Repository navigation
feat(hosting): embed concurrent repository hosts behind gateways - #66
Conversation
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
|
@codex review |
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cd713bd5ea
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Serialize encoded chunk claims and PUTs against failed-stage deletion while keeping unrelated keys concurrent. Forward job-view recovery deadlines to the shared store, and budget signed proofs plus cursor envelopes before selecting tree pages. Five regressions reproduce the three Codex findings on the previous head. Locked installation, 19 focused tests and the full 265-test host suite pass.
|
@codex review |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 933c5ce8b4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Reserve the serialized history envelope and an actual signed cursor before selecting commit records. Add a real-Git SHA-256 long-namespace regression that reproduces the former 413 and verifies bounded, complete pagination. The eight focused browse/regression tests and full 266-test host suite pass. Admission and deployment caps remain unchanged.
|
@codex review |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 606b8dd705
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Keep a shared scheduling revision so an alarm cannot erase a retry scheduled by a protected job while recovery awaits R2. Cover deferred encoded and temporary cleanup, and initialize the constructor-bypassing cleanup fixture. Verified both regressions fail on the reviewed head, 20 focused tests pass, and all 268 host tests pass without skips.
|
@codex review |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 85837c6150
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Keep the Store private until the single request/alarm initialization promise finishes orphan recovery. Otherwise a second initializer can admit live stages that the first pass removes. Both arrival-order regressions fail on the reviewed head. All 22 focused tests and 270 host tests pass without skips; accepted admission and caps are unchanged.
|
@codex review |
Git versus git-zeroCommit
Matched stored-DEFLATE/no-delta policies; persistent warm servers and OS caches. Wall time, CPU and peak RSS describe client commands, excluding servers. RSS is the largest waited process peak, not aggregate server memory. Deterministic alternating order and shared servers keep these measurements descriptive; no significance or causal claim. All original measurements are retained by Runnerless. |
|
Codex Review: Didn't find any major issues. Chef's kiss. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
An embedding product can host ordinary Git repositories with its own users/API keys, browse them over HTTP, manage lifecycle/policy, and catch up on durable push records. The module supports both a separate Worker (HTTP/service binding) and an embedded Worker whose Rust router calls the Durable Object stub directly. All six original milestones remain implemented. The accepted fifth-review admission design and caps remain unchanged. The implementation also protects shared chunk ownership during cleanup, preserves concurrent job recovery scheduling, shares cold initialization between alarms and requests, and budgets signed tree/history pages.
Delivered behavior
Grants
X-Git-Zero-Grant: v1.<namespace>.<read|write|admin>.<expiry>.<actor64>.<mac64>HMAC-SHA-256 over exact UTF-8
git-zero-grant-v1\n<namespace>\n<access>\n<expiry>\n<actor64>\n, including final LF; canonical unpadded base64url fields.This boring MAC avoids cross-language JSON canonicalization. Use the exact UTF-8 secret (at least 32 bytes), synchronized clocks and about 120-second grants; strict expiry stays
now < expiry <= now+300. Authorization is captured at DO arrival. The contract and public test vector define every route, error and forwarded stub field.Storage and browse correctness
All six defects were independently reproduced.
review-bot-regressions.test.mjshas ten cases reproducing the failures on the reviewed heads: concurrent chunk loss, warm encoded/temporary recovery, 8 KiB tree pages in both formats, SHA-256 history with an 80-character namespace, protected encoded/temporary cleanup rescheduled during an alarm, and request-first/alarm-first cold initialization overlap. The latest 22 focused recovery/lifecycle/grant/storage-policy regression tests and complete 270-test host suite pass. A constructor-bypassing cleanup fixture explicitly initializes the shared recovery marker; its assertions and deadlines are unchanged.Fifth-review fixes
admission.test.mjssaturates A, queues another A request, then admits B's clone and push within one second; the three-blob saturation/300-browse-read tests remain. Fairness is per repository, not per team. Admission-group keys are a documented limit; the gateway must enforce team-wide quotas.shared-pack.test.mjscounts exactly one/three preparations for one/three uncacheable clones and exactly one for a new key against a full cache. No N+1 discarded walk.host-module.test.mjsrejects 64/96/99 MiB, performs a real fetch at the exact 100 MiB minimum, and pushes/clones at 101 MiB. Default token mode has no isolate-wide memory bound, plus up to four idle codecs/factory (24 MiB allowance, 22.75 MiB actual linear memory).concurrent-git.test.mjspauses A's actual pack until B's pack begins, rather than mistaking v2 ls-refs for a pack; overlap, checkout/fsck and the original 30-second assertion remain.shared-pack.test.mjsabandons all followers before output and requires cleanup within its scaled window/bound. Gateway defaults: 256 MiB decoded push cap, 1 GiB codec pack cap (also fetch output), explicitly raisable; token defaults unchanged.host-module.test.mjschecks both modes.shared-pack.test.mjscuts a cache chunk inside a packet;shared-failure-git.test.mjsinjects a second-chunk PUT failure and requires real Git's “shared pack generation failed; retry fetch” remote message without framing/disconnect diagnostics.browse-scale.test.mjsrequires observed work to equal that reference and SQL < 24 + 5 × reference work, with zero writes and a reference-adjusted 1k-to-25k growth check. API overwork can no longer raise its own allowance.Eight targeted regressions copied into an isolated archive of
cf3655e9all failed there and pass with the fixes: heavy admission, invalid token pool, default caps, one/three uncacheable walks, orphan cleanup, partial-frame fatal error and real-Git fatal error. Item 6 strengthens an oracle; ordinary baseline read costs were already correct.Measurements
The fifth-review verification record contains methods, figures and limits.
Decisions and assumptions
Verification
Locked npm installation/build; full local host suite 270/270 in 407.362 s, zero failures/skips/cancellations, plus the strengthened three-blob admission focused rerun. Rust workspace tests, 96 Python tests, 39 isolated warning-denied feature builds, and public-package checks passed. The first full run exposed a test profile needing a third heavy lease; its explicit arena profile was corrected, preserving every assertion and timeout, then the whole suite passed. No Rust source or dependencies changed, and no lint/fuzz/check was weakened.
On head
65ae7457fd423da979365221fc8ea6a3f1ab9c75, all 27 applicable checks passed (both Rust/worker runs, all 16 interop shards, fuzz, benchmark, Runnerless title/commit checks, ci-toolkit and the Codex gate); lifecycle is intentionally skipped. The fifth requested Codex review returned “Didn't find any major issues.” Six VALID inline findings were reproduced, fixed with regression tests, and resolved; none were dismissed as speculative. The branch was not rebased because main stayed an ancestor. No deployment or release is part of this change.Generated repositories/packs/archives were removed; only small ignored .scratch/ evidence logs remain. No other checkout or process arguments/environments were inspected. Historical timeouts remain the Wrangler/direct-workerd variants of “matches native tree-depth filters across shared paths and history”; old resets were Wrangler tree-depth and direct-workerd wanted-closure. The exact historical cause remains unproven; original tests pass unchanged.