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 is that if I can get them partitioned (by a field call date_day) and co-located then I would avoid shuffling .

Here is what I am trying to do for each table:

def date_day_partitioner(key):
  return (key.date_day - datetime.date(2017,05,01)).days

df = sqlContext.sql("select * from hive.table")
rdd = df.rdd
rdd2 = rdd.partitionBy(100, date_day_partitioner)
df2 = sqlContext.createDataFrame(rdd2, df_log_entry.schema)

print df2.count()

Unfortunately, I can't even test my theory about co-location and avoiding shuffling, because I get the following error when I try partitionBy: ValueError: too many values to unpack

Traceback (most recent call last):
  File "/tmp/zeppelin_pyspark-118755547579363441.py", line 346, in <module>
    raise Exception(traceback.format_exc())
Exception: Traceback (most recent call last):
  File "/tmp/zeppelin_pyspark-118755547579363441.py", line 339, in <module>
    exec(code)
  File "<stdin>", line 15, in <module>
  File "/usr/lib/spark/python/pyspark/sql/dataframe.py", line 380, in count
    return int(self._jdf.count())
  File "/usr/lib/spark/python/lib/py4j-0.10.4-src.zip/py4j/java_gateway.py", line 1133, in __call__
    answer, self.gateway_client, self.target_id, self.name)
  File "/usr/lib/spark/python/pyspark/sql/utils.py", line 63, in deco
    return f(*a, **kw)
  File "/usr/lib/spark/python/lib/py4j-0.10.4-src.zip/py4j/protocol.py", line 319, in get_return_value
    format(target_id, ".", name), value)
Py4JJavaError: An error occurred while calling o115.count.
: org.apache.spark.SparkException: Job aborted due to stage failure: Task 21 in stage 6.0 failed 4 times, most recent failure: Lost task 21.3 in stage 6.0 (TID 182, ip-172-31-49-209.ec2.internal, executor 3): org.apache.spark.api.python.PythonException: Traceback (most recent call last):
  File "/mnt/yarn/usercache/zeppelin/appcache/application_1509802099365_0013/container_1509802099365_0013_01_000007/pyspark.zip/pyspark/worker.py", line 174, in main
    process()
  File "/mnt/yarn/usercache/zeppelin/appcache/application_1509802099365_0013/container_1509802099365_0013_01_000007/pyspark.zip/pyspark/worker.py", line 169, in process
    serializer.dump_stream(func(split_index, iterator), outfile)
  File "/mnt/yarn/usercache/zeppelin/appcache/application_1509802099365_0013/container_1509802099365_0013_01_000007/pyspark.zip/pyspark/serializers.py", line 138, in dump_stream
    for obj in iterator:
  File "/usr/lib/spark/python/pyspark/rdd.py", line 1752, in add_shuffle_key
ValueError: too many values to unpack
    at org.apache.spark.api.python.PythonRunner$$anon$1.read(PythonRDD.scala:193)
    at org.apache.spark.api.python.PythonRunner$$anon$1.<init>(PythonRDD.scala:234)
    at org.apache.spark.api.python.PythonRunner.compute(PythonRDD.scala:152)
    at org.apache.spark.api.python.PythonRDD.compute(PythonRDD.scala:63)
    at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:323)
    at org.apache.spark.rdd.RDD.iterator(RDD.scala:287)
    at org.apache.spark.api.python.PairwiseRDD.compute(PythonRDD.scala:390)
    at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:323)
    at org.apache.spark.rdd.RDD.iterator(RDD.scala:287)
...

I must be doing something wrong, could you please help?

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.