PIP_COMPATIBILITY.md

Compatibility with pip and pip-tools

uv is designed as a drop-in replacement for common pip and pip-tools workflows.

Informally, the intent is such that existing pip and pip-tools users can switch to uv without making meaningful changes to their packaging workflows; and, in most cases, swapping out pip install for uv pip install should "just work".

However, uv is not intended to be an exact clone of pip, and the further you stray from common pip workflows, the more likely you are to encounter differences in behavior. In some cases, those differences may be known and intentional; in others, they may be the result of implementation details; and in others, they may be bugs.

This document outlines the known differences between uv and pip, along with rationale, workarounds, and a statement of intent for compatibility in the future.

Configuration files and environment variables

uv does not read configuration files or environment variables that are specific to pip, like pip.conf or PIP_INDEX_URL.

Reading configuration files and environment variables intended for other tools has a number of drawbacks:

1. It requires bug-for-bug compatibility with the target tool, since users end up relying on bugs in the format, the parser, etc.
2. If the target tool changes the format in some way, uv is then locked-in to changing it in equivalent ways.
3. If that configuration is versioned in some way, uv would need to know which version of the target tool the user is expecting to use.
4. It prevents uv from introducing any settings or configuration that don't exist in the target tool, since otherwise pip.conf (or similar) would no longer be usable with pip.
5. It can lead to user confusion, since uv would be reading settings that don't actually affect its behavior, and many users may not expect uv to read configuration files intended for other tools.

Instead, uv supports its own environment variables, like UV_INDEX_URL. In the future, uv will also support persistent configuration in its own configuration file format (e.g., pyproject.toml or uv.toml or similar). For more, see #651.

Pre-release compatibility

By default, uv will accept pre-release versions during dependency resolution in two cases:

1. If the package is a direct dependency, and its version markers include a pre-release specifier (e.g., flask>=2.0.0rc1).
2. If all published versions of a package are pre-releases.

If dependency resolution fails due to a transitive pre-release, uv will prompt the user to re-run with --prerelease=allow, to allow pre-releases for all dependencies.

Alternatively, you can add the transitive dependency to your requirements.in file with pre-release specifier (e.g., flask>=2.0.0rc1) to opt in to pre-release support for that specific dependency.

In sum, uv needs to know upfront whether the resolver should accept pre-releases for a given package. pip, meanwhile, may respect pre-release identifiers in transitive dependencies depending on the order in which the resolver encounters the relevant specifiers (#1641).

Pre-releases are notoriously difficult to model, and are a frequent source of bugs in packaging tools. Even pip, which is viewed as a reference implementation, has a number of open questions around pre-release handling (#12469, #12470, #40505, etc.). uv's pre-release handling is intentionally limited and intentionally requires user opt-in for pre-releases, to ensure correctness.

In the future, uv may support pre-release identifiers in transitive dependencies. However, it's likely contingent on evolution in the Python packaging specifications. The existing PEPs do not cover "dependency resolution" and are instead focused on behavior for a single version specifier. As such, there are unresolved questions around the correct and intended behavior for pre-releases in the packaging ecosystem more broadly.

Local version identifiers

uv does not implement spec-compliant handling of local version identifiers (e.g., 1.2.3+local). This is considered a known limitation. Although local version identifiers are rare in published packages (and, e.g., disallowed on PyPI), they're common in the PyTorch ecosystem, and uv's approach to local versions does support typical PyTorch workflows to succeed out-of-the-box.

PEP 440 specifies that the local version segment should typically be ignored when evaluating version specifiers, with a few exceptions. For example, foo==1.2.3 should accept 1.2.3+local, but foo==1.2.3+local should not accept 1.2.3. These asymmetries are hard to model in a resolution algorithm. As such, uv treats 1.2.3 and 1.2.3+local as entirely separate versions, but respects local versions provided as direct dependencies throughout the resolution, such that if you provide foo==1.2.3+local as a direct dependency, 1.2.3+local will be accepted for any transitive dependencies that request foo==1.2.3.

