Repository navigation
chore(xtest): refactor to validate benchmark options before session setup - #640
Merged
Merged
Conversation
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 49 minutes. View limit details
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
| Check name | Status | Explanation | Resolution |
|---|---|---|---|
| Docstring Coverage |  |
Docstring coverage is 18.03% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 61 functions across 4 files. (1 skipped: … | Write docstrings for the functions missing them to satisfy the coverage threshold. |
-
Check name Status Explanation Description Check Check skipped - CodeRabbit’s high-level summary is enabled. Title check The title clearly and concisely describes the main change: refactoring benchmark option validation to occur before session setup. Linked Issues check Check skipped because no linked issues were found for this pull request. Out of Scope Changes check Check skipped because no linked issues were found for this pull request.
Full details: Docstring Coverage
-
Explanation
Docstring coverage is 18.03% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 61 functions across 4 files. (1 skipped: 1 unsupported.)
✨ Finishing Touches 💡 1
-
📝 Generate docstrings 💡
-
- Commit to this branch
- Create a new PR
🧪 Generate unit tests (beta)
-
- Commit to this branch
- Create a new PR
-
- Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
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.
A rabbit checks the bench at dawn
Each SDK name is neatly drawn
The rounds are set, the tests take flight
The harness hops through code tonight
Then carrots crown the config right
Comment @coderabbitai help to get the list of available commands.
ci: run test_bench_config.py in the offline harness check
dmihalcik-virtru
force-pushed
the
DSPX-4372-fix-bench-option-validation
branch
from
October 8, 2026 19:39
fd1a1f6 to
cc15f95
Compare
dmihalcik-virtru
changed the title
b-long
requested changes
Oct 13, 2026
b-long
left a comment
b-long
left a comment
There was a problem hiding this comment.
Review: #640
Validating options early is a clear improvement. I checked these against main: gp@main, --bench-threshold 0.9, mismatched SDKs and go@ all used to collect without error. Now they fail at configure time.
1. The pytest_configure wiring has no test
test_bench_config.py tests perf/config.py on its own. Nothing tests that pytest_configure raises UsageError, or that a lone --bench-baseline/--bench-candidate is accepted, as the description says. A fake config, like the one in test_sizes_units.py, covers this:
class _FakeConfig:
def __init__(self, **opts):
self._options = {
"--bench": True, "--sdks": "go",
"--bench-baseline": None, "--bench-candidate": None,
"--bench-threshold": stats.DEFAULT_THRESHOLD,
"--bench-min-rounds": 20, "--bench-max-rounds": 60,
"--bench-warmup": 5, "--bench-budget-seconds": 1500.0,
"--bench-seed": 0, "numprocesses": None,
} | opts
def getoption(self, name, default=None):
return self._options.get(name, default)
def configure(**opts):
conftest.pytest_configure(cast(pytest.Config, _FakeConfig(**opts)))
@pytest.mark.parametrize("opts", [
{}, {"--bench-candidate": "go@main"}, {"--bench-baseline": "go@v0.29.0"},
{"--bench-baseline": "go@v0.29.0", "--bench-candidate": "go@main"},
])
def test_accepted(opts):
configure(**opts)
@pytest.mark.parametrize(("opts", "match"), [
({"--bench-baseline": "gp@main"}, "unknown SDK 'gp'"),
({"--bench-candidate": "go@"}, "empty version"),
({"--bench-threshold": 0.9}, "threshold"),
({"--bench-baseline": "go@main", "--bench-candidate": "java@main"}, "same SDK"),
({"--bench-baseline": "go@main", "--bench-candidate": "go@main"}, "duplicate"),
])
def test_rejected(opts, match):
with pytest.raises(pytest.UsageError, match=match):
configure(**opts)All 9 cases pass against this branch. Another option is to move the block into a validate_bench_options(config) helper and test that directly.
