## Summary
The `--index-strategy` is linked to the index locations, which we
propagate to source distribution builds; so it makes sense to pass the
`--index-strategy` too.
While I was here, I made `exclude_newer` a required argument so that we
don't forget to set it via the `with_options` builder.
Closes https://github.com/astral-sh/uv/issues/4465.
## Summary
If a package lacks a source distribution, and we can't find a compatible
wheel for the current platform, we need to just _assume_ that the
package will have a valid wheel on all platforms on which it's
requested; if not, we raise an error at install time.
It's possible that we can be smarter about this over time. For example,
if the package was requested _only_ for macOS, we could verify that
there's at least one macOS-compatible wheel. See the linked issue for
more details.
Closes https://github.com/astral-sh/uv/issues/4139.
Extends #4120
Part of #2607
There should be no behavior changes here. Restructures the discovery API
to be focused on a toolchain first perspective in preparation for
exposing a `find_or_fetch` method for toolchains in
https://github.com/astral-sh/uv/pull/4138.
## Summary
Instead of checking if the target and installed version are the same, we
model the data such that the target version is only present if it was
specified by the user. This also means that we correctly say "requested
version" even if the two happen to be the same.
## Summary
This PR removes the static resolver map:
```rust
static RESOLVED_GIT_REFS: Lazy<Mutex<FxHashMap<RepositoryReference, GitSha>>> =
Lazy::new(Mutex::default);
```
With a `GitResolver` struct that we now pass around on the
`BuildContext`. There should be no behavior changes here; it's purely an
internal refactor with an eye towards making it cleaner for us to
"pre-populate" the list of resolved SHAs.
With the change, we remove the special casing of workspace dependencies
and resolve `tool.uv` for all git and directory distributions. This
gives us support for non-editable workspace dependencies and path
dependencies in other workspaces. It removes a lot of special casing
around workspaces. These changes are the groundwork for supporting
`tool.uv` with dynamic metadata.
The basis for this change is moving `Requirement` from
`distribution-types` to `pypi-types` and the lowering logic from
`uv-requirements` to `uv-distribution`. This changes should be split out
in separate PRs.
I've included an example workspace `albatross-root-workspace2` where
`bird-feeder` depends on `a` from another workspace `ab`. There's a
bunch of failing tests and regressed error messages that still need
fixing. It does fix the audited package count for the workspace tests.
## Summary
<!-- What's the purpose of the change? What does it do, and why? -->
It removes the unused result when creating the Cache with the
`from_path` constructor. I don't believe it does any io operations any
more at least.
When parsing requirements from any source, directly parse the url parts
(and reject unsupported urls) instead of parsing url parts at a later
stage. This removes a bunch of error branches and concludes the work
parsing url parts once and passing them around everywhere.
Many usages of the assembled `VerbatimUrl` remain, but these can be
removed incrementally.
Please review commit-by-commit.
## Summary
This seems to be one of the most consistent benchmark cases we have in
terms of standard deviation:
```
➜ hyperfine "target/profiling/main pip compile scripts/requirements/airflow.in" --runs 200
Benchmark 1: target/profiling/main pip compile scripts/requirements/airflow.in
Time (mean ± σ): 292.6 ms ± 6.6 ms [User: 414.1 ms, System: 194.2 ms]
Range (min … max): 282.7 ms … 320.1 ms 200 runs
```
For smaller benchmarks, scispacy and dtlssocket seem to be a bit more
consistent than our current jupyter benchmark, but it hasn't given us
any problems so I'll leave it for now.
## Summary
This PR introduces parallelism to the resolver. Specifically, we can
perform PubGrub resolution on a separate thread, while keeping all I/O
on the tokio thread. We already have the infrastructure set up for this
with the channel and `OnceMap`, which makes this change relatively
simple. The big change needed to make this possible is removing the
lifetimes on some of the types that need to be shared between the
resolver and pubgrub thread.
A related PR, https://github.com/astral-sh/uv/pull/1163, found that
adding `yield_now` calls improved throughput. With optimal scheduling we
might be able to get away with everything on the same thread here.
However, in the ideal pipeline with perfect prefetching, the resolution
and prefetching can run completely in parallel without depending on one
another. While this would be very difficult to achieve, even with our
current prefetching pattern we see a consistent performance improvement
from parallelism.
This does also require reverting a few of the changes from
https://github.com/astral-sh/uv/pull/3413, but not all of them. The
sharing is isolated to the resolver task.
## Test Plan
On smaller tasks performance is mixed with ~2% improvements/regressions
on both sides. However, on medium-large resolution tasks we see the
benefits of parallelism, with improvements anywhere from 10-50%.
