This page provides an overview of the major themes in pandas’ development. Each of these items requires a relatively large amount of effort to implement. These may be achieved more quickly with dedicated funding or interest from contributors.
An item being on the roadmap does not mean that it will necessarily happen, even with unlimited funding. During the implementation period we may discover issues preventing the adoption of the feature.
Additionally, an item not being on the roadmap does not exclude it from inclusion in pandas. The roadmap is intended for larger, fundamental changes to the project that are likely to take months or years of developer time. Smaller-scoped items will continue to be tracked on our issue tracker.
See Roadmap evolution for proposing changes to this document.
pandas Extension types allow for extending NumPy types with custom data types and array storage. pandas uses extension types internally, and provides an interface for 3rd-party libraries to define their own custom data types.
Many parts of pandas still unintentionally convert data to a NumPy array. These problems are especially pronounced for nested data.
We’d like to improve the handling of extension arrays throughout the library, making their behavior more consistent with the handling of NumPy arrays. We’ll do this by cleaning up pandas’ internals and adding new methods to the extension array interface.
String data type#
Currently, pandas stores text data in an
object -dtype NumPy array.
The current implementation has two primary drawbacks: First,
is not specific to strings: any Python object can be stored in an
array, not just strings. Second: this is not efficient. The NumPy memory model
isn’t especially well-suited to variable width text data.
To solve the first issue, we propose a new extension type for string data. This
will initially be opt-in, with users explicitly requesting
The array backing this string dtype may initially be the current implementation:
object -dtype NumPy array of Python strings.
To solve the second issue (performance), we’ll explore alternative in-memory
array libraries (for example, Apache Arrow). As part of the work, we may
need to implement certain operations expected by pandas users (for example
the algorithm used in,
Series.str.upper). That work may be done outside of
Consistent missing value handling#
Currently, pandas handles missing data differently for different data types. We
use different types to indicate that a value is missing (
None for object-dtype data – typically
strings or booleans – with missing values, and
pd.NaT for datetimelike
data). Integer data cannot store missing data or are cast to float. In addition,
pandas 1.0 introduced a new missing value sentinel,
pd.NA, which is being
used for the experimental nullable integer, boolean, and string data types.
These different missing values have different behaviors in user-facing operations. Specifically, we introduced different semantics for the nullable data types for certain operations (e.g. propagating in comparison operations instead of comparing as False).
Long term, we want to introduce consistent missing data handling for all data types. This includes consistent behavior in all operations (indexing, arithmetic operations, comparisons, etc.). There has been discussion of eventually making the new semantics the default.
Apache Arrow interoperability#
Apache Arrow is a cross-language development platform for in-memory data. The Arrow logical types are closely aligned with typical pandas use cases.
We’d like to provide better-integrated support for Arrow memory and data types within pandas. This will let us take advantage of its I/O capabilities and provide for better interoperability with other languages and libraries using Arrow.
Block manager rewrite#
We’d like to replace pandas current internal data structures (a collection of 1 or 2-D arrays) with a simpler collection of 1-D arrays.
pandas internal data model is quite complex. A DataFrame is made up of one or more 2-dimensional “blocks”, with one or more blocks per dtype. This collection of 2-D arrays is managed by the BlockManager.
The primary benefit of the BlockManager is improved performance on certain operations (construction from a 2D array, binary operations, reductions across the columns), especially for wide DataFrames. However, the BlockManager substantially increases the complexity and maintenance burden of pandas.
By replacing the BlockManager we hope to achieve
Substantially simpler code
Easier extensibility with new logical types
Better user control over memory use and layout
Option to provide a C / Cython API to pandas’ internals
See these design documents for more.
Decoupling of indexing and internals#
The code for getting and setting values in pandas’ data structures needs refactoring.
In particular, we must clearly separate code that converts keys (e.g., the argument
DataFrame.loc) to positions from code that uses these positions to get
or set values. This is related to the proposed BlockManager rewrite. Currently, the
BlockManager sometimes uses label-based, rather than position-based, indexing.
