Skip to content

Widen blocker terms to excluded membership - #604

Merged
notatallshaw merged 2 commits into
mainfrom
perf/widen-blocker-membership
Jul 30, 2026
Merged

notatallshaw merged 2 commits into
mainfrom
perf/widen-blocker-membership

Conversation

@notatallshaw

Copy link
Copy Markdown
Owner

Look-ahead blocker terms widen from the decided version's neighbor gap to the complement of the union of the dependency ranges the scanned candidates were rejected against. Each of those rejections fired because the blocker's version sat outside that candidate's range, so every blocker version outside the union repeats the same rejections and one clause excludes all of them. The widening falls back to the neighbor gap when any rejection in the group lacks a recorded range, or when the complement would not cover the group's own blocker.

pdm-skypilot-dvc, which timed out on every prior configuration, now resolves deterministically in 49s, and the conflict-heavy uv-issue-10344-airflow-py312 drops roughly half its decisions on top of the span stack. The direction came from measuring that storm's conflicts: each one arrived already asserting and carried a single excluded version, so exclusion had to move into the clause supply rather than the search order.

Stacks on #592.

@read-the-docs-community

read-the-docs-community Bot commented Jul 29, 2026 •

Copy link
Copy Markdown

Documentation build overview

📚 nab | 🛠️ Build #33825919 | 📁 Comparing d4c3aab against latest (a5594a0)

  🔍 Preview build  

No files changed.

Base automatically changed from perf/span-decided-runs to main July 29, 2026 23:48
@notatallshaw
notatallshaw marked this pull request as ready for review July 30, 2026 00:10
@notatallshaw
notatallshaw merged commit be28a2a into main Jul 30, 2026
59 checks passed
@notatallshaw
notatallshaw deleted the perf/widen-blocker-membership branch July 30, 2026 00:10
charliermarsh pushed a commit to astral-sh/uv that referenced this pull request Aug 8, 2026
This is a partial port of a technique I am adopting in nab, I have been
following @konstin's lead in
astral-sh/pubgrub#75 and extending the approach
into more places (notatallshaw/nab#357,
notatallshaw/nab#590,
notatallshaw/nab#592,
notatallshaw/nab#604).

This speeds up `uv lock` with `requires-python >= 3.8` over boto3 +
awscli by ~20% wall time, and the largest accumulated range in the run
drops from 287 segments to 2.

`add_unavailable_version` records one singleton incompatibility per
rejected version, so consecutive rejections never coalesce and the
accumulated range grows a hole per version. This widens each singleton
to the gap between its neighbors in the known versions, the same
interval decided-version dependencies use.

The win needs long runs of rejected versions. The main source is a
Python floor below the ecosystem's support boundary: boto3 and awscli
dropped 3.8 in April 2025 and now require >= 3.10, so a 3.8 floor
rejects hundreds of releases in one run, while a 3.10 floor rejects
almost nothing (14 rounds, 2 segments either way) and the change is a
no-op.

The first commit keys the pre-release hint on the requesting requirement
instead of the derived set's bounds: widened bounds land on whatever the
registry lists next, and a pre-release neighbor used to read as a
pre-release marker nobody wrote. Rejection widening makes that routine,
so the fix leads the series. The second commit merges unavailable
versions that are causes of the same derivation, and the third is the
widening itself. The fourth keeps the subsequent-index hint, which was
suppressed whenever the set covered all versions; after widening that
also covers "the index listed versions and none of them worked", which
is the case the hint is for.

The last two commits are #20809, merged in here so the error messages
land as one change. With the widening, uv no longer asks for a version
in a gap it created, so the `no_versions` incompatibility behind "only
the following versions of X are available" is no longer created there,
and steps that cross a gap stop following from the premises next to
them. The fifth commit adds that clause back while the tree is reduced,
on whichever cause rules the package out, or on the step itself when
neither does. The sixth reports the one version a package has as that
version rather than as a range over it, so a rejected `sphinx 7.2.6`
reads `sphinx==7.2.6` and not `sphinx>=7.2.6`. Neither touches the
solver.

<details>
<summary>MRE</summary>

```toml
[project]
name = "mre"
version = "0.1.0"
requires-python = ">=3.8"
dependencies = ["boto3", "awscli"]
```

Max accumulated segments (287 on main, 2 with this PR; `uv.lock`
identical):

```console
$ uv lock --exclude-newer 2026-07-01 -v 2>&1 | grep 'Searching for a compatible version' | awk -F'|' '{print NF}' | sort -n | tail -1
```

Wall (run without `-v`, which prints the accumulated ranges and
dominates the time):

```console
$ hyperfine --warmup 3 --min-runs 20 --prepare 'rm -f uv.lock' 'uv lock --exclude-newer 2026-07-01'
```

</details>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant