Repository navigation
Conversation
|
Important Review skippedToo many files! This PR contains 359 files, which is 259 over the limit of 100. To get a review, reduce the PR to 100 files or fewer by splitting it into smaller PRs or changing its base branch. Upgrade to a paid plan to raise the limit. Usage-priced reviews support at most 300 files. ⚙️ Run configuration
📒 Files selected for processing: 359
You can disable this status message by setting the
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
snissn
changed the title
|
@codex review Please review the final DATA-only/native105 implementation at |
Codex Review SummaryThis 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. |
Keep DATA page dependencies separate from target metadata in power-loss fixtures, use decoded node types in factory witnesses, and retain exact physical span coverage for absent DATA keys. Consume unpublished vacuum generation metadata only after the original cleanup scopes end.
|
Current-head qualification at The unchanged original 8-pair/128-durable-replacement comparison ran on hosted CI with both route activations and quietness passing:
Median paired time increased 102.55%; timing, bytes and allocation gates all fail. The separate raw-five comparison passes four rows and fails Hosted correctness failures cover borrowed-cohort cleanup, malformed ready-marker refusal, DATA span planning, DATA crash-closure oracle expectations, and physical node flag decoding; Windows also exposes the unsupported-vacuum branch in the cohort fixture. Source causes are being handled as one coherent repair, preserving every failed attempt and existing assertions. Remaining platform/race jobs are still running. The prior focused/public passes retain their original source attribution and do not replace these current-head failures. |
…nt main Retain selected main allocation and tracked-zombie lifetime fixes. Merge native/DATA feature and both format-fixture repairs as one candidate. SYSTEM pager layout, user DATA placement, ownership and ACK contracts remain explicit.
…format experiment
snissn
changed the title
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4466b7396a
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8e98239f07
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cf68ca24a5
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".
| if idx.nativeDocumentStats != nil { | ||
| return *idx.nativeDocumentStats |
There was a problem hiding this comment.
When a VectorIndex is loaded before subsequent inserts, updates, deletes, or a rebuild, its Search method reads the current native graph but this branch keeps returning the one-time snapshot captured by loadNativeDocumentVector. Consequently Nodes, LiveDocs, DeletedDocs, and RebuildNeeded remain stale for the lifetime of an otherwise current-searching handle; refresh these statistics from the current graph or update the cached value as mutations publish.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Accepted as a real in-scope defect. The retained native index handle captures Stats at load while Search reads the current cut. A separate two-file repair now reuses the existing admitted/validated current-cut status reader, preserves the handle-specific rebuild threshold, and leaves held-searcher status cut-bound. Its JSON/BSON lifecycle regression and independent source review are preserved in /mnt/fast4tb/gomap-5139-resume-20261017/artifacts/native-vector-stats-final01/. This patch has not run or been applied to the PR: final cf68 strict timing and allocation gates failed, so the PR is blocked/draft and execution is halted as requested. This thread intentionally remains unresolved; no stale-statistics fix or mergeability is claimed.
TreeDB publishes and recovers native state through the DATA B-tree, replacing the retired PRIMARY representation. Explicit PRIMARY requests, persisted companion namespaces and committed DATA/PRIMARY swaps fail with
ErrLegacyFormatRebuildRequiredbefore recovery writes or cleanup. Incompatible pre-alpha directories require rebuilding. Two recoverable DATA slots, immediate-parent proof, physical identity, persistent value-log dependencies and held snapshots retain their contracts.Imported ColumnStore collections persist partition-live carriers through kind105 command replay. Native JSON/BSON document vectors support CRUD, held searchers and kind103 replay. Immutable imported cuts reuse the original transaction/COW owner with allocator, sequence and complete root-identity validation. Native search retains bounded tombstone routing and deleted waypoints, shares the upper-layer routing budget, and reports
CandidateRows, document materialization and one-shot setup counters. Default, minimal and production statistics modes remain supported; explicit full-diagnostics and work-accounting requests are refused because those counters are not implemented.The final scope restores the established fresh
no_wal_fastUSER leaf default: external leaves in the persistent value log. Explicit true/false caller choices and persisted true/false layouts remain supported, including read-only reopen, environment-conflict refusal and large-payload pointer persistence. USER leaf placement does not change DATA root authority. The failed full-file META API experiment is removed; publication uses the original pinned META primitive and retains both prerequisite-data and final-META durability fences. No recovery acceptance, slot/parent or generation oracle is weakened. The deselected SYSTEM pager-format experiment remains removed.The packet also retains the Windows eligibility guards for tests requiring unavailable online vacuum, reserved-lane error precedence before resource construction, and generation-loop suppression in both parent and child of the unapplied-WAL fixture. The gateway crash test requires a successful journal acknowledgement before exit. External-leaf fixtures explicitly select their intended producer, and crash-witness allocator expectations reflect the combined sync constructor while retaining logical/durable horizons, dependency closure, cuts, seeds and complete recovery-state checks.