```
./scripts/requirements/jupyter.in
Benchmark 1: ./target/profiling/baseline (resolve-warm)
Time (mean ± σ): 29.2 ms ± 1.8 ms [User: 20.3 ms, System: 29.8 ms]
Range (min … max): 26.4 ms … 36.0 ms 91 runs
Benchmark 2: ./target/profiling/parallel (resolve-warm)
Time (mean ± σ): 25.5 ms ± 1.0 ms [User: 19.5 ms, System: 25.5 ms]
Range (min … max): 23.6 ms … 27.8 ms 99 runs
Summary
./target/profiling/parallel (resolve-warm) ran
1.15 ± 0.08 times faster than ./target/profiling/baseline (resolve-warm)
```
```
./scripts/requirements/boto3.in
Benchmark 1: ./target/profiling/baseline (resolve-warm)
Time (mean ± σ): 487.1 ms ± 6.2 ms [User: 464.6 ms, System: 61.6 ms]
Range (min … max): 480.0 ms … 497.3 ms 10 runs
Benchmark 2: ./target/profiling/parallel (resolve-warm)
Time (mean ± σ): 430.8 ms ± 9.3 ms [User: 529.0 ms, System: 77.2 ms]
Range (min … max): 417.1 ms … 442.5 ms 10 runs
Summary
./target/profiling/parallel (resolve-warm) ran
1.13 ± 0.03 times faster than ./target/profiling/baseline (resolve-warm)
```
```
./scripts/requirements/airflow.in
Benchmark 1: ./target/profiling/baseline (resolve-warm)
Time (mean ± σ): 478.1 ms ± 18.8 ms [User: 482.6 ms, System: 205.0 ms]
Range (min … max): 454.7 ms … 508.9 ms 10 runs
Benchmark 2: ./target/profiling/parallel (resolve-warm)
Time (mean ± σ): 308.7 ms ± 11.7 ms [User: 428.5 ms, System: 209.5 ms]
Range (min … max): 287.8 ms … 323.1 ms 10 runs
Summary
./target/profiling/parallel (resolve-warm) ran
1.55 ± 0.08 times faster than ./target/profiling/baseline (resolve-warm)
```
## Summary
This PR consolidates the concurrency limits used throughout `uv` and
exposes two limits, `UV_CONCURRENT_DOWNLOADS` and
`UV_CONCURRENT_BUILDS`, as environment variables.
Currently, `uv` has a number of concurrent streams that it buffers using
relatively arbitrary limits for backpressure. However, many of these
limits are conflated. We run a relatively small number of tasks overall
and should start most things as soon as possible. What we really want to
limit are three separate operations:
- File I/O. This is managed by tokio's blocking pool and we should not
really have to worry about it.
- Network I/O.
- Python build processes.
Because the current limits span a broad range of tasks, it's possible
that a limit meant for network I/O is occupied by tasks performing
builds, reading from the file system, or even waiting on a `OnceMap`. We
also don't limit build processes that end up being required to perform a
download. While this may not pose a performance problem because our
limits are relatively high, it does mean that the limits do not do what
we want, making it tricky to expose them to users
(https://github.com/astral-sh/uv/issues/1205,
https://github.com/astral-sh/uv/issues/3311).
After this change, the limits on network I/O and build processes are
centralized and managed by semaphores. All other tasks are unbuffered
(note that these tasks are still bounded, so backpressure should not be
a problem).
We now use the getters and setters everywhere.
There were some places where we wanted to build a `MarkerEnvironment`
out of whole cloth, usually in tests. To facilitate those use cases, we
add a `MarkerEnvironmentBuilder` that provides a convenient constructor.
It's basically like a `MarkerEnvironment::new`, but with named
parameters. That's useful here because there are so many fields (and
they many have the same type).
This commit touches a lot of code, but the conceptual change here is
pretty simple: make it so we can run the resolver without providing a
`MarkerEnvironment`. This also indicates that the resolver should run in
universal mode. That is, the effect of a missing marker environment is
that all marker expressions that reference the marker environment are
evaluated to `true`. That is, they are ignored. (The only markers we
evaluate in that context are extras, which are the only markers that
aren't dependent on the environment.)
One interesting change here is that a `Resolver` no longer needs an
`Interpreter`. Previously, it had only been using it to construct a
`PythonRequirement`, by filling in the installed version from the
`Interpreter` state. But we now construct a `PythonRequirement`
explicitly since its `target` Python version should no longer be tied to
the `MarkerEnvironment`. (Currently, the marker environment is mutated
such that its `python_full_version` is derived from multiple sources,
including the CLI, which I found a touch confusing.)
