How does Spark Word2Vec merge each partition's results? - scala

Increasing numPartitions for Spark's Word2Vec makes it faster but less accurate since it fits each partition separately, reducing the context available for each word, before merging the results.
How exactly does it merge the results from multiple partitions? Is it just an average of the vectors? Looking to better understand how this affects the accuracy.
Looking at the source code, I think the merging is happening here:
val synAgg = partial.reduceByKey { case (v1, v2) =>
blas.saxpy(vectorSize, 1.0f, v2, 1, v1, 1)
v1
}.collect()
Which looks like just a vector sum (effectively an average). partial comes from:
val sentences: RDD[Array[Int]] = dataset.mapPartitions { sentenceIter =>
// Each sentence will map to 0 or more Array[Int]
sentenceIter.flatMap { sentence =>
// Sentence of words, some of which map to a word index
val wordIndexes = sentence.flatMap(bcVocabHash.value.get)
// break wordIndexes into trunks of maxSentenceLength when has more
wordIndexes.grouped(maxSentenceLength).map(_.toArray)
}
}
val newSentences = sentences.repartition(numPartitions).cache()
val partial = newSentences.mapPartitionsWithIndex { case (idx, iter) =>
// ... long calculation (skip-gram training, etc.)
}
But I'm not a Word2Vec/Spark ML/Scala expert, so hoping someone more knowledgeable can verify.

Related

Spark find previous value on each iteration of RDD

I've following code :-
val rdd = sc.cassandraTable("db", "table").select("id", "date", "gpsdt").where("id=? and date=? and gpsdt>? and gpsdt<?", entry(0), entry(1), entry(2) , entry(3))
val rddcopy = rdd.sortBy(row => row.get[String]("gpsdt"), false).zipWithIndex()
rddcopy.foreach { records =>
{
val previousRow = (records - 1)th row
val currentRow = records
// Some calculation based on both rows
}
}
So, Idea is to get just previous \ next row on each iteration of RDD. I want to calculate some field on current row based on the value present on previous row. Thanks,
EDIT II: Misunderstood question below is how to get tumbling window semantics but sliding window is needed. considering this is a sorted RDD
import org.apache.spark.mllib.rdd.RDDFunctions._
sortedRDD.sliding(2)
should do the trick. Note however that this is using a DeveloperAPI.
alternatively you can
val l = sortedRdd.zipWithIndex.map(kv => (kv._2, kv._1))
val r = sortedRdd.zipWithIndex.map(kv => (kv._2-1, kv._1))
val sliding = l.join(r)
rdd joins should be inner joins (IIRC) thus dropping the edge cases where the tuples would be partially null
OLD STUFF:
how do you do identify the previous row? RDDs do not have any sort of stable ordering by themselves. if you have an incrementing dense key you could add a new column that get's calculated the following way if (k % 2 == 0) k / 2 else (k-1)/2 this should give you a key that has the same value for two successive keys. Then you could just group by.
But to reiterate there is no really sensible notion of previous in most cases for RDDs (depending on partitioning, datasource etc.)
EDIT: so now that you have a zipWithIndex and an ordering in your set you can do what I mentioned above. So now you have an RDD[(Int, YourData)] and can do
rdd.map( kv => if (kv._1 % 2 == 0) (kv._1 / 2, kv._2) else ( (kv._1 -1) /2, kv._2 ) ).groupByKey.foreach (/* your stuff here /*)
if you reduce at any point consider using reduceByKey rather than groupByKey().reduce

Spark: flatMap/reduceByKey seems to be quite slow with Long keys on some distributions

I'm using Spark to process some corpora and I need to count the occurrence of each 2-gram. I started with counting tuples (wordID1, wordID2) and it worked fine except for the large memory usage and gc overhead due to the substantial number of small tuple objects. Then I tried to pack a pair of Ints into a Long, and the gc overhead did reduce greatly, but the run time also increased several times.
I ran some small experiments with random data on different distributions. It seems that the performance issue only occurs on exponential distributed data.
// lines of word IDs
val data = (1 to 5000).par.map({ _ =>
(1 to 1000) map { _ => (-1000 * Math.log(Random.nextDouble)).toInt }
}).seq
// count Tuples, fast
sc parallelize(data) flatMap { line =>
val first = line.iterator
val second = line.iterator.drop(1)
for (pair <- first zip(second))
yield (pair, 1L)
} reduceByKey { _ + _ } count()
// count Long, slow
sc parallelize(data) flatMap { line =>
val first = line.iterator
val second = line.iterator.drop(1)
for ((a, b) <- first zip(second))
yield ((a.toLong << 32) | b, 1L)
} reduceByKey { _ + _ } count()
The job is split into two stages, flatMap() and count(). When counting Tuple2s, flatMap() takes about 6s and count() takes about 2s, while when counting Longs, flatMap() takes 18s and count() takes 10s.
It doesn't make sense to me as Longs should impose less overhead than Tuple2. Does spark has some specializations for Long keys, which happen to perform even worse for some specific distributions?
Thanks to #SarveshKumarSingh's hint, I finally solved the problem. It is not the Spark's specialization for Long that trigger the issue, but Java's, and Spark doesn't address it properly.
Java's hashCode() for Long is quite simple and strongly dependent on the two halves of the values, and Spark's default HashPartitioner simply partition keys according their hashCode() values modulo the partition number. These make Spark's default partitioning quite sensitive to the distribution of Long keys, especially when the number of partitions is relatively small. And in my case, the situation deteriorates as the Long keys are constructed via concatenating pairs of Ints.
The solutions would be quite straightforward as we just need to somehow "shuffle" the keys, which makes the keys with similar frequencies distributed evenly.
The simplest way is to map each key into another unique value using some perfect hash function, and convert it back when the original key is required. This approach involves only small code changes, but might not perform very well. I achieved performance similar to the count-by-tuple approach using the following mappings.
val newKey = oldKey * 6364136223846793005L + 1442695040888963407L
val oldKey = (newKey - 1442695040888963407L) * -4568919932995229531L
A more effective way is to substitute the default HashPartitioner. I used the following partitioner between flatMap and reduceByKey and achieved two times performance boost on real world data.
val prevRDD = // ... flatMap ...
val nParts = prevRDD.partitioner match {
case Some(p) => p.numPartitions
case None => prevRDD.partitions.size
}
prevRDD partitionBy (new Partitioner {
override def getPartition(key: Any): Int = {
val rawMod = LongHash(key.asInstanceOf[Long]) % numPartitions
rawMod + (if (rawMod < 0) numPartitions else 0)
}
override def numPartitions: Int = nParts
}) reduceByKey { _ + _ }
def LongHash(v: Long) = { // the 64bit mix function from Murmurhash3
var k = v
k ^= k >> 33
k *= 0xff51afd7ed558ccdL
k ^= k >> 33
k *= 0xc4ceb9fe1a85ec53L
k ^= k >> 33
k.toInt
}

