PDEP-1: Purpose and guidelines#

PDEP definition, purpose and scope#

A PDEP (pandas enhancement proposal) is a proposal for a major change in pandas, in a similar way as a Python PEP or a NumPy NEP.

Bug fixes and conceptually minor changes (e.g. adding a parameter to a function) are out of the scope of PDEPs. A PDEP should be used for changes that are not immediate and not obvious, when everybody in the pandas community needs to be aware of the possibility of an upcoming change. Such changes require detailed documentation before being implemented and frequently lead to a significant discussion within the community.

PDEPs are appropriate for user facing changes, internal changes and significant discussions. Examples of topics worth a PDEP could include substantial API changes, breaking behavior changes, moving a module from pandas to a separate repository, or a refactoring of the pandas block manager. It is not always trivial to know which issue has enough scope to require the full PDEP process. Some simple API changes have sufficient consensus among the core team, and minimal impact on the community. On the other hand, if an issue becomes controversial, i.e. it generated a significant discussion, one could suggest opening a PDEP to formalize and document the discussion, making it easier for the wider community to participate. For context, see the list of issues that could have been a PDEP.

PDEP guidelines#

Target audience#

A PDEP is a public document available to anyone, but the main stakeholders to consider when writing a PDEP are:

  • The core development team, who will have the final decision on whether a PDEP is approved or not

  • Contributors to pandas and other related projects, and experienced users. Their feedback is highly encouraged and appreciated, to make sure all points of views are taken into consideration

  • The wider pandas community, in particular users, who may or may not have feedback on the proposal, but should know and be able to understand the future direction of the project

PDEP authors#

Anyone can propose a PDEP, but a core member should be engaged to advise on a proposal made by non-core contributors. To submit a PDEP as a community member, please propose the PDEP concept on an issue, and find a pandas team member to collaborate with. They can advise you on the PDEP process and should be listed as an advisor on the PDEP when it is submitted to the PDEP repository.

Workflow#

The possible states of a PDEP are:

  • Under discussion

  • Accepted

  • Implemented

  • Rejected

Next is described the workflow that PDEPs can follow.

Submitting a PDEP#

Proposing a PDEP is done by creating a PR adding a new file to web/pdeps/. The file is a markdown file, you can use web/pdeps/0001.md as a reference for the expected format.

The initial status of a PDEP will be Status: Under discussion. This will be changed to Status: Accepted when the PDEP is ready and has the approval of the core team.

Accepted PDEP#

A PDEP can only be accepted by the core development team, if the proposal is considered worth implementing. Decisions will be made based on the process detailed in the pandas governance document. In general, more than one approval will be needed before the PR is merged. And there should not be any Request changes review at the time of merging.

Once a PDEP is accepted, any contributions can be made toward the implementation of the PDEP, with an open-ended completion timeline. Development of pandas is difficult to understand and forecast, being that the contributors to pandas are a mix of volunteers and developers paid from different sources, with different priorities. For companies, institutions or individuals with interest in seeing a PDEP being implemented, or to in general see progress to the pandas roadmap, please check how you can help in the contributing page.

Implemented PDEP#

Once a PDEP is implemented and available in the main branch of pandas, its status will be changed to Status: Implemented, so there is visibility that the PDEP is not part of the roadmap and future plans, but a change that has already happened. The first pandas version in which the PDEP implementation is available will also be included in the PDEP header with for example Implemented: v2.0.0.

Rejected PDEP#

A PDEP can be rejected when the final decision is that its implementation is not in the best interests of the project. Rejected PDEPs are as useful as accepted PDEPs, since there are discussions that are worth having, and decisions about changes to pandas being made. They will be merged with Status: Rejected, so there is visibility on what was discussed and what was the outcome of the discussion. A PDEP can be rejected for different reasons, for example good ideas that are not backward-compatible, and the breaking changes are not considered worth implementing.

Invalid PDEP#

For submitted PDEPs that do not contain proper documentation, are out of scope, or are not useful to the community for any other reason, the PR will be closed after discussion with the author, instead of merging them as rejected. This is to avoid adding noise to the list of rejected PDEPs, which should contain documentation as good as an accepted PDEP, but where the final decision was to not implement the changes.

Evolution of PDEPs#

Most PDEPs are not expected to change after they are accepted. Once there is agreement on the changes, and they are implemented, the PDEP will be only useful to understand why the development happened, and the details of the discussion.

But in some cases, a PDEP can be updated. For example, a PDEP defining a procedure or a policy, like this one (PDEP-1). Or cases when after attempting the implementation, new knowledge is obtained that makes the original PDEP obsolete, and changes are required. When there are specific changes to be made to the original PDEP, this will be edited, its Revision: X label will be increased by one, and a note will be added to the PDEP-N history section. This will let readers understand that the PDEP has changed and avoid confusion.

List of issues that could have been PDEPs for context#

Clear examples for potential PDEPs:#

  • Adding a new parameter to many existing methods, or deprecating one in many places. For example:

    • The numeric_only deprecation (GH-28900) affected many methods and could have been a PDEP.

  • Adding a new data type has impact on a variety of places that need to handle the data type. Such wide-ranging impact would require a PDEP. For example:

  • A significant (breaking) change in existing behavior. For example:

  • Support of new Python features with a wide impact on the project. For example:

    • Supporting typing within pandas vs. creation of pandas-stubs (GH-43197, GH-45253)

  • New required dependency.

  • Removing module from the project or splitting it off to a separate repository:

    • Moving rarely used I/O connectors to a separate repository GH-28409

  • Significant changes to contributors’ processes are not going to have an impact on users, but they do benefit from structured discussion among the contributors. For example:

    • Changing the build system to meson (GH-49115)

Borderline examples:#

Small changes to core functionality, such as DataFrame and Series, should always be considered as a PDEP candidate as it will likely have a big impact on users. But the same types of changes in other functionalities would not be good PDEP candidates. That said, any discussion, no matter how small the change, which becomes controversial is a PDEP candidate. Consider if more attention and/or a formal decision-making process would help. Following are some examples we hope can help clarify our meaning here:

  • API breaking changes, or discussion thereof, could be a PDEP. For example:

    • value_counts result rename (GH-49497). The scope does not justify a PDEP at first, but later a discussion about whether it should be executed as a breaking change or with deprecation emerges, which could benefit from the PDEP process.

  • Adding new methods or parameters to an existing method typically will not require a PDEP for non-core features. For example:

    • Both dropna(percentage) (GH-35299), and Timestamp.normalize() (GH-8794) would not have required a PDEP.

    • On the other hand, DataFrame.assign() might. While it is a single method without backwards compatibility concerns, it is also a core feature and the discussion should be highly visible.

  • Deprecating or removing a single method would not require a PDEP in most cases.

    • That said, DataFrame.append (GH-35407) is an example of deprecations on core features that would be a good candidate for a PDEP.

  • Changing the default value of parameters in a core pandas method is another edge case. For example:

    • Such changes for dropna in DataFrame.groupby and Series.groupby could be a PDEP.

  • New top level modules and/or exposing internal classes. For example:

    • Add pandas.api.typing (GH-48577) is relatively small and would not necessarily require a PDEP.

PDEP-1 History#

  • 3 August 2022: Initial version (GH-47938)

  • 15 February 2023: Version 2 (GH-51417) clarifies the scope of PDEPs and adds examples