2. A spec for an SDK outside --sdks is still silently dropped
--sdks java --bench-candidate go@main -> collects 7 java cells; go@main ignored
go is a valid name, so it passes the new check. _spec_for then drops the spec, because it matches no SDK being benchmarked. This is the same silent green run the PR targets. It already happens on main, so it can be a follow-up, but it fits here. Suggested test, which fails today:
def test_spec_for_unbenchmarked_sdk_rejected():
with pytest.raises(pytest.UsageError):
configure(**{"--sdks": "java", "--bench-candidate": "go@main"})3. Pin the new --bench-baseline "" behavior
An empty spec used to mean "use the default". Now it raises unknown SDK '', which fits the PR's aim of rejecting specs that would otherwise be silently ignored. Since this changes behavior from main, please pin it with a test:
def test_empty_spec_rejected():
with pytest.raises(pytest.UsageError, match="unknown SDK ''"):
configure(**{"--bench-baseline": ""})Reject a --bench-baseline/--bench-candidate naming an SDK outside --sdks, which was otherwise dropped per-SDK and silently measured the defaults. Pin the empty-spec rejection and the lone-spec acceptance.
CPython 3.15.0 has no markupsafe wheel, so uv sync built it from source and the --frozen --no-build test step refused to install it.
dmihalcik-virtru
force-pushed
the
DSPX-4372-fix-bench-option-validation
branch
from
October 13, 2026 13:36
81b7300 to
7bed50d
Compare
|
X-Test Failure Report |
Member
Author
b-long
approved these changes
Oct 13, 2026
b-long
left a comment
b-long
left a comment
There was a problem hiding this comment.
Assuming CI jobs pass, this PR is approved 👍
dmihalcik-virtru
added a commit
that referenced
this pull request
Oct 13, 2026
Follow-up to #640, which pinned \`scriptcheck\` to 3.14 after the CPython 3.15.0 release broke it. ### Why it broke The CI Python version wasn't pinned (\`requires-python\` is just \`>=3.14\`), so uv picked up 3.15 as soon as it shipped. markupsafe 3.0.3 had no cp315 wheel, so \`uv sync\` built it from source. The \`uv run --frozen --no-build\` step then refused to install it. ### Changes - **Pin 3.15 everywhere CI sets up Python:** \`check.yml\`'s \`setup-uv\`, plus \`xtest.yml\`'s \`setup-python\` (3.14 → 3.15) and its three \`setup-uv\` steps that had no pin. Those three were also floating to 3.15 and building pydantic-core from source with Rust. - **Relock** so every compiled dependency has a cp315 wheel: - pydantic 2.14.1 / pydantic-core 2.50.1 in \`xtest\`, \`otdf-local\` and \`otdf-sdk-mgr\` - markupsafe 3.0.4 in \`xtest\` \`requires-python\` stays at \`>=3.14\`. This changes what CI runs, not the minimum for local development. ### Verification Locally on CPython 3.15.0, I ran the same steps as \`scriptcheck\` against fresh environments: - xtest: ruff, ruff format, pyright, and the \`--frozen --no-build\` offline harness tests (299 passed) - otdf-local: lint, pyright, and \`-m "not integration"\` tests (29 passed) - otdf-sdk-mgr: lint, pyright, and \`-m "not integration"\` tests (138 passed) A scan of all three lock files found no compiled package without a cp315 or abi3 wheel. On this PR, the X-Test `xct` matrix (go, java, js; platform `main` and v0.28.0) passes on 3.15. The `bench` and `zip64` jobs are opt-in and were skipped, so their `setup-uv` pins have not run yet. Signed-off-by: Dave Mihalcik <dmihalcik@virtru.com>
dmihalcik-virtru
added a commit
that referenced
this pull request
Oct 14, 2026
## Problem The benchmark report's `p (BH)` column always showed the **slower**-tail p-value (`p_adjusted` / `p_value`). For a real speedup that tail is ~1, so an `IMPROVED` row read e.g. `ratio 0.50 … p 1.000 … IMPROVED` — the p-value contradicted the verdict on the same row. The verdict itself was correct (it already used `p_adjusted_faster`); only the rendering was wrong. ## Fix `_p_cell` now shows the one-sided tail in the direction of the observed ratio: the faster tail when `ratio < 1`, the slower tail otherwise, each preferring the BH-adjusted value and falling back to unadjusted (controls stay outside the BH families). The table footer notes that p is one-sided. Ported from the closed #621 branch. Stage-1 follow-up to #625 (DSPX-4372). ## Tests New offline `xtest/test_bench_report.py` (added to the `check.yml` harness list): improvement → faster tail, slowdown/no-change → slower tail, unadjusted fallback, `<0.001` floor, NaN → `-`. Note: the `check.yml` list edit touches the same lines as #640 (which adds `test_bench_config.py`); whichever lands second needs a trivial rebase. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Bug Fixes** * Performance reports now show the p-value for the observed direction of change: faster for improvements and slower for regressions or no change. The footer clarifies that p-values are one-sided. * **Tests** * Added offline coverage for report p-values, including adjusted and unadjusted values, very small values, and empty comparisons. * Expanded the offline test run to include benchmark measurement, runner, and report tests. <!-- end of auto-generated comment: release notes by coderabbit.ai -->
This was referenced Oct 19, 2026
