Custom Partitioner in Pyspark 2.1.0

Question

I read that RDDs with the same partitioner will be co-located. This is important to me because I want to join several large Hive tables that are not partitioned. My theory i

Answer 1

It's happening because you are not applying partitionBy on key-value pair rdd. Your rdd must be in key-value pair. Also, your key type should be integer. I don't have sample data for your hive table. So let's demonstrate the fact using below hive table:

I have created a below dataframe using hive table :

df = spark.table("udb.emp_details_table");
+------+--------+--------+----------------+
|emp_id|emp_name|emp_dept|emp_joining_date|
+------+--------+--------+----------------+
|     1|     AAA|      HR|      2018-12-06|
|     1|     BBB|      HR|      2017-10-26|
|     2|     XXX|   ADMIN|      2018-10-22|
|     2|     YYY|   ADMIN|      2015-10-19|
|     2|     ZZZ|      IT|      2018-05-14|
|     3|     GGG|      HR|      2018-06-30|
+------+--------+--------+----------------+

Now, I wish to partition my dataframe and want to keep the similar keys in one partition. So, I have converted my dataframe to rdd as you can only apply partitionBy on rdd for re-partitioning.

    myrdd = df.rdd
    newrdd = myrdd.partitionBy(10,lambda k: int(k[0]))
    newrdd.take(10)

I got the same error:

 File "/usr/hdp/current/spark2-client/python/pyspark/rdd.py", line 1767, in add_shuffle_key
    for k, v in iterator:
ValueError: too many values to unpack 

Hence, we need to convert our rdd into key-value pair to use paritionBy

keypair_rdd = myrdd.map(lambda x : (x[0],x[1:]))

Now,you can see that rdd has been converted to key value pair and you can therefore distribute your data in partitions according to keys available.

[(u'1', (u'AAA', u'HR', datetime.date(2018, 12, 6))), 
(u'1', (u'BBB', u'HR', datetime.date(2017, 10, 26))), 
(u'2', (u'XXX', u'ADMIN', datetime.date(2018, 10, 22))), 
(u'2', (u'YYY', u'ADMIN', datetime.date(2015, 10, 19))), 
(u'2', (u'ZZZ', u'IT', datetime.date(2018, 5, 14))), 
(u'3', (u'GGG', u'HR', datetime.date(2018, 6, 30)))]

Using a paritionBy on key-value rdd now:

newrdd = keypair_rdd.partitionBy(5,lambda k: int(k[0]))

Lets take a look at the partitions. Data is grouped and similar keys are stored into similar partitions now. Two of them are empty.

>>> print("Partitions structure: {}".format(newrdd.glom().map(len).collect()))
Partitions structure: [0, 2, 3, 1, 0]

Now lets say I want to custom partitioning my data. So I have created below function to keep keys '1' and '3' in similar partition.

def partitionFunc(key):
    import random
    if key == 1 or key == 3:
        return 0
    else:
        return random.randint(1,2)

newrdd = keypair_rdd.partitionBy(5,lambda k: partitionFunc(int(k[0])))

>>> print("Partitions structure: {}".format(newrdd.glom().map(len).collect()))
Partitions structure: [3, 3, 0, 0, 0]

As you can see now that keys 1 and 3 are stored in one partition and rest on other.

I hope this helps. You can try to partitionBy your dataframe. Make sure to convert it into key value pair and keeping key as type integer.