CODEWITHSUNDEEP
pandasdataframepyspark
David Kon
David Kon

Reputation: 65

Pyspark - remove duplicates from dataframe keeping the last appearance

I'm trying to dedupe a spark dataframe leaving only the latest appearance. The duplication is in three variables:

NAME
ID
DOB

I succeeded in Pandas with the following:

df_dedupe = df.drop_duplicates(subset=['NAME','ID','DOB'], keep='last', inplace=False)

But in spark I tried the following:

df_dedupe = df.dropDuplicates(['NAME', 'ID', 'DOB'], keep='last')

I get this error:

TypeError: dropDuplicates() got an unexpected keyword argument 'keep'

Any ideas?

Upvotes: 1

Views: 15112

Answers (2)

David Kon
David Kon

Reputation: 65

Thanks for your help. I followed your directives but the outcome was not as expected:

d1 = [('Bob', '10', '1542189668', '0', '0'),  ('Alice', '10', '1425298030', '154', '39'), ('Bob', '10', '1542189668', '178', '42')]
df1 = spark.createDataFrame(d1, ['NAME', 'ID', 'DOB' , 'Height' , 'ShoeSize'])
df_dedupe = df1.dropDuplicates(['NAME', 'ID', 'DOB'])
df_reverse = df1.sort((["NAME", "ID", "DOB"]), ascending= False)
df_dedupe.join(df_reverse,['NAME','ID','DOB'],'inner')
df_dedupe.show(100, False)

The outcome was:

+-----+---+----------+------+--------+    
|NAME |ID |DOB       |Height|ShoeSize|
+-----+---+----------+------+--------+
|Bob  |10 |1542189668|0     |0       |
|Alice|10 |1425298030|154   |39      |
+-----+---+----------+------+--------+

Showing the "Bob" with corrupted data.

Finally, I changed my approach and converted the DF to Pandas and then back to spark:

p_schema = StructType([StructField('NAME',StringType(),True),StructField('ID',StringType(),True),StructField('DOB',StringType(),True),StructField('Height',StringType(),True),StructField('ShoeSize',StringType(),True)])
d1 = [('Bob', '10', '1542189668', '0', '0'),  ('Alice', '10', '1425298030', '154', '39'), ('Bob', '10', '1542189668', '178', '42')]
df = spark.createDataFrame(d1, p_schema)
pdf = df.toPandas()
df_dedupe = pdf.drop_duplicates(subset=['NAME','ID','DOB'], keep='last', inplace=False)

df_spark = spark.createDataFrame(df_dedupe, p_schema)
df_spark.show(100, False)

This finally brought the correct "Bob":

+-----+---+----------+------+--------+
|NAME |ID |DOB       |Height|ShoeSize|
+-----+---+----------+------+--------+
|Alice|10 |1425298030|154   |39      |
|Bob  |10 |1542189668|178   |42      |
+-----+---+----------+------+--------+

Of course, I'd still like to have a purely Spark solution but the lack of indexing seems to be problematic with Spark.

Thanks!

Upvotes: 1

Manrique
Manrique

Reputation: 2221

As you can see in http://spark.apache.org/docs/2.1.0/api/python/pyspark.sql.html the documentation of the function dropDuplicates(subset=None), it only allows a subset as parameter. Why would you like to keep the last, if they're all equal?

EDIT

As @W-B pointed, u want the other columns. My solution will be to sort the original dataframe in reverse order, and use the df_dedupe on the three repeated columns to make an inner join and only preserve the last values.

df_dedupe.join(original_df,['NAME','ID','DOB'],'inner')

Upvotes: 0

Related Questions