dmihalcik-virtru
added a commit
that referenced
this pull request
Oct 21, 2026
Stages are milestones, not pull requests. Records the stack each stage lands as, with branch names; runs stages 2 and 3 as parallel tracks from main, since they share no imports; and adds a size budget and refactor-first rule to the merge discipline. Also brings the plan up to date: the work that landed as separate fixes (#638, #640, #641, #647), #622's payload sizes, item 7's move into stage 3, the neutral-nightly gap (DSPX-5260), and scratch-disk scheduling (DSPX-5253). Signed-off-by: Dave Mihalcik <dmihalcik@virtru.com>
dmihalcik-virtru
added a commit
that referenced
this pull request
Oct 30, 2026
… record benchmark arms (#626) ## TL;DR Adds **`otdf-sdk-mgr install benchmark <sdk> <ref>...`**: one command that resolves any number of SDK refs (releases, branches, SHAs, `pr:N`) to commits, installs them side by side under `xtest/sdk/<sdk>/dist/`, and writes a manifest recording exactly which commit is in each directory. It makes comparing SDK variants a local, one-command job instead of a CI dispatch or a hand-built set of `dist/` directories. It is part 2c of the [DSPX-4372-t1 plan](#657), based on `main` now that #667 and #668 (the manifest contract) have merged, and it is the preparation step CI will share (#660). ## Why: quicker local testing of multiple SDK variants Today, there are two ways to compare SDK builds: - **Dispatch the xtest workflow with `run-benchmarks`.** That is about 45 minutes per SDK per attempt, and a mis-set input (`main` rather than `main latest`) spends the whole run before it says so. - **Reproduce locally**, per `xtest/perf/README.md`: "whatever two builds you want, side by side under `sdk/<name>/dist/`". You install each build yourself and need to know: - whether each ref is a release or a source build, and the right `--ref` for it; - which directory it will land in: `pr:123` builds into `refs--pull--123--head`, and `otdfctl/v0.24.0` and `v0.24.0` share `dist/v0.24.0`; - how to provision `otdfctl` and the platform separately. Nothing records which commit a directory holds. A `dist/main` left from last week is measured, and labelled, as today's `main`. With this PR: ```bash otdf-sdk-mgr install benchmark go v0.38.0 main pr:123 --platform main --otdfctl v0.38.1 ``` - **Any number of variants in one command.** Refs are kept in order, and the first one is the reference. Stage 3 (#658) measures up to four. - **Fails fast.** Every ref, and the `--otdfctl` pin, is resolved to a full commit before anything builds, so a mistyped branch or tag costs a lookup rather than minutes of compilation. Two refs that would share one `dist/` directory at different commits are rejected before either builds. (A full SHA is not checked to exist; see Follow-ups.) - **Doesn't measure a commit against itself.** If every ref resolves to one commit, preparation stops with `neutral` instead of benchmarking A against A. - **Cheap re-runs.** A source build records its commit in `dist/<slug>/.version` and is reused only when that record matches. Re-running after pushing to a branch rebuilds the branch, not everything. The record is all it checks, so after a hand `make` of local edits, delete `dist/<slug>` (the README says so). - **A manifest instead of hand-built env vars.** `xtest/sdk/benchmark.installed.json` records, for each arm, the requested alias, repository, commit, `dist/` directory and install method, plus the platform and `otdfctl` pins. `otdf-local env` exports its path and the `otdfctl` pin (#642), and from #670 `pytest --bench` reads the arms from it. Until then, name the arms with `--bench-baseline`/`--bench-candidate`, as the README now says. CI runs this same command (#660), so a local result and a CI result come from the same preparation. ## What's in this PR - **`install benchmark`** (`cli_install.py`, `bench_prepare.py`) resolves, checks for collisions, installs releases via `install_release` and source refs via `cmd_tip`, then installs the platform and the `otdfctl` pin. - **Separate provisioning pin.