Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
32 changes: 24 additions & 8 deletions ROADMAP.md
Original file line number Diff line number Diff line change
Expand Up @@ -1036,12 +1036,24 @@ P28 review follow-ons, the port de-flake, the clippy allow-list) are recorded in
`--allow-private-branches` (the daemon needs non-interactive git auth); (9) a
download budget: every fetch into the cache runs under a watchdog that kills
git over its budget (disk bytes per fetch; not RAM inside `index-pack`, not the
total cache). (4) validated against a private GitHub repository on 2026-11-21 (custom
total cache); (11) fewer round trips per push: `Bucket::put_new_batch` and
`exists_batch` (one `git push` per chunk of at most 64 MiB of values, one
create-only lease per ref, never `+` or `--atomic`; one `ls-remote` per chunk
of exact refs) carry the pack and idx and the pre-commit presence check, the
view after a won manifest CAS is built from the manifest just written instead
of read back, and `has_object` trusts a 2 s memo of an empty remote instead of
probing once per object. Measured with a logging `SC_GIT` shim against a local
bare repository, one small commit through `sc push`: a first push 19 network
git calls → 10, a second push 11 → 8 (the open-time manifest read included;
the three probes of an empty remote, at open, in `list_refs` and in
`update_ref`, remain because only `has_object` may use the memo). Not in scope:
`ref_guard::check_ref_move` still downloads the old tip's pack to compare kinds
on every push after the first (bandwidth, not round trips; the idx carries no
kind). (4) validated against a private GitHub repository on 2026-11-21 (custom
refs, lease CAS with one winner of two racing pushers, filtered fetch,
compaction 18 → 5 refs); still unexercised there: per-push and per-blob limits,
rulesets on custom refs, and CAS across GitHub replicas beyond one race. Still
deferred: (11) fewer round trips per push (a first small push to GitHub took
56 s, one HTTPS round trip per git call); (10) a group-leading supervisor for the serve-mode watchdog: run git
deferred: (10) a group-leading supervisor for the serve-mode watchdog: run git
under a std-only `sh` leader (`sh -c 'git "$@"; echo $? > <status file>; read
_' sh <args>`, stdin piped from sc) so the process group id stays pinned after
git exits, and the watchdog can group-kill a helper that outlives git (the
Expand All @@ -1057,12 +1069,16 @@ P28 review follow-ons, the port de-flake, the clippy allow-list) are recorded in
does not support them and sc-forge is SHA-1, so the value is low and the cost
moderate (the cache's object format would have to be detected before the first
fetch); (S2, from review) the ssh user name that `remote list` prints is not a
credential, and `http(s)` userinfo is refused at validation. Known issue,
recorded and not fixed: a push to `refs/sc/K` fails with "dst refspec matches
credential, and `http(s)` userinfo is refused at validation. Declined: a push to `refs/sc/K` fails with "dst refspec matches
more than one" when a decoy branch such as `refs/heads/refs/sc/K` exists on the
host (found while building `serve --store`). Whoever can push to the
repository can plant such a decoy, so it is a writer-only denial of service on
that key; reads compare exact full ref names and are not affected.
host (found while building `serve --store`; a create goes to the decoy and is
lease-rejected). Git has no option to turn off the push destination's DWIM
lookup, so a fix would mean avoiding refspecs for these keys altogether. Writers
who can push branches can also push `refs/sc/*` on hosts where custom refs are
not separately protected (verified for sc-forge, unverified for GitHub
rulesets), and a writer can already delete or overwrite any bucket ref, so the
decoy adds no capability; reads compare exact full ref names (`exact_only`) and
are not affected.
- **Serve-side persistent pack cache (P36c follow-on).** A bucket-backed
serve instance re-downloads packs per connection; a content-addressed
on-disk cache in the serve home would make warm instances cheap without
Expand Down
Loading
Loading