Following system colour scheme Selected dark colour scheme Selected light colour scheme

Python Enhancement Proposals

PEP 759 – External Wheel Hosting

Author:
Barry Warsaw <barry at python.org>, Ethan Smith <ethan at ethanhs.me>
PEP-Delegate:
Donald Stufft <donald at python.org>
Discussions-To:
Discourse thread
Status:
Draft
Type:
Standards Track
Topic:
Packaging
Created:
01-Oct-2024
Post-History:

Table of Contents

Abstract

This PEP proposes a mechanism by which projects hosted on pypi.org can safely host wheel artifacts on external sites other than PyPI. This PEP explicitly does not propose external hosting of projects, packages, or their metadata. That functionality is already available by externally hosting independent package indexes. Because this PEP only provides a mechanism for projects to customize the download URL for specific released wheel artifacts, dependency resolution as already implemented by common installer tools such as pip and uv does not need to change.

This PEP defines what it means to be “safe” in this context, along with a new package upload file format called a .rim file. It defines how .rim files affect the metadata returned for a package’s Simple Repository API in both HTML and JSON formats, and how traditional wheels can easily be turned into .rim files.

Rationale

The Python Package Index, hosted at https://pypi.org, imposes default limits on upload artifact file size (100 MiB) and total project size (10 GiB). Most projects can comfortably fit within these limits during the lifetime of the project, through years of uploads. A few projects have encountered these limits, and have been granted both file size and project size exceptions, allowing them to continue uploading new releases without having to take more drastic measures, such as removing files which may potentially still be in use by consumers (e.g. through version pins).

A related workaround is the “wheel stub” approach, which provides an indirect link between PyPI and an external third party package index, where such limitations can be avoided. Wheel stubs are source distributions (a.k.a. “sdists”) which utilize a PEP 517 build backend that, instead of turning source code into a binary wheel, performs some logic to calculate the URL for an existing, externally hosted wheel to download and install. This approach works, but it obscures the connection between PyPI, the sdist, and the externally hosted wheel, since there is no way to present this information to pip or other such tools.

Historical context

In 2013, PEP 438 proposed a “backward-compatible two-phase transition process” to modify several aspects of release file hosting on PyPI. As this PEP describes, PyPI originally supported only project and release registration without also allowing for artifact file hosting. As such, most projects hosted release file artifacts elsewhere. Artifact hosting was later added, but the mix of externally and PyPI-hosted files led to a wide range of usability and potential security-related problems. PEP 438 was an attempt to provide several facilities to allow external hosting while promoting a PyPI-first hosting preference.

PEP 438 was complex, with three different “hosting modes”, rel metadata in the simple HTML index pages to signify hosting locations, and a two-phase transition plan affecting PyPI and installer tools. PEP 438 was ultimately retracted in 2015 by PEP 470, which acknowledges that PEP 438 did succeed in…

bringing about more people to utilize PyPI’s repository features, an altogether good thing given the global CDN powering PyPI providing speed ups for a lot of people[…]

Instead of external hosting, PEP 470 promoted the use of explicit multiple repositories, providing full package indexing and artifact hosting, and enabled through installer tool support, such as pip install --extra-index-url allowing pip to essentially treat multiple repositories as one single global repository for package installation resolution. Because this has been the blessed norm for so many years, all Python package installation tools support querying multiple indexes for dependency resolution.

The problem with multiple indexes

Why then does this PEP propose to allow a more limited form of external hosting, and how does this proposal avoid the problems documented in PEP 470?

One well-known problem that consolidating multiple indexes enables is dependency confusion attacks, to which Python can be particularly vulnerable, due to the algorithm that pip install uses for resolving package dependencies and preferred versions. The uv tool addresses this by supporting an additional index strategy option, whereby users can select between, e.g. a pip-compatible strategy, and a more limited strategy that prevents such dependency confusion attacks.

PEP 708 provides additional background about dependency confusion attacks, and takes a different approach to preventing them. At its core, PEP 708 allows repository owners to indicate that projects track across different repositories, which allows installers to determine how to treat the global package namespace when combined across multiple repositories. PEP 708 has been provisionally accepted, pending several required conditions as outlined in PEP 708, some of which may have an indeterminate future. As PEP 708 itself says, this won’t by itself solve dependency confusion attacks, but is one way to provide enough information to installers to help minimize these attacks.