** `--otdfctl` is recorded apart from the measured arms. It joins the collision check only when it shares the arms' `dist/` tree, which is only the case for Go. - **Collision and neutral checks** run before anything builds. They, `dist_slug`, and the manifest's data model are in #668; this PR is the command that uses them. - **Manifest** (schema `v1alpha1`, defined in #668). The status is `success`, `neutral`, `partial` or `failed`, with a `status_message`: - `failed` means nothing was installed; `partial` means some installs completed but not all. - The manifest is written on every exit path, an exception or Ctrl-C included. - The exit code is 0 for `success` or `neutral`, 1 otherwise. **Size.** About 640 non-test lines, over the plan's ~400-line budget. About 315 of them are code; the rest are docstrings, comments and the README section. Resolution, installation and the manifest's exit paths are one command: a PR that resolved without installing would ship a command that does nothing. Two thirds of the diff is tests. **Since the first review:** - Java release arms resolved nothing, because the wrong tag infix was used. Fixed. - The manifest now has one write site: `status` is required, reads check `schema_version`, and writes create the parent directory and replace the file atomically. - JS release arms land in `dist/vX.Y.Z`. - `platform` is rejected as an SDK. - The platform pin records the directory it was installed to. - A first-arm failure is now `failed`, not `partial`. **Since review of #668:** - **Abbreviated commits are rejected.** A ref that resolves to a commit that isn't a full SHA (in practice, an abbreviated SHA that no branch, tag or PR head points at) now fails before anything installs: "give the full commit SHA for …", with a `failed` manifest and exit code 1. Before, `abc1234` and its own full SHA weren't recognised as one build, so the commit was built twice and reported as `success`. Reusing a `dist/` directory and detecting a moved ref now compare full SHAs exactly. - An empty ref list fails instead of writing a `success` manifest with no arms. Every manifest written sets `metadata.sdk` and passes #668's strict reader. - Platform and `otdfctl` install failures are caught narrowly (`PlatformInstallError`, `BenchPrepareError`, `OSError`). Anything else aborts with its traceback, and the manifest records `partial` "Preparation aborted", as arm installs already did. A platform whose `.version` has no full commit is `partial`. **Since the 2026-10-29 review:** - An arm whose stale `dist/` cannot be removed (`OSError`) is now `failed`/`partial` with the error, like the `otdfctl` pin, rather than "Preparation aborted". - If the *first* manifest write fails, the CLI no longer prints "Manifest: …" pointing at an earlier run's file. `prepare_benchmark` raises `BenchPrepareError` for that case. `RuntimeError` is no longer caught around the whole run: `xtest/sdk` is located up front, `--manifest` or not. - The CLI tests now go through Typer (`CliRunner().invoke(install_app, …)`), so a lost registration or renamed option fails them. They also assert a clean `typer.Exit`, since `CliRunner` reports an uncaught exception as exit 1 too. - New tests: - npm-only js releases: one version is `neutral` ("whose commit is unknown"), two versions are not; - a dist of unknown provenance is emptied before rebuilding; - an arm `OSError`; - a failed final write. - Comments: two were wrong and are now corrected: - the repeated-builds example `v0.38.0 main` is in fact neutral, so it is now `v0.38.0 main pr:123`; - the `otdfctl` pin "never provisions the fixtures it is measured against" contradicted `check_collision`. Stale ones fixed: js tags are `sdk-vX.Y.Z`; the go fallback affects `v0.30.x`, not "older tags"; the java Makefile sources `java/<slug>.env`. The `# type: ignore` is gone, and the CLI help example now uses `v0.38.0`/`v0.38.1`. - README: - says reuse trusts `.version`, and that a release `dist/` is reused as is; - says a full SHA is not checked; - says `pytest --bench` doesn't read the manifest yet. **Validation** (on this commit): - `uv run pytest` 404 passed; ruff check, ruff format and pyright are clean, with no `noqa` or `type: ignore`. - Each new test was checked against a deliberate break of the code it covers: the unregistered command, the missing `rmtree`, the uncaught arm `OSError`, the unwrapped first write and the neutral message all fail at least one test. - Live, from an empty `xtest/sdk/go`: - `install benchmark go v0.38.0 main --otdfctl v0.38.1` gave `success`, exit 0. The arms were `v0.38.0` at `308002c2` (release) and `main` at `ef874839` (source). `dist/main/.version` matches `git -C src/main rev-parse HEAD`. The pin was `v0.38.1` at `99b51145`. - The same command re-run gave `success` in 1.4 s, with nothing rebuilt. - `install benchmark go 0e4b0d1 main` gave `failed`, "give the full commit SHA for 0e4b0d1", with exit 