To take an example from the PyTorch ecosystem, it's common to specify torch==2.0.0+cu118 and torchvision==0.15.1+cu118 as direct dependencies. torchvision @ 0.15.1+cu118 declares a dependency on torch==2.0.0. In this case, uv would recognize that torch==2.0.0+cu118 satisfies the specifier, since it was provided as a direct dependency.

As compared to pip, the main differences in observed behavior are as follows:

• In general, local versions must be provided as direct dependencies. Resolution may succeed for transitive dependencies that request a non-local version, but this is not guaranteed.
• If only local versions exist for a package foo at a given version (e.g., 1.2.3+local exists, but 1.2.3 does not), uv pip install foo==1.2.3 will fail, while pip install foo==1.2.3 will resolve to an arbitrary local version.

Packages that exist on multiple indexes

In both uv and pip, users can specify multiple package indexes from which to search for the available versions of a given package. However, uv and pip differ in how they handle packages that exist on multiple indexes.

For example, imagine that a company publishes an internal version of requests on a private index (--extra-index-url), but also allows installing packages from PyPI by default. In this case, the private requests would conflict with the public requests on PyPI.

When uv searches for a package across multiple indexes, it will iterate over the indexes in order (preferring the --extra-index-url over the default index), and stop searching as soon as it finds a match. This means that if a package exists on multiple indexes, uv will limit its candidate versions to those present in the first index that contains the package.

pip, meanwhile, will combine the candidate versions from all indexes, and select the best version from the combined set, though it makes no guarantees around the order in which it searches indexes, and expects that packages are unique up to name and version, even across indexes.

uv's behavior is such that if a package exists on an internal index, it should always be installed from the internal index, and never from PyPI. The intent is to prevent "dependency confusion" attacks, in which an attacker publishes a malicious package on PyPI with the same name as an internal package, thus causing the malicious package to be installed instead of the internal package. See, for example, the torchtriton attack from December 2022.

As of v0.1.29, users can opt in to pip-style behavior for multiple indexes via the --index-strategy unsafe-any-match command-line option, or the UV_INDEX_STRATEGY environment variable. When enabled, uv will search for each package across all indexes, and consider all available versions when resolving dependencies, prioritizing the --extra-index-url indexes over the default index URL. (Versions that are duplicated across indexes will be ignored.)

In the future, uv will support pinning packages to dedicated indexes (see: #171). Additionally, PEP 708 is a provisional standard that aims to address the "dependency confusion" issue across package registries and installers.

Transitive direct URL dependencies for constraints and overrides

While uv does support URL dependencies (e.g., black @ https://...), it does not support transitive (i.e., "nested") direct URL dependencies for constraints and overrides.

Specifically, if a constraint or override is defined using a direct URL dependency, and the constrained package has a direct URL dependency of its own, uv may reject that transitive direct URL dependency during resolution.

uv also makes the assumption that non-URL dependencies won't introduce URL dependencies (i.e., that dependencies fetched from a registry will not themselves have direct URL dependencies). If a non-URL dependency does introduce a URL dependency, uv will reject the URL dependency during resolution.

If uv rejects a transitive URL dependency in either case, the best course of action is to provide the URL dependency as a direct dependency in the requirements.in file, rather than as a constraint, override, or transitive dependency.

Virtual environments by default

uv pip install and uv pip sync are designed to work with virtual environments by default.

Specifically, uv will always install packages into the currently active virtual environment, or search for a virtual environment named .venv in the current directory or any parent directory (even if it is not activated).

This differs from pip, which will install packages into a global environment if no virtual environment is active, and will not search for inactive virtual environments.

In uv, you can install into non-virtual environments by providing a path to a Python executable via the --python /path/to/python option, or via the --system flag, which installs into the first Python interpreter found on the PATH, like pip.

In other words, uv inverts the default, requiring explicit opt-in to installing into the system Python, which can lead to breakages and other complications, and should only be done in limited circumstances.

For more, see "Installing into arbitrary Python environments".

Resolution strategy

For a given set of dependency specifiers, it's often the case that there is no single "correct" set of packages to install. Instead, there are many valid sets of packages that satisfy the specifiers.

