Groupby and apply pandas vs dask

Question

there is something that I quite don't understand about dask.dataframe behavior. Let say I want to replicate this from pandas

import pandas as pd
import dask.dataframe as dd
import random

s = "abcd"
lst = 10*[0]+list(range(1,6))
n = 100
df = pd.DataFrame({"col1": [random.choice(s) for i in range(n)],
                   "col2": [random.choice(lst) for i in range(n)]})
# I will need an hash in dask
df["hash"] = 2*df.col1
df = df[["hash","col1","col2"]]

def fun(data):
    if data["col2"].mean()>1:
        data["col3"]=2
    else:
        data["col3"]=1
    return(data)

df1 = df.groupby("col1").apply(fun)
df1.head()

this returns

  hash col1  col2  col3
0   dd    d     0     1
1   aa    a     0     2
2   bb    b     0     1
3   bb    b     0     1
4   aa    a     0     2

In Dask I tried

def fun2(data):
    if data["col2"].mean()>1:
        return 2
    else:
        return 1

ddf = df.copy()
ddf.set_index("hash",inplace=True)
ddf = dd.from_pandas(ddf, npartitions=2)

gpb = ddf.groupby("col1").apply(fun2, meta=pd.Series())

where the groupby lead to the same result as in pandas but I'm having hard time merging the result on a new column preserving the hash index. I'd like to have the following result

      col1  col2  col3
hash           
aa      a     5     2
aa      a     0     2
aa      a     0     2
aa      a     0     2
aa      a     4     2

UPDATE

Playing with merge I found this solution

ddf1 = dd.merge(ddf, gpb.to_frame(), 
                left_on="col1",
                left_index=False, right_index=True)
ddf1 = ddf1.rename(columns={0:"col3"})

I'm not sure how this is going to work if I have to a groupby over several columns. Plus is not exactly elegant.

Your solution involves merging a computed dask df against another df, which goes against two performance suggestions from the docs: http://dask.pydata.org/en/latest/dataframe-performance.html#joins — kuanb, Jul 15 '17 at 09:49
Hi @knuab, thank you. I already had a look on that doc but I couldn't find any other way to solve my problem. — rpanai, Jul 17 '17 at 13:17

score 4 · Accepted Answer · edited May 10 '19 at 18:45

How about using join?

This is your dask code with the exception of naming the Series pd.Series(name='col3')

def fun2(data):
    if data["col2"].mean()>1:
        return 2
    else:
        return 1

ddf = df.copy()
ddf.set_index("hash",inplace=True)
ddf = dd.from_pandas(ddf, npartitions=2)

gpb = ddf.groupby("col1").apply(fun2, meta=pd.Series(name='col3'))

then the join

ddf.join(gpb.to_frame(), on='col1')
print(ddf.compute().head())
      col1  col2  col3
hash                 
cc      c     0     2
cc      c     0     2
cc      c     0     2
cc      c     2     2
cc      c     0     2

Groupby and apply pandas vs dask

1 Answers1