This section will provide a look into some of pandas internals. It’s primarily intended for developers of pandas itself.


In pandas there are a few objects implemented which can serve as valid containers for the axis labels:

  • Index: the generic “ordered set” object, an ndarray of object dtype assuming nothing about its contents. The labels must be hashable (and likely immutable) and unique. Populates a dict of label to location in Cython to do O(1) lookups.

  • MultiIndex: the standard hierarchical index object

  • DatetimeIndex: An Index object with Timestamp boxed elements (impl are the int64 values)

  • TimedeltaIndex: An Index object with Timedelta boxed elements (impl are the in64 values)

  • PeriodIndex: An Index object with Period elements

There are functions that make the creation of a regular index easy:

  • date_range(): fixed frequency date range generated from a time rule or DateOffset. An ndarray of Python datetime objects

  • period_range(): fixed frequency date range generated from a time rule or DateOffset. An ndarray of Period objects, representing timespans

The motivation for having an Index class in the first place was to enable different implementations of indexing. This means that it’s possible for you, the user, to implement a custom Index subclass that may be better suited to a particular application than the ones provided in pandas.

From an internal implementation point of view, the relevant methods that an Index must define are one or more of the following (depending on how incompatible the new object internals are with the Index functions):

  • get_loc(): returns an “indexer” (an integer, or in some cases a slice object) for a label

  • slice_locs(): returns the “range” to slice between two labels

  • get_indexer(): Computes the indexing vector for reindexing / data alignment purposes. See the source / docstrings for more on this

  • get_indexer_non_unique(): Computes the indexing vector for reindexing / data alignment purposes when the index is non-unique. See the source / docstrings for more on this

  • reindex(): Does any pre-conversion of the input index then calls get_indexer

  • union(), intersection(): computes the union or intersection of two Index objects

  • insert(): Inserts a new label into an Index, yielding a new object

  • delete(): Delete a label, yielding a new object

  • drop(): Deletes a set of labels

  • take(): Analogous to ndarray.take


Internally, the MultiIndex consists of a few things: the levels, the integer codes, and the level names:

In [1]: index = pd.MultiIndex.from_product(
   ...:     [range(3), ["one", "two"]], names=["first", "second"]
   ...: )

In [2]: index
MultiIndex([(0, 'one'),
            (0, 'two'),
            (1, 'one'),
            (1, 'two'),
            (2, 'one'),
            (2, 'two')],
           names=['first', 'second'])

In [3]: index.levels
Out[3]: FrozenList([[0, 1, 2], ['one', 'two']])

In [4]:
Out[4]: FrozenList([[0, 0, 1, 1, 2, 2], [0, 1, 0, 1, 0, 1]])

In [5]: index.names
Out[5]: FrozenList(['first', 'second'])

You can probably guess that the codes determine which unique element is identified with that location at each layer of the index. It’s important to note that sortedness is determined solely from the integer codes and does not check (or care) whether the levels themselves are sorted. Fortunately, the constructors from_tuples() and from_arrays() ensure that this is true, but if you compute the levels and codes yourself, please be careful.


pandas extends NumPy’s type system with custom types, like Categorical or datetimes with a timezone, so we have multiple notions of “values”. For 1-D containers (Index classes and Series) we have the following convention:

  • cls._values refers is the “best possible” array. This could be an ndarray or ExtensionArray.

So, for example, Series[category]._values is a Categorical.

Subclassing pandas data structures#

This section has been moved to Subclassing pandas data structures.