Neither pip nor uv make any guarantees about the exact set of packages that will be installed; only that the resolution will be consistent, deterministic, and compliant with the specifiers. As such, in some cases, pip and uv will yield different resolutions; however, both resolutions should be equally valid.

For example, consider:

# requirements.txt
starlette
fastapi

At time of writing, the most recent starlette version is 0.37.2, and the most recent fastapi version is 0.110.0. However, fastapi==0.110.0 also depends on starlette, and introduces an upper bound: starlette>=0.36.3,<0.37.0.

If a resolver prioritizes including the most recent version of starlette, it would need to use an older version of fastapi that excludes the upper bound on starlette. In practice, this requires falling back to fastapi==0.1.17:

# This file was autogenerated by uv via the following command:
#    uv pip compile -
annotated-types==0.6.0
    # via pydantic
anyio==4.3.0
    # via starlette
fastapi==0.1.17
idna==3.6
    # via anyio
pydantic==2.6.3
    # via fastapi
pydantic-core==2.16.3
    # via pydantic
sniffio==1.3.1
    # via anyio
starlette==0.37.2
    # via fastapi
typing-extensions==4.10.0
    # via
    #   pydantic
    #   pydantic-core

Alternatively, if a resolver prioritizes including the most recent version of fastapi, it would need to use an older version of starlette that satisfies the upper bound. In practice, this requires falling back to starlette==0.36.3:

#    uv pip compile -
annotated-types==0.6.0
    # via pydantic
anyio==4.3.0
    # via starlette
fastapi==0.110.0
idna==3.6
    # via anyio
pydantic==2.6.3
    # via fastapi
pydantic-core==2.16.3
    # via pydantic
sniffio==1.3.1
    # via anyio
starlette==0.36.3
    # via fastapi
typing-extensions==4.10.0
    # via
    #   fastapi
    #   pydantic
    #   pydantic-core

When uv resolutions differ from pip in undesirable ways, it's often a sign that the specifiers are too loose, and that the user should consider tightening them. For example, in the case of starlette and fastapi, the user could require fastapi>=0.110.0.

pip check

At present, uv pip check will surface the following diagnostics:

• A package has no METADATA file, or the METADATA file can't be parsed.
• A package has a Requires-Python that doesn't match the Python version of the running interpreter.
• A package has a dependency on a package that isn't installed.
• A package has a dependency on a package that's installed, but at an incompatible version.
• Multiple versions of a package are installed in the virtual environment.

In some cases, uv pip check will surface diagnostics that pip check does not, and vice versa. For example, unlike uv pip check, pip check will not warn when multiple versions of a package are installed in the current environment.

Strictness and spec enforcement

uv tends to be stricter than pip, and will often reject packages that pip would install. For example, uv omits packages with invalid version specifiers in its metadata, which pip similarly plans to exclude in a future release.

In some cases, uv implements lenient behavior for popular packages that are known to have specific spec compliance issues.

If uv rejects a package that pip would install due to a spec violation, the best course of action is to first attempt to install a newer version of the package; and, if that fails, to report the issue to the package maintainer.

pip command-line options and subcommands

uv does not support the complete set of pip's command-line options and subcommands, although it does support a large subset.

Missing options and subcommands are prioritized based on user demand and the complexity of the implementation, and tend to be tracked in individual issues. For example:

• --trusted-host
• --user

If you encounter a missing option or subcommand, please search the issue tracker to see if it has already been reported, and if not, consider opening a new issue. Feel free to upvote any existing issues to convey your interest.

Registry authentication

uv does not support pip's auto or import options for --keyring-provider. At present, only the subproces option is supported.

Unlike pip, uv does not enable keyring authentication by default.

Unlike pip, uv does not wait until a request returns a HTTP 401 before searching for authentication. uv attaches authentication to all requests for hosts with credentials available.

Legacy features

uv does not support features that are considered legacy or deprecated in pip. For example, uv does not support .egg-style distributions.

uv does not plan to support features that the pip maintainers explicitly recommend against, like --target.