Non-unique column names in fixed format HDF5

[filmor]

This leads to an error on read, writing is possible however. Since the format has a fundamental problem with this case if there are columns with the same name but different data types this is not fixable.

However, df.to_hdf should just not write the respective dataframe if the column index is not unique.

[jreback]

can u put up an example for the tests

[filmor]

df = pd.DataFrame(columns=["a"])
df2 = pd.concat([df, df], axis=1)
df2.to_hdf("test.h5", "test")
pd.read_hdf("test.h5", "test")

Only the last call fails with an InvalidIndexError: Reindexing only valid with uniquely valued Index objects in Index.get_indexer. You could silence the error by using get_indexer_non_unique instead but as described I think the reading will be broken if you have different datatypes in the columns.

It can be fixed by adding a check for the uniqueness of the column index in the fixed format. Note that this problem does not occur for the table format.

[jreback]

perfect. want to do a PR to validate and raise ValueError (could also be NotImplementedError) on writing?

[filmor]

I'll build one with tests, yes. Is it enough to fix this in BlockManagerFixed.write?

[jreback]

@filmor I believe so. Put in a test with table format as well (if you can't find a test with that is similar; if you do then put the tests together)