问题
I read that NaNs are problematic, but the following causes an actual corruption of my data, rather than an error. Is this a bug? Have I missed something basic in the documentation? I would like the second command to give an error or to give the same response as the first command:
ipdb> df
year PRuid QC data
18 2007 nonQC 0 8.014261
19 2008 nonQC 0 7.859152
20 2010 nonQC 0 7.468260
21 1985 10 NaN 0.861403
22 1985 11 NaN 0.878531
23 1985 12 NaN 0.842704
24 1985 13 NaN 0.785877
25 1985 24 1 0.730625
26 1985 35 NaN 0.816686
27 1985 46 NaN 0.819271
28 1985 47 NaN 0.807050
ipdb> df.set_index(['year','PRuid','QC']).reset_index()
year PRuid QC data
0 2007 nonQC 0 8.014261
1 2008 nonQC 0 7.859152
2 2010 nonQC 0 7.468260
3 1985 10 1 0.861403
4 1985 11 1 0.878531
5 1985 12 1 0.842704
6 1985 13 1 0.785877
7 1985 24 1 0.730625
8 1985 35 1 0.816686
9 1985 46 1 0.819271
10 1985 47 1 0.807050
The value of "QC" is actually changed to 1 from NaN where it should be NaN.
Btw, for symmetry I added the ".reset_index()", but the data corruption is introduced by set_index.
And in case this is interesting, the version is:
pd.version
<module 'pandas.version' from '/usr/lib/python2.6/site-packages/pandas-0.10.1-py2.6-linux-x86_64.egg/pandas/version.pyc'>
回答1:
So this was a bug. By the end of May 2013, pandas 0.11.1 should be released with the bug fix (see comments on this question). In the mean time, I avoided using a value with NaNs in any multiindex, for instance by using some other flag value (-99) for the NaNs in the column 'QC'.
来源：https://stackoverflow.com/questions/16511319/trouble-with-nans-set-index-reset-index-corrupts-data