Pandas Series value_counts working differently for different counts

Question

For example:

df1 = pd.DataFrame(np.repeat(np.arange(1,7),3), columns=['A'])

df1.A.value_counts(sort=False)
1    3
2    3
3    3
4    3
5    3
6    3
Name: A, dtype: int64

---

df2 = pd.DataFrame(np.repeat(np.arange(1,7),100), columns=['A'])

df2.A.value_counts(sort=False)
1    100
2    100
3    100
4    100
5    100
6    100
Name: A, dtype: int64

---

In the above examples the value_counts works perfectly and give the required result. whereas when coming to larger dataframes it is giving a different output. Here the A values are already sorted and counts are also same, but the order of index that is A changed after value_counts. Why is it doing correctly for small counts but not for large counts:

df3 = pd.DataFrame(np.repeat(np.arange(1,7),1000), columns=['A'])

df3.A.value_counts(sort=False)
4    1000
1    1000
5    1000
2    1000
6    1000
3    1000
Name: A, dtype: int64

Here I can do df3.A.value_counts(sort=False).sort_index() or df3.A.value_counts(sort=False).reindex(df.A.unique()). I want to know the reason why it is behaving differently for different counts?

Using:

Numpy version :1.15.2
Pandas version :0.23.4

Answer 1

This is actually a known problem.

If you browse through the source code -

1. C:\ProgramData\Anaconda3\Lib\site-packages\pandas\core\algorithims.py line 581 is the original implementation
2. It calls _value_counts_arraylike for int64 values when bins=None
3. This function makes a call - keys, counts = htable.value_count_int64(values, dropna)

If you then look at the htable implementation you will conclude that the keys are in an arbitrary order, subject to how the hashtable works.

Its not a guarantee of ANY kind of ordering. Typically this routine sorts by biggest values, and that is almost always what you want.

I guess they can change this to have sort=False mean original ordering. I don't know if this would actually break anything (and done internally this isn't very costly as the uniques are already known).

The order is changed from pandas/hashtable.pyx.build_count_table_object(). Resizing of the pymap moves the entries by hashing values.

Here is the full discussion