say "keep the current pin, drop the rest."
Fix direction
Either would resolve it, and they compose:
refresh_currentsucceeds, remove versioned directories under.hyperd/other than the one just installed. Doing it after the new pin is verified andcurrent/is refreshed keeps the operation safe — a failed download must not leave the tree with no usable engine.make clean-hyperd-stale(plus thebuild.ps1equivalent) that removes superseded versioned directories while leavingcurrent/and the active pin intact. Useful independently of (1), since it also cleans up trees that predate the change.Worth considering whether to keep exactly one previous pin, which would make an A/B benchmark against the prior release — the procedure the
update-hyperd-releaseskill documents — possible without re-downloading. That argues for pruning to N most recent rather than to one.Provenance
Noticed while correcting the
HYPERD_PATHguidance (#298). Verified againstmainat05e993f.