NOT MERGEABLE — final candidate failed the unchanged strict durable-cost gate; execution is halted at the user-requested boundary.
Final exact
cf68ca24a5fa769a5da252c5729e889853d5be11versus PRIMARYb03c73fa010987b8ce4c904ff6dca2d2497be393: all eight quiet pairs / 16 original samples are retained, with no retries or exclusions. All five raw microbenchmarks passed. The public durable comparison failed timing and allocation limits:Paired median time increased 118.938883% against the +5% limit; median allocations increased 3.703704% against the no-growth limit. Bytes and both actual route-activation gates passed. This result is a product performance failure, not a flaky-test waiver. Exact failed job; artifact
12304129258retains original binaries, harness/bindings, samples and the failure-only diagnostic trace. Remaining hosted checks may finish after halt; they cannot clear this failed gate.Automatic exact-cf68 review also found stale
VectorIndex.Statson retained current-searching handles. The finding is accepted and remains unresolved. A separate two-file current-cut Stats repair has independent source acceptance, but is unqualified and not applied to this PR. It is preserved locally for a future authorized continuation; it is not a current-head fix or mergeability claim.Historical hosted cost evidence remains source-bound below. Exact prior head
8e98239f075173815d149b8cf5fdc56b58ceaa58completed 54 hosted checks successfully and failed two: the strict durable replacement timing check within the raw-path job, and its aggregate required gate. The five raw microbenchmarks passed. The quiet eight-pair public durable replacement comparison failed:b03c73fa8e98239fPaired median time increased 99.109630%, exceeding the unchanged +5% limit. Bytes, objects, quiet-host checks, actual route activation and both complete public lifecycles passed. Distinct binary/runtime identities remain recorded rather than labeled equivalent. Earlier failures, including
237aa0c4(+97.993812%) and the preceding +99.74148% result, remain retained. The failure-only same-binary traces confirmed the attempted META primitive activated, but both DATA persistence phases remained expensive. Those traces include setup/seed/Close and are diagnostics, not additional acceptance samples or exact timed-loop attribution. No filesystem-cause or performance-improvement claim is made.Final source is
cf68ca24a5fa769a5da252c5729e889853d5be11. All ten focused normal controls passed at this exact commit with zero skips, missing tests or failures: six profile/reopen/DATA recovery/refusal controls, the original public durable replacement lifecycle, and three JSON/BSON vector lifecycle/search controls. Independent GPT-6.1 Sol source review accepted exact commit cf68ca2 / tree 5c75b4c6c95a8d79f7aa42fada5347760994a9f0 under #4879, binding the reviewed components and final source postimages. This source acceptance does not certify performance or hosted CI. At source237aa0c4, the contract guard, 24 cost-runner Python controls, 34 focused normal controls with83 subtests, 74 affected race controls with59 subtests, original history/vacuum lifecycle, and all eight original crash-witness rows with the maintained whole-state comparator passed. These receipts remain bound to237, rather than being presented as current-head or full-Q acceptance. Automatic review findings already dispositioned remain retained; new-head findings still require resolution. No manual hosted trigger is requested under #4879.Focused commands used Go 1.26.9,
GOWORK=off,GOMAXPROCS=2, default GC, and-p=1:This PR does not claim final Q/P5 or Dgraph/Badger acceptance. The final strict cycle has failed. This PR remains open and blocked; no further blind performance variant, retry, new PR or full qualification campaign is selected. PR #5190 is merged. Execution is halted as requested; complete engine/Q and Dgraph acceptance remain open.
Refs #5142, #5144, #5171 and #5139.