The takeOrdered Method from pyspark.RDD gets the N elements from an RDD ordered in ascending order or as specified by the optional key function as described here pyspark.RDD.takeOrdered. The example shows the following code with one key:
>>> sc.parallelize([10, 1, 2, 9, 3, 4, 5, 6, 7], 2).takeOrdered(6, key=lambda x: -x)
[10, 9, 7, 6, 5, 4]
Is it also possible to define more keys e.g. x,y,z for data that has 3 columns?
The keys should be in different order such as x= asc, y= desc, z=asc. That means if the first value x of two rows are equal then the second value y should be used in descending order.
For numeric you could write:
n = 1
rdd = sc.parallelize([
(-1, 99, 1), (-1, -99, -1), (5, 3, 8), (-1, 99, -1)
])
rdd.takeOrdered(n, lambda x: (x[0], -x[1], x[2]))
# [(-1, 99, -1)]
For other objects you can define some type of record type and define your own set of rich comparison methods:
class XYZ(object):
slots = ["x", "y", "z"]
def __init__(self, x, y, z):
self.x, self.y, self.z = x, y, z
def __eq__(self, other):
if not isinstance(other, XYZ):
return False
return self.x == other.x and self.y == other.y and self.z == other.z
def __lt__(self, other):
if not isinstance(other, XYZ):
raise ValueError(
"'<' not supported between instances of 'XYZ' and '{0}'".format(
type(other)
))
if self.x == other.x:
if self.y == other.y:
return self.z < other.z
else:
return self.y > other.y
else:
return self.x < other.x
def __repr__(self):
return "XYZ({}, {}, {})".format(self.x, self.y, self.z)
@classmethod
def from_tuple(cls, xyz):
x, y, z = xyz
return cls(x, y, z)
and then:
from xyz import XYZ
rdd.map(XYZ.from_tuple).takeOrdered(n)
# [XYZ(-1, 99, -1)]
In practice just use SQL:
from pyspark.sql.functions import asc, desc
rdd.toDF(["x", "y", "z"]).orderBy(asc("x"), desc("y"), asc("z")).take(n)
# [Row(x=-1, y=99, z=-1)]