1 and no arms. - `--manifest` under a regular file gave "Error: cannot write the manifest …", exit 1, and no "Manifest:" line. - The platform pin (`--platform`) was not exercised live on this commit; its last live run predates the review fixes. ## 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 (merged) | `main` | | 2b: manifest contract | #668 (merged) | `main` | | 2c: `install benchmark` | **#626 (this PR)** | `main` | | 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 | #667 and #668 were split out of this PR so the fixes and the contract can be reviewed on their own. Split out of an earlier version so they could land first: #638, #639, #640 and #641, all merged. #647, the golden-key configuration CI needs once `otdf-local` starts the platform, is still open. **Related, independent of this stack:** #684 makes go release installs put a real `otdfctl` binary in `dist/`. Before it, a go release arm ran `go run` on every call (about 370 ms wall and 1 s CPU, against 54 ms for the binary), while source-built candidates ran a binary. That made release baselines look slower than they are. ## Follow-ups None of these blocks merge. All but the "Re-run" message are now noted in the code or the README. - **Go repository URL.** Go arms are all recorded under the monorepo URL. The monorepo carries `otdfctl/v0.10.0`–`v0.29.0` at the archived repo's commits, but not `v0.30.x`, which `resolve` answers from the archived repo. That makes `v0.30.x` arms the ones recorded with a commit the monorepo lacks. The fix needs `resolve` to report the repo it matched. - **Fallback on any error.** Go resolution falls back to the archived otdfctl repo whenever the monorepo lookup fails, so a transient error can resolve `main` or `pr:N` against the wrong repo. A source arm then fails the read-back, but the error says "the ref moved". - **Unhelpful "Re-run" advice.** "The ref moved… Re-run" is wrong for a worktree that `is_mutable_ref` treats as fixed (full SHAs; web-sdk `release/x.y.z` branches). Checkout never resets those, so a re-run fails the same way. The message should name `src/<slug>` and say to delete it. - **Reuse trusts `.version`.** The SDK Makefiles rebuild into `dist/<slug>` without touching `.version`, and `_built_commit` doesn't check for a dirty worktree. The README now tells users to delete the directory; a code check (`git status --porcelain`, a binary hash) would be better. - **Release dists reused as is.** A release arm adopts an existing `dist/<tag>` (`install_release`'s skip), including a half-finished install. - **Full SHAs unverified.** A typo fails at checkout after the arms before it have built, giving `partial`. - **`--platform` not resolved up front.** It is not resolved before the arms build: `--platform latest` becomes `service/vlatest` and fails after every build. The release/source label for the pin is also a `startswith("v")` guess; both installers build from source. **Downstream:** #642 is rebased onto this commit with no conflicts (it touches only `otdf-local` and `xtest/perf/README.md`). Its local-run example uses `--otdfctl v0.30.0`, one of the `v0.30.x` tags above; `v0.38.1` would avoid that. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> <!-- coderabbit-ai-generated:v1 --> ## Summary by CodeRabbit * **New Features** * Added an `install benchmark` command to prepare SDK benchmarks from ordered references, with optional platform and CLI tool artifacts. * Benchmark preparation validates references and installation conflicts, reuses matching builds, and records progress and results in a manifest. Results distinguish successful, neutral, partial, and failed preparations. * **Documentation** * Added guidance for preparing benchmarks and troubleshooting common test failures. <!-- 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.
Split out of #626.
--bench*options were validated late: a bad--bench-thresholdsurfaced from thebench_configfixture after payloads were built and CLIs installed, and an unknown SDK in an arm spec was skipped silently byselect_arms.xtest/perf/config.py— pytest-free module owningBenchConfig(moved fromperf/runner.py, re-imported there sofrom perf.runner import BenchConfigstill resolves) and the request rules: SDK-name validation, arm ceiling, duplicate rejection, finite positive budgets/timeouts,max_rounds >= min_rounds.conftest.pytest_configurechecks every--bench*option before session setup. A full baseline/candidate pair gets the paired rules (same SDK, no duplicates); either flag alone is still allowed —select_armsdefaults the other side, as the option help says — and is checked only for a valid SDK name.Tests:
xtest/test_bench_config.py; alltest_bench*.pypass (167).Summary by CodeRabbit