While there can still be valid use cases for standing up a totally independent package index (such as providing richer platform support for GPUs until a fully formed variant proposal is accepted), this PEP takes a different, simpler approach and doesn’t replace any of the existing, proposed, or approved package index cooperation specifications.

This PEP also preserves the core purpose of PyPI, and allows it to remain the traditional, canonical, centralized index of all Python packages.

Addressing PyPI limits

This proposal also addresses the problem of size limits imposed by PyPI, where there is a default artifact size limit of 100 MiB and a default overall project size limit of 10 GiB. Most packages and artifacts can easily fit in these limits, even for packages containing binary extension modules for a variety of platforms. A small, but important class of packages routinely exceed these limits, requiring them to submit PyPI exception request support tickets. It’s not necessarily difficult to get resolution on such exceptions, but it is a special process that can take some time to resolve, and the criteria for granting such exceptions aren’t well documented.

Reducing operational complexity

Setting up and maintaining an entire package index can be a complex operational solution, both time and resource intensive. This is especially true if the main purpose of such an index is just to avoid file size limitations. The external index approach also imposes a tricky UX on consumers of projects on the external index, requiring them to understand how CLI options such as --external-index-url work, along with the security implications of such flags. It would be much easier for both producers and consumers of large wheel packages to just set up and maintain a simple web server, capable of serving individual files with no more complex API than HTTP GET. Such an interface is also easily cacheable or placed behind a CDN. Simple HTTP servers are also much easier to audit for security purposes, easier to proxy, and usually take much less resources to run, support, and maintain. Even something like Amazon S3 could be used to host external wheels.

This PEP proposes an approach that favors such operational simplicity.

Specification

A new type of uploadable file is defined, called a “RIM” (i.e. .rim), or “Remote Installable Metadata” file. The name evokes the image of a wheel with the tire removed, and emphasizes that .rim files are easily derived from .whl files. The process of turning a .whl into a .rim is outlined below. The file name format exactly matches the wheel file naming format specification, except that RIM files use the suffix .rim. This means that all the tags used to discriminate .whl files also distinguish between different .rim files, and thus can be used during dependency resolution steps, exactly as .whl files are today. In this respect, .whl and .rim files are interchangeable.

The content of a .rim file is nearly identical to .whl files, however .rim files MUST contain only the .dist-info directory from a wheel. No other top-level file or directory is allowed in the .rim zip file. The .dist-info directory MUST contain a single additional file in addition to those allowed in a .whl file’s .dist-info directory: a file called EXTERNAL-HOSTING.json.

This is a JSON file contains containing the following keys:

version
This is the file format version, which for this PEP MUST be 1.0.
owner
This MUST name the PyPI organization owner of this externally hosted file, for reasons which will be described in detail below.
uri
This is a single URL naming the location of the physical .whl file hosted on an external site. This URL MUST use the https scheme.
size
This is an integer value describing the size in bytes of the physical .whl file on the remote host.
hashes
This is a dictionary of the format described in PEP 694, used to capture both the PEP 694 of the physical .whl file, with the same constraints as proposed in that PEP. Since these hashes are immutable once uploaded to PyPI, they serve as a critical validation that the externally hosted wheel hasn’t been corrupted or compromised.

Effects of the RIM file

The only effect of a .rim file is to change the download URL for the wheel artifact in both the HTML and JSON interfaces in the simple repository API. In the HTML page for a package release, the href attribute MUST be the value of the uri key, including a #<hashname>=<hashvalue> fragment. this hash fragment MUST be in exactly the same format as described the PEP 376 originated signed wheel file format in the .dist-info/RECORD file. The exact same rules for selection of hash algorithm and encoding is used here.

Similarly in the JSON response the url key pointing to the download file must be the value of the uri key, and the hashes dictionary MUST be included with values populated from the hashes dictionary provided above.

In all other respects, a compliant package index should treat .rim files the same as .whl files, with some other minor exceptions as outlined below. For example, .rim files can be deleted and yanked (PEP 592) just like any .whl file, with the exact same semantics (i.e. deletions are permanent). When a .rim is deleted, an index MUST NOT allow a matching .whl or .rim file to be (re-)uploaded.

Availability order

Externally hosted wheels MUST be available before the corresponding .rim file is uploaded to PyPI, otherwise a publishing race condition is introduced, although this requirement MAY be relaxed for .rim files uploaded to a PEP 694 staged release.

