1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
|
package spark.rdd
import spark.{Utils, OneToOneDependency, RDD, SparkContext, Partition, TaskContext}
import java.io.{ObjectOutputStream, IOException}
private[spark] class ZippedPartitionsPartition(
idx: Int,
@transient rdds: Seq[RDD[_]])
extends Partition {
override val index: Int = idx
var partitionValues = rdds.map(rdd => rdd.partitions(idx))
def partitions = partitionValues
@throws(classOf[IOException])
private def writeObject(oos: ObjectOutputStream) {
// Update the reference to parent split at the time of task serialization
partitionValues = rdds.map(rdd => rdd.partitions(idx))
oos.defaultWriteObject()
}
}
abstract class ZippedPartitionsBaseRDD[V: ClassManifest](
sc: SparkContext,
var rdds: Seq[RDD[_]])
extends RDD[V](sc, rdds.map(x => new OneToOneDependency(x))) {
override def getPartitions: Array[Partition] = {
val sizes = rdds.map(x => x.partitions.size)
if (!sizes.forall(x => x == sizes(0))) {
throw new IllegalArgumentException("Can't zip RDDs with unequal numbers of partitions")
}
val array = new Array[Partition](sizes(0))
for (i <- 0 until sizes(0)) {
array(i) = new ZippedPartitionsPartition(i, rdds)
}
array
}
override def getPreferredLocations(s: Partition): Seq[String] = {
// Note that as number of rdd's increase and/or number of slaves in cluster increase, the computed preferredLocations below
// become diminishingly small : so we might need to look at alternate strategies to alleviate this.
// If there are no (or very small number of preferred locations), we will end up transferred the blocks to 'any' node in the
// cluster - paying with n/w and cache cost.
// Maybe pick a node which figures max amount of time ?
// Choose node which is hosting 'larger' of some subset of blocks ?
// Look at rack locality to ensure chosen host is atleast rack local to both hosting node ?, etc (would be good to defer this if possible)
val splits = s.asInstanceOf[ZippedPartitionsPartition].partitions
val rddSplitZip = rdds.zip(splits)
// exact match.
val exactMatchPreferredLocations = rddSplitZip.map(x => x._1.preferredLocations(x._2))
val exactMatchLocations = exactMatchPreferredLocations.reduce((x, y) => x.intersect(y))
// Remove exact match and then do host local match.
val exactMatchHosts = exactMatchLocations.map(Utils.parseHostPort(_)._1)
val matchPreferredHosts = exactMatchPreferredLocations.map(locs => locs.map(Utils.parseHostPort(_)._1))
.reduce((x, y) => x.intersect(y))
val otherNodeLocalLocations = matchPreferredHosts.filter { s => !exactMatchHosts.contains(s) }
otherNodeLocalLocations ++ exactMatchLocations
}
override def clearDependencies() {
super.clearDependencies()
rdds = null
}
}
class ZippedPartitionsRDD2[A: ClassManifest, B: ClassManifest, V: ClassManifest](
sc: SparkContext,
f: (Iterator[A], Iterator[B]) => Iterator[V],
var rdd1: RDD[A],
var rdd2: RDD[B])
extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2)) {
override def compute(s: Partition, context: TaskContext): Iterator[V] = {
val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions
f(rdd1.iterator(partitions(0), context), rdd2.iterator(partitions(1), context))
}
override def clearDependencies() {
super.clearDependencies()
rdd1 = null
rdd2 = null
}
}
class ZippedPartitionsRDD3
[A: ClassManifest, B: ClassManifest, C: ClassManifest, V: ClassManifest](
sc: SparkContext,
f: (Iterator[A], Iterator[B], Iterator[C]) => Iterator[V],
var rdd1: RDD[A],
var rdd2: RDD[B],
var rdd3: RDD[C])
extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2, rdd3)) {
override def compute(s: Partition, context: TaskContext): Iterator[V] = {
val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions
f(rdd1.iterator(partitions(0), context),
rdd2.iterator(partitions(1), context),
rdd3.iterator(partitions(2), context))
}
override def clearDependencies() {
super.clearDependencies()
rdd1 = null
rdd2 = null
rdd3 = null
}
}
class ZippedPartitionsRDD4
[A: ClassManifest, B: ClassManifest, C: ClassManifest, D:ClassManifest, V: ClassManifest](
sc: SparkContext,
f: (Iterator[A], Iterator[B], Iterator[C], Iterator[D]) => Iterator[V],
var rdd1: RDD[A],
var rdd2: RDD[B],
var rdd3: RDD[C],
var rdd4: RDD[D])
extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2, rdd3, rdd4)) {
override def compute(s: Partition, context: TaskContext): Iterator[V] = {
val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions
f(rdd1.iterator(partitions(0), context),
rdd2.iterator(partitions(1), context),
rdd3.iterator(partitions(2), context),
rdd4.iterator(partitions(3), context))
}
override def clearDependencies() {
super.clearDependencies()
rdd1 = null
rdd2 = null
rdd3 = null
rdd4 = null
}
}
|
