5/5 fix: stabilise duration_based_chunks across collection orders#124
Open
vokracko wants to merge 5 commits intojerry-git:masterfrom
Open
5/5 fix: stabilise duration_based_chunks across collection orders#124vokracko wants to merge 5 commits intojerry-git:masterfrom
vokracko wants to merge 5 commits intojerry-git:masterfrom
Conversation
The previous setup relied on pytest-cov (``--cov pytest_split --cov tests`` in addopts) which has a long-standing limitation when measuring a pytest plugin: the plugin imports happen during pytest's plugin discovery, *before* pytest-cov starts tracing. Anything that runs at import time (class bodies, decorators, type aliases, enum definitions) is reported as uncovered even though it obviously executed - the plugin wouldn't load otherwise. The ``CoverageWarning: Module pytest_split was previously imported, but not measured`` made that explicit. The cookiecutter template the project was generated from worked around this by adding ``--cov tests`` to the addopts, inflating the denominator with 100%-covered test files so the average reached 90%. That pulled the number up but stopped reporting the actual source coverage. Switch to ``coverage run -m pytest`` followed by ``coverage report``. Running ``coverage`` from the outside means tracing starts before Python imports anything at all, so plugin import-time code is counted. Real source coverage jumps from ~74% to 99%; threshold goes from 90 to 95 to reflect that.
Tests previously asserted each TestGroup field separately (`first.selected`, `first.deselected`, `first.duration`). Consolidate to a single equality check against an expected TestGroup, with concrete values for every field rather than only ``selected``.
Tighten the algorithm contract so each one receives a single data argument: a dict mapping pytest items to their durations. Promote the helper that builds the dict to a public name (compute_durations) and inline the two helpers that only it used. The plugin builds the dict once and passes it to the algorithm. Algorithm bodies otherwise consume the same (item, duration) stream they did before, so behaviour is unchanged. Sets up follow-ups that separate group membership from within-group ordering.
Algorithms now own only group membership; the plugin owns the order in which selected/deselected items are emitted to pytest. LeastDurationAlgorithm drops the original_index bookkeeping. Its body collapses to a name + duration sort followed by the heap-based balanced assignment, with selected ending up in heap-pop order. The plugin's collection hook then rebuilds the chosen group's selected and deselected lists in pytest's collection order, keyed on nodeid. That's now the single source of truth for preserving collection order, so subsequent algorithm changes don't have to worry about it. End-to-end behaviour is unchanged: pytest still executes each group's tests in the order it collected them.
Group membership for duration_based_chunks previously depended on the order pytest happened to collect tests, so a parametrize fed a hash-randomised iterable (set, frozenset, dict.keys()) under PYTHONHASHSEED=random could move tests between groups across runs, producing churn in committed durations files and CI group manifests even when nothing changed. The plugin now sorts the durations dict by nodeid before invoking the algorithm. Membership for both algorithms is now a pure function of the (nodeid, duration) multiset and the requested split count. LeastDurationAlgorithm drops its now-redundant nodeid pre-sort (stable sort by duration desc on a canonically-ordered dict has the same effect). DurationBasedChunksAlgorithm walks the canonical- ordered dict unchanged. Within-group execution order is still pytest's collection order, so the only user-visible change is that duration_based_chunks membership distribution shifts once. .test_durations file format is unchanged.
This was referenced Apr 29, 2026
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
1 participant
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.
Fifth and final PR in the stack. This is the user-visible fix:
duration_based_chunksgroup memberships are now stable across pytest collection orders. Closes #25.Motivation.
duration_based_chunkspreviously consumed pytest's collected order directly, so a parametrize fed a hash-randomised iterable (set,frozenset,dict.keys()) underPYTHONHASHSEED=randomcould move tests between groups across runs — churning committed.test_durationsfiles and CI group manifests with no real code change. 0.8.0 fixed this forleast_duration(sort-by-name) but the same fix was never extended toduration_based_chunks. The previous four PRs were preparation; this one swaps in canonical (nodeid-sorted) ordering at the plugin level so membership becomes a pure function of the (nodeid, duration) multiset and the requested split count.What changes. The plugin sorts the durations dict by nodeid before invoking the algorithm.
LeastDurationAlgorithmdrops its now-redundant internal nodeid sort.DurationBasedChunksAlgorithmis unchanged — it walks the dict in iteration order, which is now canonical thanks to the plugin step.Within-group execution order is unchanged — pytest still runs each group's tests in collection order, courtesy of the helper added in 4/5. The user-visible change is a one-time shift in
duration_based_chunksmembership distribution; the.test_durationsfile format is unchanged. README's algorithms table updated.Stack.
coverage run→ 1/5 build: measure coverage on source only via coverage run #120TestGroupdirectly → 2/5 refactor: assert TestGroup directly in algorithm tests #121duration_based_chunks— closes Splits invalid when collection order not deterministic #25