The change in behavior can now be observed through the
`--unstable-uv-lock-file` flag. First, without it:
```
$ cat requirements.in
anyio>=4.3.0 ; sys_platform == "linux"
anyio<4 ; sys_platform == "darwin"
$ cargo run -qp uv -- pip compile -p3.10 requirements.in
anyio==4.3.0
exceptiongroup==1.2.1
# via anyio
idna==3.7
# via anyio
sniffio==1.3.1
# via anyio
typing-extensions==4.11.0
# via anyio
```
And now with it:
```
$ cargo run -qp uv -- pip compile -p3.10 requirements.in --unstable-uv-lock-file
x No solution found when resolving dependencies:
`-> Because you require anyio>=4.3.0 and anyio<4, we can conclude that the requirements are unsatisfiable.
```
This is expected at this point because the marker expressions are being
explicitly ignored, *and* there is no forking done yet to account for
the conflict.
## Introduction
PEP 621 is limited. Specifically, it lacks
* Relative path support
* Editable support
* Workspace support
* Index pinning or any sort of index specification
The semantics of urls are a custom extension, PEP 440 does not specify
how to use git references or subdirectories, instead pip has a custom
stringly format. We need to somehow support these while still stying
compatible with PEP 621.
## `tool.uv.source`
Drawing inspiration from cargo, poetry and rye, we add `tool.uv.sources`
or (for now stub only) `tool.uv.workspace`:
```toml
[project]
name = "albatross"
version = "0.1.0"
dependencies = [
"tqdm >=4.66.2,<5",
"torch ==2.2.2",
"transformers[torch] >=4.39.3,<5",
"importlib_metadata >=7.1.0,<8; python_version < '3.10'",
"mollymawk ==0.1.0"
]
[tool.uv.sources]
tqdm = { git = "https://github.com/tqdm/tqdm", rev = "cc372d09dcd5a5eabdc6ed4cf365bdb0be004d44" }
importlib_metadata = { url = "https://github.com/python/importlib_metadata/archive/refs/tags/v7.1.0.zip" }
torch = { index = "torch-cu118" }
mollymawk = { workspace = true }
[tool.uv.workspace]
include = [
"packages/mollymawk"
]
[tool.uv.indexes]
torch-cu118 = "https://download.pytorch.org/whl/cu118"
```
See `docs/specifying_dependencies.md` for a detailed explanation of the
format. The basic gist is that `project.dependencies` is what ends up on
pypi, while `tool.uv.sources` are your non-published additions. We do
support the full range or PEP 508, we just hide it in the docs and
prefer the exploded table for easier readability and less confusing with
actual url parts.
This format should eventually be able to subsume requirements.txt's
current use cases. While we will continue to support the legacy `uv pip`
interface, this is a piece of the uv's own top level interface. Together
with `uv run` and a lockfile format, you should only need to write
`pyproject.toml` and do `uv run`, which generates/uses/updates your
lockfile behind the scenes, no more pip-style requirements involved. It
also lays the groundwork for implementing index pinning.
## Changes
This PR implements:
* Reading and lowering `project.dependencies`,
`project.optional-dependencies` and `tool.uv.sources` into a new
requirements format, including:
* Git dependencies
* Url dependencies
* Path dependencies, including relative and editable
* `pip install` integration
* Error reporting for invalid `tool.uv.sources`
* Json schema integration (works in pycharm, see below)
* Draft user-level docs (see `docs/specifying_dependencies.md`)
It does not implement:
* No `pip compile` testing, deprioritizing towards our own lockfile
* Index pinning (stub definitions only)
* Development dependencies
* Workspace support (stub definitions only)
* Overrides in pyproject.toml
* Patching/replacing dependencies
One technically breaking change is that we now require user provided
pyproject.toml to be valid wrt to PEP 621. Included files still fall
back to PEP 517. That means `pip install -r requirements.txt` requires
it to be valid while `pip install -r requirements.txt` with `-e .` as
content falls back to PEP 517 as before.
## Implementation
The `pep508` requirement is replaced by a new `UvRequirement` (name up
for bikeshedding, not particularly attached to the uv prefix). The still
existing `pep508_rs::Requirement` type is a url format copied from pip's
requirements.txt and doesn't appropriately capture all features we
want/need to support. The bulk of the diff is changing the requirement
type throughout the codebase.
We still use `VerbatimUrl` in many places, where we would expect a
parsed/decomposed url type, specifically:
* Reading core metadata except top level pyproject.toml files, we fail a
step later instead if the url isn't supported.
* Allowed `Urls`.