Wheels can override RIMs

Indexes MUST reject .rim files if a matching .whl file already exists with the exact same file name tags. However, indexes MAY accept a .whl file if a matching .rim file exists, as long as that .rim file hasn’t been deleted or yanked. This allows uploaders to replace an externally hosted wheel file with an index hosted wheel file, but the converse is prohibited. Since the default is to host wheels on the same package index that contains the package metadata, it is not allowed to “downgrade” an existing wheel file once uploaded. When a .whl replaces a .rim, the index MUST provide download URLs for the package using its own hosted file service. When uploading the overriding .whl file, the package index MUST validate the hash from the existing .rim file, and these hashes must match or the overriding upload MUST be rejected.

PyPI API bump unnecessary

It’s likely that the changes are backward compatible enough that a bump in the PyPI repository version is not necessary. Since .rim files are essentially changes only to the upload API, package resolvers and package installers can continue to function with the APIs they’ve always supported.

External hosting resiliency

One of the key concerns leading to PEP 438’s revocation in PEP 470 was potential user confusion when an external index disappeared. From PEP 470:

This confusion comes down to end users of projects not realizing if a project is hosted on PyPI or if it relies on an external service. This often manifests itself when the external service is down but PyPI is not. People will see that PyPI works, and other projects works, but this one specific one does not. They oftentimes do not realize who they need to contact in order to get this fixed or what their remediation steps are.

While the problem of external wheel hosting service going down is not directly solved by this PEP, several safeguards are in place to greatly reduce the potential burden on PyPI administrators.

This PEP thus proposes that:

  • External wheel hosting is only allowed for packages which are owned by organization accounts. External hosting is an organization-wide setting.
  • Organization accounts do not automatically gain the ability to externally host wheels; this feature MUST be explicitly enabled by PyPI admins at their discretion. Since this will not be a common request, we don’t expect the overhead to be nearly as burdensome as PEP 541 resolutions, account recovery requests, or even file/project size increase requests. External hosting requests would be handled in the same manner as those requests, i.e. via the PyPI GitHub support tracker.
  • Organization accounts requesting external wheel hosting MUST register their own support contact URI, be it a mailto URI for a contact email address, or the URL to the organization’s support tracker. Such a contact URI is optional for organizations which do not avail themselves of external wheel file hosting.

Combined with the EXTERNAL-HOSTING.json file’s owner key, this allows for installer tools to unambiguously redirect any download errors away from the PyPI support admins and squarely to the organization’s support admins.

While the exact mechanics of storing and retrieving this organization support URL will be defined separately, for the sake of example, let’s say a package foo externally hosts wheel files on `https://foo.example.com <https://foo.example.com>`__ and that host becomes unreachable. When an installer tool tries to download and install the package foo wheel, the download step will fail. The installer would then be able to query PyPI to provide a useful error message to the end user:

  • The installer downloads the .rim file and reads the owner key from the EXTERNAL-HOSTING.json file inside the .rim zip file.
  • The installer queries PyPI for the support URI for the organization owner of the externally hosted wheel.
  • An informative error message would then be displayed, e.g.:
    The externally hosted wheel file foo-....whl could not be downloaded. Please contact support@foo.example.com for help. Do not report this to the PyPI administrators.

Dismounting wheels

It is generally very easy to produce a .rim file from an existing .whl file. This could be done efficiently by a PEP 518 build backend with an additional command line option, or a separate tool which takes a .whl file as input and creates the associated .rim file. To complete the analogy, the act of turning a .whl into a .rim is called “dismounting”. The steps such a tool would take are:

  • Accept as input the source .whl file, the organization owner of the package, and URL at which the .whl will be hosted, and the support URI to report download problems from. These could in fact be captured in the pyproject.toml file, but that specification is out of scope for this PEP.
  • Unzip the .whl and create the .rim zip archive.
  • Omit from the .rim file any path in the .whl that isn’t rooted at the .dist-info directory.
  • Calculate the hash of the source .whl file.
  • Add the EXTERNAL-HOSTING.json file containing the JSON keys and values as described above, to the .rim archive.

Changes to tools

Theoretically, installer tools shouldn’t need any changes, since when they have identified the wheel to download and install, they simply consult the download URLs returned by PyPI’s Simple API. In practice though, tools such as pip and uv may have constrained lists of hosts they will allow downloads from, such as PyPI’s own pythonhosted.org domain.

