Caveats and Gotchas¶
Using If/Truth Statements with Pandas¶
Pandas follows the numpy convention of raising an error when you try to convert something to a bool. This happens in a if or when using the boolean operations, and, or, or not. It is not clear what the result of
>>> if Series([False, True, False]):
...
should be. Should it be True because it’s not zero-length? False because there are False values? It is unclear, so instead, pandas raises a ValueError:
>>> if pd.Series([False, True, False]):
print("I was true")
Traceback
...
ValueError: The truth value of an array is ambiguous. Use a.empty, a.any() or a.all().
If you see that, you need to explicitly choose what you want to do with it (e.g., use any(), all() or empty). or, you might want to compare if the pandas object is None
>>> if pd.Series([False, True, False]) is not None:
print("I was not None")
>>> I was not None
or return if any value is True.
>>> if pd.Series([False, True, False]).any():
print("I am any")
>>> I am any
To evaluate single-element pandas objects in a boolean context, use the method .bool():
In [1]: Series([True]).bool()
True
In [2]: Series([False]).bool()
False
In [3]: DataFrame([[True]]).bool()
True
In [4]: DataFrame([[False]]).bool()
False
Bitwise boolean¶
Bitwise boolean operators like == and != will return a boolean Series, which is almost always what you want anyways.
>>> s = pd.Series(range(5))
>>> s == 4
0 False
1 False
2 False
3 False
4 True
dtype: bool
See boolean comparisons for more examples.
NaN, Integer NA values and NA type promotions¶
Choice of NA representation¶
For lack of NA (missing) support from the ground up in NumPy and Python in general, we were given the difficult choice between either
- A masked array solution: an array of data and an array of boolean values indicating whether a value
- Using a special sentinel value, bit pattern, or set of sentinel values to denote NA across the dtypes
For many reasons we chose the latter. After years of production use it has proven, at least in my opinion, to be the best decision given the state of affairs in NumPy and Python in general. The special value NaN (Not-A-Number) is used everywhere as the NA value, and there are API functions isnull and notnull which can be used across the dtypes to detect NA values.
However, it comes with it a couple of trade-offs which I most certainly have not ignored.
Support for integer NA¶
In the absence of high performance NA support being built into NumPy from the ground up, the primary casualty is the ability to represent NAs in integer arrays. For example:
In [5]: s = Series([1, 2, 3, 4, 5], index=list('abcde'))
In [6]: s
a 1
b 2
c 3
d 4
e 5
dtype: int64
In [7]: s.dtype
dtype('int64')
In [8]: s2 = s.reindex(['a', 'b', 'c', 'f', 'u'])
In [9]: s2
a 1
b 2
c 3
f NaN
u NaN
dtype: float64
In [10]: s2.dtype
dtype('float64')
This trade-off is made largely for memory and performance reasons, and also so that the resulting Series continues to be “numeric”. One possibility is to use dtype=object arrays instead.
NA type promotions¶
When introducing NAs into an existing Series or DataFrame via reindex or some other means, boolean and integer types will be promoted to a different dtype in order to store the NAs. These are summarized by this table:
| Typeclass | Promotion dtype for storing NAs |
|---|---|
| floating | no change |
| object | no change |
| integer | cast to float64 |
| boolean | cast to object |
While this may seem like a heavy trade-off, in practice I have found very few cases where this is an issue in practice. Some explanation for the motivation here in the next section.
Why not make NumPy like R?¶
Many people have suggested that NumPy should simply emulate the NA support present in the more domain-specific statistical programming langauge R. Part of the reason is the NumPy type hierarchy:
| Typeclass | Dtypes |
|---|---|
| numpy.floating | float16, float32, float64, float128 |
| numpy.integer | int8, int16, int32, int64 |
| numpy.unsignedinteger | uint8, uint16, uint32, uint64 |
| numpy.object_ | object_ |
| numpy.bool_ | bool_ |
| numpy.character | string_, unicode_ |
The R language, by contrast, only has a handful of built-in data types: integer, numeric (floating-point), character, and boolean. NA types are implemented by reserving special bit patterns for each type to be used as the missing value. While doing this with the full NumPy type hierarchy would be possible, it would be a more substantial trade-off (especially for the 8- and 16-bit data types) and implementation undertaking.