* `PackageId` with a custom `CanonicalUrl` comparison, instead of
canonicalizing urls eagerly.
* `PubGrubPackage`: We eventually convert the `VerbatimUrl` back to a
`Dist` (`Dist::from_url`), instead of remembering the url.
* Source dist types: We use verbatim url even though we know and require
that these are supported urls we can and have parsed.
I tried to make improve the situation be replacing `VerbatimUrl`, but
these changes would require massive invasive changes (see e.g.
https://github.com/astral-sh/uv/pull/3253). A main problem is the ref
`VersionOrUrl` and applying overrides, which assume the same
requirement/url type everywhere. In its current form, this PR increases
this tech debt.
I've tried to split off PRs and commits, but the main refactoring is
still a single monolith commit to make it compile and the tests pass.
## Demo
Adding
d1ae3b85d5/pyproject.json
as json schema (v7) to pycharm for `pyproject.toml`, you can try the IDE
support already:
[dove.webm](c293c272-c80b-459d-8c95-8c46a8d198a1)
Scott schafer got me the idea: We can avoid repeating the path for
workspaces dependencies everywhere if we declare them in the virtual
package once and treat them as workspace dependencies from there on.
First, replace all usages in files in-place. I used my editor for this.
If someone wants to add a one-liner that'd be fun.
Then, update directory and file names:
```
# Run twice for nested directories
find . -type d -print0 | xargs -0 rename s/puffin/uv/g
find . -type d -print0 | xargs -0 rename s/puffin/uv/g
# Update files
find . -type f -print0 | xargs -0 rename s/puffin/uv/g
```
Then add all the files again
```
# Add all the files again
git add crates
git add python/uv
# This one needs a force-add
git add -f crates/uv-trampoline
```
This PR tweaks the representation of `Tags` in order to offer a
faster implementation of `WheelFilename::is_compatible`. We now use a
nested map of tags that lets us avoid looping over every supported
platform tag. As the code comments suggest, that is the essential gain.
We still do not mind looping over the tags in each wheel name since they
tend to be quite small. And pushing our thumb on that side of things can
make things worse overall since it would likely slow down WheelFilename
construction itself.
For micro-benchmarks, we improve considerably for compatibility
checking:
$ critcmp base test3
group base test3
----- ---- -----
build_platform_tags/burntsushi-archlinux 1.00 46.2±0.28µs ? ?/sec 2.48
114.8±0.45µs ? ?/sec
wheelname_parsing/flyte-long-compatible 1.00 624.8±3.31ns 174.0 MB/sec
1.01 629.4±4.30ns 172.7 MB/sec
wheelname_parsing/flyte-long-incompatible 1.00 743.6±4.23ns 165.4 MB/sec
1.00 746.9±4.62ns 164.7 MB/sec
wheelname_parsing/flyte-short-compatible 1.00 526.7±4.76ns 54.3 MB/sec
1.01 530.2±5.81ns 54.0 MB/sec
wheelname_parsing/flyte-short-incompatible 1.00 540.4±4.93ns 60.0 MB/sec
1.01 545.7±5.31ns 59.4 MB/sec
wheelname_parsing_failure/flyte-long-extension 1.00 13.6±0.13ns 3.2
GB/sec 1.01 13.7±0.14ns 3.2 GB/sec
wheelname_parsing_failure/flyte-short-extension 1.00 14.0±0.20ns 1160.4
MB/sec 1.01 14.1±0.14ns 1146.5 MB/sec
wheelname_tag_compatibility/flyte-long-compatible 11.33 159.8±2.79ns
680.5 MB/sec 1.00 14.1±0.23ns 7.5 GB/sec
wheelname_tag_compatibility/flyte-long-incompatible 237.60
1671.8±37.99ns 73.6 MB/sec 1.00 7.0±0.08ns 17.1 GB/sec
wheelname_tag_compatibility/flyte-short-compatible 16.07 223.5±8.60ns
128.0 MB/sec 1.00 13.9±0.30ns 2.0 GB/sec
wheelname_tag_compatibility/flyte-short-incompatible 149.83 628.3±2.13ns
51.6 MB/sec 1.00 4.2±0.10ns 7.6 GB/sec
We do regress slightly on the time it takes for `Tags::new` to run, but
this is somewhat expected. And in absolute terms, 114us is perfectly
acceptable given that it's only executed ~once for each `puffin`
invocation.
Ad hoc benchmarks indicate an overall 25% perf improvement in `puffin
pip-compile` times. This roughly corresponds with how much time
`is_compatible` was taking. Indeed, profiling confirms that it has
virtually disappeared from the profile.
Fixes#157