In this case, such tools will need to relax those constraints, but the exact policy for this is left to the installer tools themselves. Any number of approaches could be implemented, such as downloading the .rim file and verifying the EXTERNAL-HOSTING.json metadata, or simply trusting the external downloads for any wheel with a matching checksum. They could also query PyPI for the project’s organization owner and support URI before trusting the download. They could warn the user when externally hosted wheel files are encountered, and/or require the use of a command line option to enable additional download hosts. Any of these verification policies could be chosen in configuration files.

Installer tools should also probably provide better error messages when externally hosted wheels cannot be downloaded, e.g. because a host is unreachable. As described above, such tools could query enough metadata from PyPI to provide clear and distinct error messages pointing users to the package’s external hosting support email or issue tracker.

Constraints for external hosting services

The following constraints lead to reliable and compatible external wheel hosting services:

  • External wheels MUST be served over HTTPS, with a certificate signed by Mozilla’s root certificate store. This ensures compatibility with pip and uv. At the time of this writing, pip 24.2 on Python 3.10 or newer uses the system certificate store in addition to the Mozilla store provided by the third party certifi Python package. uv uses the Mozilla store provided by the webpki-roots crate, but not the system store unless the --native-tls flag is given [1]. The PyPI administrators may modify this requirement in the future, but compatibility with popular installers will not be compromised.
  • External wheel hosts SHOULD use a content delivery network (CDN), just as PyPI does.
  • External wheel hosts MUST commit to a stable URL for all wheels they host.
  • Externally hosted wheels MUST NOT be removed from an external wheel host unless the corresponding .rim file is deleted from PyPI first, and MUST NOT remove external wheels for yanked releases.
  • External wheel hosts MUST support HTTP range requests.
  • External wheel hosts SHOULD support the HTTP/2 protocol.

Security

Several factors as described in this proposal should mitigate security concerns with externally hosted wheels, such as:

  • Wheel file checksums MUST be included in .rim files, and once uploaded cannot be changed. Since the checksum stored on PyPI is immutable and required, it is not possible to spoof an external wheel file, even if the owning organization lost control of their hosting domain.
  • Externally hosted wheels MUST be served over HTTPS.
  • In order to serve externally hosted wheels, organizations MUST be approved by the PyPI admins.

When users identify malware or vulnerabilities in PyPI-hosted projects, they can now report this using the malware reporting facilities on PyPI, as also described in this blog post. The same process can be used to report security issues in externally hosted wheels, and the same remediation process should be used. In addition, since organizations with external hosting enabled MUST provide a support contact URI, that URI can be used in some cases to report the security issue to the hosting organization. Such organization reporting won’t make sense for malware, but could indeed be a very useful way to report security vulnerabilities in externally hosted wheels.

Rejected ideas

Several ideas were considered and rejected.

  • Requiring digital signatures on externally hosted wheel files, either in addition to or other than hashes. We deem this unnecessary since the checksum requirement should be enough to validate that the metadata on PyPI for a wheel exactly matches the downloaded wheel. The added complexity of key management outweighs any additional benefit such digital signatures might convey.
  • Hash verification on .rim file uploads. PyPI could verify that the hash in the uploaded .rim file matches the externally hosted wheel before it accepts the upload, but this requires downloading the external wheel and performing the checksum, which also implies that the upload of the .rim file cannot be accepted until this external .whl file is downloaded and verified. This increases PyPI bandwidth and slows down the upload query, although PEP 694 draft uploads could potentially mitigate these concerns. Still, the benefit is not likely worth the additional complexity.
  • Periodic verification of the download URLs by the index. PyPI could try to periodically ensure that the external wheel host or the external .whl file itself is still available, e.g. via an HTTP HEAD request. This is likely overkill and without also providing the file’s checksum in the response [2], may not provide much additional benefit.
  • This PEP could allow for an organization to provide fallback download hosts, such that a secondary is available if the primary goes down. We believe that DNS-based replication is a much better, well-known technique, and probably much more resilient anyway.
  • .rim file replacement. While it is allowed for .whl files to replace existing .rim files, as long as a) the .rim file hasn’t been deleted or yanked, b) the checksums match, we do not allow replacing .whl files with .rim files, nor do we allow a .rim file to overwrite an existing .rim file. This latter could be a technique to change the hosting URL for an externally hosted .whl; however, we do not think this is a good idea. There are other ways to “fix” an external host URL as described above, and we do not want to encourage mass re-uploads of existing .rim files.

Footnotes

Source: https://github.com/python/peps/blob/main/peps/pep-0759.rst

Last modified: 2024-10-09 20:11:15 GMT