How to pivot on multiple columns in Spark SQL?

Question

I need to pivot more than one column in a pyspark dataframe. Sample dataframe,

 >>> d = [(100,1,23,10),(100,2,45,11),(100,3,67,12),(100,4,78,13),(101,1,23,10),(101,2,45,13),(101,3,67,14),(101,4,78,15),(102,1,23,10),(102,2,45,11),(102,3,67,16),(102,4,78,18)]
>>> mydf = spark.createDataFrame(d,['id','day','price','units'])
>>> mydf.show()
+---+---+-----+-----+
| id|day|price|units|
+---+---+-----+-----+
|100|  1|   23|   10|
|100|  2|   45|   11|
|100|  3|   67|   12|
|100|  4|   78|   13|
|101|  1|   23|   10|
|101|  2|   45|   13|
|101|  3|   67|   14|
|101|  4|   78|   15|
|102|  1|   23|   10|
|102|  2|   45|   11|
|102|  3|   67|   16|
|102|  4|   78|   18|
+---+---+-----+-----+

Now,if I need to get price column into a row for each id based on day, then I can use pivot method as,

>>> pvtdf = mydf.withColumn('combcol',F.concat(F.lit('price_'),mydf['day'])).groupby('id').pivot('combcol').agg(F.first('price'))
>>> pvtdf.show()
+---+-------+-------+-------+-------+
| id|price_1|price_2|price_3|price_4|
+---+-------+-------+-------+-------+
|100|     23|     45|     67|     78|
|101|     23|     45|     67|     78|
|102|     23|     45|     67|     78|
+---+-------+-------+-------+-------+

so when I need units column as well to be transposed as price, either I got to create one more dataframe as above for units and then join both using id.But, when I have more columns as such, I tried a function to do it,

>>> def pivot_udf(df,*cols):
...     mydf = df.select('id').drop_duplicates()
...     for c in cols:
...        mydf = mydf.join(df.withColumn('combcol',F.concat(F.lit('{}_'.format(c)),df['day'])).groupby('id').pivot('combcol').agg(F.first(c)),'id')
...     return mydf
...
>>> pivot_udf(mydf,'price','units').show()
+---+-------+-------+-------+-------+-------+-------+-------+-------+
| id|price_1|price_2|price_3|price_4|units_1|units_2|units_3|units_4|
+---+-------+-------+-------+-------+-------+-------+-------+-------+
|100|     23|     45|     67|     78|     10|     11|     12|     13|
|101|     23|     45|     67|     78|     10|     13|     14|     15|
|102|     23|     45|     67|     78|     10|     11|     16|     18|
+---+-------+-------+-------+-------+-------+-------+-------+-------+

Need suggestions on ,if it is good practice to do so and if any other better way of doing it. Thanks in advance!

Answer 1

Here's a non-UDF way involving a single pivot (hence, just a single column scan to identify all the unique dates).

dff = mydf.groupBy('id').pivot('day').agg(F.first('price').alias('price'),F.first('units').alias('unit'))

Here's the result (apologies for the non-matching ordering and naming):

+---+-------+------+-------+------+-------+------+-------+------+               
| id|1_price|1_unit|2_price|2_unit|3_price|3_unit|4_price|4_unit|
+---+-------+------+-------+------+-------+------+-------+------+
|100|     23|    10|     45|    11|     67|    12|     78|    13|
|101|     23|    10|     45|    13|     67|    14|     78|    15|
|102|     23|    10|     45|    11|     67|    16|     78|    18|
+---+-------+------+-------+------+-------+------+-------+------+

We just aggregate both on the price and the unit column after pivoting on the day.

If naming required as in question,

dff.select([F.col(c).name('_'.join(x for x in c.split('_')[::-1])) for c in dff.columns]).show()

+---+-------+------+-------+------+-------+------+-------+------+
| id|price_1|unit_1|price_2|unit_2|price_3|unit_3|price_4|unit_4|
+---+-------+------+-------+------+-------+------+-------+------+
|100|     23|    10|     45|    11|     67|    12|     78|    13|
|101|     23|    10|     45|    13|     67|    14|     78|    15|
|102|     23|    10|     45|    11|     67|    16|     78|    18|
+---+-------+------+-------+------+-------+------+-------+------+

Answer 2

The solution in the question is the best I could get. The only improvement would be to cache the input dataset to avoid double scan, i.e.

mydf.cache
pivot_udf(mydf,'price','units').show()

Answer 3

As in spark 1.6 version I think that's the only way because pivot takes only one column and there is second attribute values on which you can pass the distinct values of that column that will make your code run faster because otherwise spark has to run that for you, so yes that's the right way to do it.