什么是上传一个很大的CSV文件笔记本蟒蛇大熊猫工作的最快方法？

Question

我正在尝试上传一个250MB的csv文件.基本上有400万行和6列时间序列数据(1分钟).通常的程序是:

location = r'C:\Users\Name\Folder_1\Folder_2\file.csv'
df = pd.read_csv(location)

这个过程大约需要20分钟!!! 非常初步我已经探索了以下选项

• 以块的形式上传,然后将块放在一起.
• HDF5
• '羽毛'
• '泡菜'

我想知道是否有人比较了这些选项(或更多)并且有明显的赢家.如果没有人回答,将来我会发布我的结果.我现在没有时间.

Answer 1

以下是我对DF的读写比较结果(形状:4000000 x 6,内存大小183.1 MB,未压缩CSV大小 - 492 MB).

以下存储格式比较:( ,CSV,,CSV.gzip [各种压缩]):PickleHDF5

                  read_s  write_s  size_ratio_to_CSV
storage
CSV               17.900    69.00              1.000
CSV.gzip          18.900   186.00              0.047
Pickle             0.173     1.77              0.374
HDF_fixed          0.196     2.03              0.435
HDF_tab            0.230     2.60              0.437
HDF_tab_zlib_c5    0.845     5.44              0.035
HDF_tab_zlib_c9    0.860     5.95              0.035
HDF_tab_bzip2_c5   2.500    36.50              0.011
HDF_tab_bzip2_c9   2.500    36.50              0.011

读

写作/储蓄

与未压缩的CSV文件相关的文件大小比率

原始数据:

CSV:

In [68]: %timeit df.to_csv(fcsv)
1 loop, best of 3: 1min 9s per loop

In [74]: %timeit pd.read_csv(fcsv)
1 loop, best of 3: 17.9 s per loop

CSV.gzip:

In [70]: %timeit df.to_csv(fcsv_gz, compression='gzip')
1 loop, best of 3: 3min 6s per loop

In [75]: %timeit pd.read_csv(fcsv_gz)
1 loop, best of 3: 18.9 s per loop

泡菜:

In [66]: %timeit df.to_pickle(fpckl)
1 loop, best of 3: 1.77 s per loop

In [72]: %timeit pd.read_pickle(fpckl)
10 loops, best of 3: 173 ms per loop

HDF(format='fixed')[默认]:

In [67]: %timeit df.to_hdf(fh5, 'df')
1 loop, best of 3: 2.03 s per loop

In [73]: %timeit pd.read_hdf(fh5, 'df')
10 loops, best of 3: 196 ms per loop

HDF(format='table'):

In [37]: %timeit df.to_hdf('D:\\temp\\.data\\37010212_tab.h5', 'df', format='t')
1 loop, best of 3: 2.6 s per loop

In [38]: %timeit pd.read_hdf('D:\\temp\\.data\\37010212_tab.h5', 'df')
1 loop, best of 3: 230 ms per loop

HDF(format='table', complib='zlib', complevel=5):

In [40]: %timeit df.to_hdf('D:\\temp\\.data\\37010212_tab_compress_zlib5.h5', 'df', format='t', complevel=5, complib='zlib')
1 loop, best of 3: 5.44 s per loop

In [41]: %timeit pd.read_hdf('D:\\temp\\.data\\37010212_tab_compress_zlib5.h5', 'df')
1 loop, best of 3: 854 ms per loop

HDF(format='table', complib='zlib', complevel=9):

In [36]: %timeit df.to_hdf('D:\\temp\\.data\\37010212_tab_compress_zlib9.h5', 'df', format='t', complevel=9, complib='zlib')
1 loop, best of 3: 5.95 s per loop

In [39]: %timeit pd.read_hdf('D:\\temp\\.data\\37010212_tab_compress_zlib9.h5', 'df')
1 loop, best of 3: 860 ms per loop

HDF(format='table', complib='bzip2', complevel=5):

