Repository navigation
chore(xtest): measure benchmark cells over a tuple of arms - #669
Draft
dmihalcik-virtru wants to merge 1 commit into
Draft
dmihalcik-virtru wants to merge 1 commit into
dmihalcik-virtru wants to merge 1 commit into
Conversation
The runner, statistics and report take a tuple of arms with a reference instead of a baseline/candidate pair. Option handling and arm selection still produce two arms, so a run measures what it measured before; part 3b adds more. - runner: Arm.name is the arm's id -- the installed dist tag -- rather than its role. CellResult records arm_ids, arm_labels and the reference; samples are keyed by arm id. run_cell checks for two or more arms with unique ids, records a round only once every arm has completed, and stops on precision only when every candidate's gated intervals are narrow enough. analyze() makes one contrast per candidate against the reference, keyed by contrast_key(). - stats: worst_control_key() picks the least reassuring of a control's A/A contrasts as the noise floor, ties broken by order. With two arms there is exactly one. - report: the JSON artifact is schema 2. Each cell names its arms and reference, and its statistics are under contrasts, one per candidate; a two-arm cell keeps the baseline/candidate labels. nothing_measured is recorded. The step summary has one row per candidate and metric. - fixtures: build_arms names each arm after its build and numbers the control's two copies of the baseline, since ids key the sample vectors. Output names append their suffix rather than using with_suffix, which would turn v0.29.0 into v0.29 and give two builds one output file. Part 3a of the DSPX-4372-t1 plan. Signed-off-by: Dave Mihalcik <dmihalcik@virtru.com>
|
Important Draft PR not reviewedDraft PRs are not automatically reviewed by default.
To automatically review draft PRs, update your CodeRabbit configuration: reviews:
auto_review:
drafts: true
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 |
|
This was referenced Oct 21, 2026
Draft
dmihalcik-virtru
changed the title
dmihalcik-virtru
added a commit
that referenced
this pull request
Oct 23, 2026
…pted js installs (#667) Part 2a of the [DSPX-4372-t1 plan](#657), split out of #626: fixes the benchmark needs, which every `otdf-sdk-mgr` install benefits from. **Based on `main`.** ### What it does - **Bare repos refresh their branches.** `git clone --bare` records no fetch refspec, so the `fetch --all --tags` before every checkout refreshed tags only. Every branch stayed where the first clone found it, and `install tip main` or `--platform main` built that stale commit. `checkout.BARE_REFRESH` now fetches origin's branches into `refs/remotes/origin/*` with `--prune`, for the SDK and platform bare repos alike, and `checkout.upstream_ref` maps a branch name onto them. Fetching into the bare repo's own `refs/heads/*` instead would move branches that other worktrees have checked out, leaving their HEAD ahead of their files, and would fail for good once upstream replaced a branch `foo` with `foo/bar`. - **Worktrees are detached** and reset to the refreshed ref. A branch origin no longer has is refused rather than built from the clone's stale copy, and `HEAD` means origin's default branch as of the clone; a platform `HEAD` worktree used to reset to itself. - **Go source rebuilds work a second time.** `checkout_go_from_platform` reset an existing worktree to a `FETCH_HEAD` fetched into the bare repo, which no worktree can see, so every rebuild after the first failed with `pathspec 'FETCH_HEAD' did not match`. It now resets to the refreshed ref, like the SDK and platform paths. - **`install tip` reports checkout failures as errors.** A refused branch, or a failed git command with git's stderr, now prints `Error: …` and exits 1 for go, js and java, as the platform path already did, instead of a traceback. - **An interrupted js install leaves nothing behind.** #663 already removes a partial `dist/` when npm fails. A Ctrl-C mid-install still left a directory holding only `cli.sh`, which `install_release` then skipped as installed from then on. Both js install paths now remove it unless the install finished. - `AGENTS.md` claimed the bare clone mapped `refs/heads/*`, which is what hid the first bug. It and the code comments also said `git pull` fails in a worktree; it merges origin's default branch into whatever is checked out. ### Validation - `uv run pytest` in `otdf-sdk-mgr`: 207 passed, and 321 at the top of this track (#642). ruff, ruff format and pyright are clean. - `test_checkout.py` runs real git against a local origin. Every shared behaviour is parametrized over the