An alternate approach is that of using masked arrays. A masked array is an array of data with an associated boolean mask denoting whether each value should be considered NA or not. I am personally not in love with this approach as I feel that overall it places a fairly heavy burden on the user and the library implementer. Additionally, it exacts a fairly high performance cost when working with numerical data compared with the simple approach of using NaN. Thus, I have chosen the Pythonic “practicality beats purity” approach and traded integer NA capability for a much simpler approach of using a special value in float and object arrays to denote NA, and promoting integer arrays to floating when NAs must be introduced.
Integer indexing¶
Label-based indexing with integer axis labels is a thorny topic. It has been discussed heavily on mailing lists and among various members of the scientific Python community. In pandas, our general viewpoint is that labels matter more than integer locations. Therefore, with an integer axis index only label-based indexing is possible with the standard tools like .ix. The following code will generate exceptions:
s = Series(range(5))
s[-1]
df = DataFrame(np.random.randn(5, 4))
df
df.ix[-2:]
This deliberate decision was made to prevent ambiguities and subtle bugs (many users reported finding bugs when the API change was made to stop “falling back” on position-based indexing).
Label-based slicing conventions¶
Non-monotonic indexes require exact matches¶
Endpoints are inclusive¶
Compared with standard Python sequence slicing in which the slice endpoint is not inclusive, label-based slicing in pandas is inclusive. The primary reason for this is that it is often not possible to easily determine the “successor” or next element after a particular label in an index. For example, consider the following Series:
In [11]: s = Series(randn(6), index=list('abcdef'))
In [12]: s
a 1.337122
b -1.531095
c 1.331458
d -0.571329
e -0.026671
f -1.085663
dtype: float64
Suppose we wished to slice from c to e, using integers this would be
In [13]: s[2:5]
c 1.331458
d -0.571329
e -0.026671
dtype: float64
However, if you only had c and e, determining the next element in the index can be somewhat complicated. For example, the following does not work:
s.ix['c':'e'+1]
A very common use case is to limit a time series to start and end at two specific dates. To enable this, we made the design design to make label-based slicing include both endpoints:
In [14]: s.ix['c':'e']
c 1.331458
d -0.571329
e -0.026671
dtype: float64
This is most definitely a “practicality beats purity” sort of thing, but it is something to watch out for if you expect label-based slicing to behave exactly in the way that standard Python integer slicing works.
Miscellaneous indexing gotchas¶
Reindex versus ix gotchas¶
Many users will find themselves using the ix indexing capabilities as a concise means of selecting data from a pandas object:
In [15]: df = DataFrame(randn(6, 4), columns=['one', 'two', 'three', 'four'],
....: index=list('abcdef'))
....:
In [16]: df
one two three four
a -1.114738 -0.058216 -0.486768 1.685148
b 0.112572 -1.495309 0.898435 -0.148217
c -1.596070 0.159653 0.262136 0.036220
d 0.184735 -0.255069 -0.271020 1.288393
e 0.294633 -1.165787 0.846974 -0.685597
f 0.609099 -0.303961 0.625555 -0.059268
[6 rows x 4 columns]
In [17]: df.ix[['b', 'c', 'e']]
one two three four
b 0.112572 -1.495309 0.898435 -0.148217
c -1.596070 0.159653 0.262136 0.036220
e 0.294633 -1.165787 0.846974 -0.685597
[3 rows x 4 columns]
This is, of course, completely equivalent in this case to using th reindex method:
In [18]: df.reindex(['b', 'c', 'e'])
one two three four
b 0.112572 -1.495309 0.898435 -0.148217
c -1.596070 0.159653 0.262136 0.036220
e 0.294633 -1.165787 0.846974 -0.685597
[3 rows x 4 columns]
Some might conclude that ix and reindex are 100% equivalent based on this. This is indeed true except in the case of integer indexing. For example, the above operation could alternately have been expressed as:
In [19]: df.ix[[1, 2, 4]]
one two three four
b 0.112572 -1.495309 0.898435 -0.148217
c -1.596070 0.159653 0.262136 0.036220
e 0.294633 -1.165787 0.846974 -0.685597
[3 rows x 4 columns]
If you pass [1, 2, 4] to reindex you will get another thing entirely:
In [20]: df.reindex([1, 2, 4])
one two three four
1 NaN NaN NaN NaN
2 NaN NaN NaN NaN
4 NaN NaN NaN NaN
[3 rows x 4 columns]
So it’s important to remember that reindex is strict label indexing only. This can lead to some potentially surprising results in pathological cases where an index contains, say, both integers and strings:
In [21]: s = Series([1, 2, 3], index=['a', 0, 1])
In [22]: s
a 1
0 2
1 3
dtype: int64
In [23]: s.ix[[0, 1]]
0 2
1 3
dtype: int64
In [24]: s.reindex([0, 1])
0 2
1 3
dtype: int64
Because the index in this case does not contain solely integers, ix falls back on integer indexing. By contrast, reindex only looks for the values passed in the index, thus finding the integers 0 and 1. While it would be possible to insert some logic to check whether a passed sequence is all contained in the index, that logic would exact a very high cost in large data sets.
Reindex potentially changes underlying Series dtype¶
The use of reindex_like can potentially change the dtype of a Series.
series = pandas.Series([1, 2, 3])
x = pandas.Series([True])
x.dtype
x = pandas.Series([True]).reindex_like(series)
x.dtype
This is because reindex_like silently inserts NaNs and the dtype changes accordingly. This can cause some issues when using numpy ufuncs such as numpy.logical_and.
See the this old issue for a more detailed discussion.
Timestamp limitations¶
Minimum and maximum timestamps¶
Since pandas represents timestamps in nanosecond resolution, the timespan that can be represented using a 64-bit integer is limited to approximately 584 years:
In [25]: begin = Timestamp.min
In [26]: begin
Timestamp('1677-09-22 00:12:43.145225', tz=None)
In [27]: end = Timestamp.max
In [28]: end
Timestamp('2262-04-11 23:47:16.854775807', tz=None)
If you need to represent time series data outside the nanosecond timespan, use PeriodIndex:
In [29]: span = period_range('1215-01-01', '1381-01-01', freq='D')
In [30]: span
<class 'pandas.tseries.period.PeriodIndex'>
freq: D
[1215-01-01, ..., 1381-01-01]
length: 60632
Parsing Dates from Text Files¶
When parsing multiple text file columns into a single date column, the new date column is prepended to the data and then index_col specification is indexed off of the new set of columns rather than the original ones:
In [31]: print(open('tmp.csv').read())
KORD,19990127, 19:00:00, 18:56:00, 0.8100
KORD,19990127, 20:00:00, 19:56:00, 0.0100
KORD,19990127, 21:00:00, 20:56:00, -0.5900
KORD,19990127, 21:00:00, 21:18:00, -0.9900
KORD,19990127, 22:00:00, 21:56:00, -0.5900
KORD,19990127, 23:00:00, 22:56:00, -0.5900
In [32]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}
In [33]: df = read_csv('tmp.csv', header=None,
....: parse_dates=date_spec,
....: keep_date_col=True,
....: index_col=0)
....:
# index_col=0 refers to the combined column "nominal" and not the original
# first column of 'KORD' strings
In [34]: df
actual 0 1 2 3 \
nominal
1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 19990127 19:00:00 18:56:00
1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 19990127 20:00:00 19:56:00
1999-01-27 21:00:00 1999-01-27 20:56:00 KORD 19990127 21:00:00 20:56:00
1999-01-27 21:00:00 1999-01-27 21:18:00 KORD 19990127 21:00:00 21:18:00
1999-01-27 22:00:00 1999-01-27 21:56:00 KORD 19990127 22:00:00 21:56:00
1999-01-27 23:00:00 1999-01-27 22:56:00 KORD 19990127 23:00:00 22:56:00
4
nominal
1999-01-27 19:00:00 0.81
1999-01-27 20:00:00 0.01
1999-01-27 21:00:00 -0.59
1999-01-27 21:00:00 -0.99
1999-01-27 22:00:00 -0.59
1999-01-27 23:00:00 -0.59
[6 rows x 6 columns]
Differences with NumPy¶
For Series and DataFrame objects, var normalizes by N-1 to produce unbiased estimates of the sample variance, while NumPy’s var normalizes by N, which measures the variance of the sample. Note that cov normalizes by N-1 in both pandas and NumPy.