In [42]: %timeit df.to_hdf('D:\\temp\\.data\\37010212_tab_compress_bzip2_l5.h5', 'df', format='t', complevel=5, complib='bzip2')
1 loop, best of 3: 36.5 s per loop

In [43]: %timeit pd.read_hdf('D:\\temp\\.data\\37010212_tab_compress_bzip2_l5.h5', 'df')
1 loop, best of 3: 2.5 s per loop

HDF(format='table', complib='bzip2', complevel=9):

In [42]: %timeit df.to_hdf('D:\\temp\\.data\\37010212_tab_compress_bzip2_l9.h5', 'df', format='t', complevel=9, complib='bzip2')
1 loop, best of 3: 36.5 s per loop

In [43]: %timeit pd.read_hdf('D:\\temp\\.data\\37010212_tab_compress_bzip2_l9.h5', 'df')
1 loop, best of 3: 2.5 s per loop

PS我无法feather在我的Windows笔记本上测试

DF信息:

In [49]: df.shape
Out[49]: (4000000, 6)

In [50]: df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 4000000 entries, 0 to 3999999
Data columns (total 6 columns):
a    datetime64[ns]
b    datetime64[ns]
c    datetime64[ns]
d    datetime64[ns]
e    datetime64[ns]
f    datetime64[ns]
dtypes: datetime64[ns](6)
memory usage: 183.1 MB

In [41]: df.head()
Out[41]:
                    a                   b                   c  \
0 1970-01-01 00:00:00 1970-01-01 00:01:00 1970-01-01 00:02:00
1 1970-01-01 00:01:00 1970-01-01 00:02:00 1970-01-01 00:03:00
2 1970-01-01 00:02:00 1970-01-01 00:03:00 1970-01-01 00:04:00
3 1970-01-01 00:03:00 1970-01-01 00:04:00 1970-01-01 00:05:00
4 1970-01-01 00:04:00 1970-01-01 00:05:00 1970-01-01 00:06:00

                    d                   e                   f
0 1970-01-01 00:03:00 1970-01-01 00:04:00 1970-01-01 00:05:00
1 1970-01-01 00:04:00 1970-01-01 00:05:00 1970-01-01 00:06:00
2 1970-01-01 00:05:00 1970-01-01 00:06:00 1970-01-01 00:07:00
3 1970-01-01 00:06:00 1970-01-01 00:07:00 1970-01-01 00:08:00
4 1970-01-01 00:07:00 1970-01-01 00:08:00 1970-01-01 00:09:00

文件大小:

{ .data }  » ls -lh 37010212.*                                                                          /d/temp/.data
-rw-r--r-- 1 Max None 492M May  3 22:21 37010212.csv
-rw-r--r-- 1 Max None  23M May  3 22:19 37010212.csv.gz
-rw-r--r-- 1 Max None 214M May  3 22:02 37010212.h5
-rw-r--r-- 1 Max None 184M May  3 22:02 37010212.pickle
-rw-r--r-- 1 Max None 215M May  4 10:39 37010212_tab.h5
-rw-r--r-- 1 Max None 5.4M May  4 10:46 37010212_tab_compress_bzip2_l5.h5
-rw-r--r-- 1 Max None 5.4M May  4 10:51 37010212_tab_compress_bzip2_l9.h5
-rw-r--r-- 1 Max None  17M May  4 10:42 37010212_tab_compress_zlib5.h5
-rw-r--r-- 1 Max None  17M May  4 10:36 37010212_tab_compress_zlib9.h5

结论:

Pickle并且HDF5更快,但HDF5更方便 - 你可以在里面存储多个表/帧,你可以有条件地读取你的数据(看看read_hdf()中的where参数),你也可以存储压缩的数据(- 更快,- 提供更好的压缩比)等zlibbzip2

PS,如果你可以建立/使用feather-format- 与HDF5和相比,它应该更快Pickle

PPS:不要将Pickle用于大数据帧,因为您最终可能会遇到SystemError:错误返回无异常设置错误消息.它也在这里和这里描述.