SDK, go and platform entry points: new and existing worktrees follow their branch, `HEAD` (including beside a tag named `main`), a deleted branch is refused, tags and SHAs are checked out and then reused, and a force-pushed `refs/pull/N/head` is followed. This replaces the stubbed command-string test for platform PR refs. SDK-only tests guard other worktrees, force pushes and `x` → `x/y` renames; platform-only tests cover the helper-scripts worktree. A `CliRunner` test covers the `install tip` error path. - Mutation-checked: `symbolic-ref --short`, treating every non-origin name as a deleted branch, dropping go's explicit PR-ref fetch or its `+`, swallowing go's refusal, and dropping the `install tip` catch each fail a test. - Two new tests interrupt each js install path. - `isolated_git` keeps those tests off the developer's git config, any inherited `GIT_DIR`, and the network (`GIT_ALLOW_PROTOCOL=file`). ### Follow-ups Older problems found while reviewing this PR are fixed in a stack of drafts based on it: #673 (don't reset a directory that isn't a worktree) → #674 (branch names vs versions) → #675 (tags moved upstream) → #676 (`checkout go` / `--all`). ### Stack ([DSPX-4372-t1 plan](#657)) Two parallel tracks from `main`; they share no imports and meet in stage 5. | Part | PR | Base | |---|---|---| | 2a: bare-repo and js install fixes | **#667 (this PR)** | `main` | | 2b: manifest contract | #668 | 2a | | 2c: `install benchmark` | #626 | 2b | | 2d: `otdf-local env` exports | #642 | 2c | | 3a: arms as a tuple | #669 | `main` | | 3b: two to four arms | #658 | 3a | | 3c: pytest reads the manifest | #670 | 3b | | 4: candidate-to-candidate conclusions | #659 | 3c | | 5, 6: CI and report | #660, #661 | not yet moved onto these stacks | <!-- This is an auto-generated comment: release notes by coderabbit.ai --> <!-- coderabbit-ai-generated:v1 --> ## Summary by CodeRabbit * **Bug Fixes** * SDK and platform worktrees now follow updated branches and pull request refs, including after force-pushes. Missing branches produce a clear error instead of a traceback. * Interrupted JavaScript installations no longer leave partial distribution files behind. * **Documentation** * Updated platform installation guidance for branch refreshes, detached worktrees, and pull request refs. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Signed-off-by: Dave Mihalcik <dmihalcik@virtru.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 3a of the DSPX-4372-t1 plan, split out of #658. Based on
main: stage 3 changes onlyxtestand imports nothing from stage 2, so it can be reviewed alongside #626.What it does
The runner, statistics and report take a tuple of arms with a reference instead of a baseline/candidate pair. Options and arm selection still produce two arms, so a run measures what it measured before. #658 adds more.
CellResultrecordsarm_ids,arm_labelsand thereference. A round is recorded only once every arm completes, and the precision rule stops only when every candidate's gated intervals are narrow enough.analyze()makes one contrast per candidate against the reference, keyed bycontrast_key().worst_control_key()takes the least reassuring of a control's A/A contrasts as the noise floor. With two arms there is exactly one.schema2: samples keyed by arm id, each cell'sarmsandreference, per-candidatecontrasts, andnothing_measured. Two-arm cells keep thebaseline/candidatelabels. The step summary has one row per candidate and metric.build_armsnames each arm after its build and numbers the control's two copies of the baseline.analyze()registers control keys before RSS censoring, so a censored control-only run is still NOTHING MEASURED.Changes from #658 as it was
build_armscalledwith_suffix(".tdf")on names that now end in a version, which turnedv0.29.0intov0.29.tdf. Both copies of a release in the control cell, or two releases such asv0.29.0andv0.29.1, wrote one output file. It now appends the suffix. A new test with a release baseline caught this; the old tests used dot-free names.run_celldeleted every arm's output in afinally. feat(xtest): benchmark payloads up to 1 GiB, with per-dispatch budget #622, now onmain, made that cleanup opt-in through--bench-prune-scratchso that local runs keep their outputs to inspect, and the unconditional delete contradicted it (and re-indented most ofrun_cell).Compatibility
schema. Version 1's per-cellmetricsblock is nowcontrasts.Validation
uv run pytest test_bench_*.py: 256 passed. ruff, ruff format and pyright are clean.Stack (DSPX-4372-t1 plan)
Two parallel tracks from
main; they share no imports and meet in stage 5.maininstall benchmarkotdf-local envexportsmain