Thread-safety¶
As of pandas 0.11, pandas is not 100% thread safe. The known issues relate to the DataFrame.copy method. If you are doing a lot of copying of DataFrame objects shared among threads, we recommend holding locks inside the threads where the data copying occurs.
See this link for more information.
HTML Table Parsing¶
There are some versioning issues surrounding the libraries that are used to parse HTML tables in the top-level pandas io function read_html.
Issues with lxml
- Benefits
- Drawbacks
- lxml does not make any guarantees about the results of it’s parse unless it is given strictly valid markup.
- In light of the above, we have chosen to allow you, the user, to use the lxml backend, but this backend will use html5lib if lxml fails to parse
- It is therefore highly recommended that you install both BeautifulSoup4 and html5lib, so that you will still get a valid result (provided everything else is valid) even if lxml fails.
Issues with BeautifulSoup4 using lxml as a backend
- The above issues hold here as well since BeautifulSoup4 is essentially just a wrapper around a parser backend.
Issues with BeautifulSoup4 using html5lib as a backend
- Benefits
- html5lib is far more lenient than lxml and consequently deals with real-life markup in a much saner way rather than just, e.g., dropping an element without notifying you.
- html5lib generates valid HTML5 markup from invalid markup automatically. This is extremely important for parsing HTML tables, since it guarantees a valid document. However, that does NOT mean that it is “correct”, since the process of fixing markup does not have a single definition.
- html5lib is pure Python and requires no additional build steps beyond its own installation.
- Drawbacks
- The biggest drawback to using html5lib is that it is slow as molasses. However consider the fact that many tables on the web are not big enough for the parsing algorithm runtime to matter. It is more likely that the bottleneck will be in the process of reading the raw text from the url over the web, i.e., IO (input-output). For very large tables, this might not be true.
Issues with using Anaconda
- Anaconda ships with lxml version 3.2.0; the following workaround for Anaconda was successfully used to deal with the versioning issues surrounding lxml and BeautifulSoup4.
Note
Unless you have both:
- A strong restriction on the upper bound of the runtime of some code that incorporates read_html()
- Complete knowledge that the HTML you will be parsing will be 100% valid at all times
then you should install html5lib and things will work swimmingly without you having to muck around with conda. If you want the best of both worlds then install both html5lib and lxml. If you do install lxml then you need to perform the following commands to ensure that lxml will work correctly:
# remove the included version conda remove lxml # install the latest version of lxml pip install 'git+git://github.com/lxml/lxml.git' # install the latest version of beautifulsoup4 pip install 'bzr+lp:beautifulsoup'Note that you need bzr and git installed to perform the last two operations.
Byte-Ordering Issues¶
Occasionally you may have to deal with data that were created on a machine with a different byte order than the one on which you are running Python. To deal with this issue you should convert the underlying NumPy array to the native system byte order before passing it to Series/DataFrame/Panel constructors using something similar to the following:
In [35]: x = np.array(list(range(10)), '>i4') # big endian
In [36]: newx = x.byteswap().newbyteorder() # force native byteorder
In [37]: s = Series(newx)
See the NumPy documentation on byte order for more details.