We propose that it should only work with positional indexing, and the translation of keys
to positions should be entirely done at a higher level.
Indexing is a complicated API with many subtleties. This refactor will require care and attention. The following principles should inspire refactoring of indexing code and should result on cleaner, simpler, and more performant code.
1. Label indexing must never involve looking in an axis twice for the same label(s). This implies that any validation step must either:
limit validation to general features (e.g. dtype/structure of the key/index), or
reuse the result for the actual indexing.
2. Indexers must never rely on an explicit call to other indexers.
For instance, it is OK to have some internal method of
.loc call some
internal method of
__getitem__ (or of their common base class),
but never in the code flow of
3. Execution of positional indexing must never involve labels (as currently, sadly, happens).
That is, the code flow of a getter call (or a setter call in which the right hand side is non-indexed)
.iloc should never involve the axes of the object in any way.
4. Indexing must never involve accessing/modifying values (i.e., act on
.values) more than once.
The following steps must hence be clearly decoupled:
find positions we need to access/modify on each axis
(if we are accessing) derive the type of object we need to return (dimensionality)
actually access/modify the values
(if we are accessing) construct the return object
5. As a corollary to the decoupling between 4.i and 4.iii, any code which deals on how data is stored (including any combination of handling multiple dtypes, and sparse storage, categoricals, third-party types) must be independent from code that deals with identifying affected rows/columns, and take place only once step 4.i is completed.
In particular, such code should most probably not live in
… and must not depend in any way on the type(s) of axes (e.g. no
6. As a corollary to point 1.i, ``Index`` (sub)classes must provide separate methods for any desired validity check of label(s) which does not involve actual lookup, on the one side, and for any required conversion/adaptation/lookup of label(s), on the other.
7. Use of trial and error should be limited, and anyway restricted to catch only exceptions
which are actually expected (typically
In particular, code should never (intentionally) raise new exceptions in the
exceptportion of a
8. Any code portion which is not specific to setters and getters must be shared,
and when small differences in behavior are expected (e.g. getting with
.loc raises for
missing labels, setting still doesn’t), they can be managed with a specific parameter.
Numba is a JIT compiler for Python code. We’d like to provide
ways for users to apply their own Numba-jitted functions where pandas accepts user-defined functions
and in groupby and window contexts). This will improve the performance of
user-defined-functions in these operations by staying within compiled code.
pandas uses airspeed velocity to monitor for performance regressions. ASV itself is a fabulous tool, but requires some additional work to be integrated into an open source project’s workflow.
The asv-runner organization, currently made up of pandas maintainers, provides tools built on top of ASV. We have a physical machine for running a number of project’s benchmarks, and tools managing the benchmark runs and reporting on results.
We’d like to fund improvements and maintenance of these tools to
Be more stable. Currently, they’re maintained on the nights and weekends when a maintainer has free time.
Tune the system for benchmarks to improve stability, following https://pyperf.readthedocs.io/en/latest/system.html
Build a GitHub bot to request ASV runs before a PR is merged. Currently, the benchmarks are only run nightly.
pandas continues to evolve. The direction is primarily determined by community interest. Everyone is welcome to review existing items on the roadmap and to propose a new item.
Each item on the roadmap should be a short summary of a larger design proposal. The proposal should include
Short summary of the changes, which would be appropriate for inclusion in the roadmap if accepted.
Motivation for the changes.
An explanation of why the change is in scope for pandas.
Detailed design: Preferably with example-usage (even if not implemented yet) and API documentation
API Change: Any API changes that may result from the proposal.
That proposal may then be submitted as a GitHub issue, where the pandas maintainers can review and comment on the design. The pandas mailing list should be notified of the proposal.
When there’s agreement that an implementation would be welcome, the roadmap should be updated to include the summary and a link to the discussion issue.
This section records now completed items from the pandas roadmap.
We improved the pandas documentation