Selecting every 3rd element from a huge RDD [duplicate]

I'm looking for a way to split an RDD into two or more RDDs. The closest I've seen is Scala Spark: Split collection into several RDD? which is still a single RDD.
If you're familiar with SAS, something like this:
data work.split1, work.split2;
set work.preSplit;
if (condition1)
output work.split1
else if (condition2)
output work.split2
run;
which resulted in two distinct data sets. It would have to be immediately persisted to get the results I intend...
It is not possible to yield multiple RDDs from a single transformation*. If you want to split a RDD you have to apply a filter for each split condition. For example:
def even(x): return x % 2 == 0
def odd(x): return not even(x)
rdd = sc.parallelize(range(20))
rdd_odd, rdd_even = (rdd.filter(f) for f in (odd, even))
If you have only a binary condition and computation is expensive you may prefer something like this:
kv_rdd = rdd.map(lambda x: (x, odd(x)))
kv_rdd.cache()
rdd_odd = kv_rdd.filter(lambda kv: kv[1]).keys()
rdd_even = kv_rdd.filter(lambda kv: not kv[1]).keys()
It means only a single predicate computation but requires additional pass over all data.
It is important to note that as long as an input RDD is properly cached and there no additional assumptions regarding data distribution there is no significant difference when it comes to time complexity between repeated filter and for-loop with nested if-else.
With N elements and M conditions number of operations you have to perform is clearly proportional to N times M. In case of for-loop it should be closer to (N + MN) / 2 and repeated filter is exactly NM but at the end of the day it is nothing else than O(NM). You can see my discussion** with Jason Lenderman to read about some pros-and-cons.
At the very high level you should consider two things:
Spark transformations are lazy, until you execute an action your RDD is not materialized
Why does it matter? Going back to my example:
rdd_odd, rdd_even = (rdd.filter(f) for f in (odd, even))
If later I decide that I need only rdd_odd then there is no reason to materialize rdd_even.
If you take a look at your SAS example to compute work.split2 you need to materialize both input data and work.split1.
RDDs provide a declarative API. When you use filter or map it is completely up to Spark engine how this operation is performed. As long as the functions passed to transformations are side effects free it creates multiple possibilities to optimize a whole pipeline.
At the end of the day this case is not special enough to justify its own transformation.
This map with filter pattern is actually used in a core Spark. See my answer to How does Sparks RDD.randomSplit actually split the RDD and a relevant part of the randomSplit method.
If the only goal is to achieve a split on input it is possible to use partitionBy clause for DataFrameWriter which text output format:
def makePairs(row: T): (String, String) = ???
data
.map(makePairs).toDF("key", "value")
.write.partitionBy($"key").format("text").save(...)
* There are only 3 basic types of transformations in Spark:
RDD[T] => RDD[T]
RDD[T] => RDD[U]
(RDD[T], RDD[U]) => RDD[W]
where T, U, W can be either atomic types or products / tuples (K, V). Any other operation has to be expressed using some combination of the above. You can check the original RDD paper for more details.
** https://chat.stackoverflow.com/rooms/91928/discussion-between-zero323-and-jason-lenderman
*** See also Scala Spark: Split collection into several RDD?
As other posters mentioned above, there is no single, native RDD transform that splits RDDs, but here are some "multiplex" operations that can efficiently emulate a wide variety of "splitting" on RDDs, without reading multiple times:
http://silex.freevariable.com/latest/api/#com.redhat.et.silex.rdd.multiplex.MuxRDDFunctions
Some methods specific to random splitting:
http://silex.freevariable.com/latest/api/#com.redhat.et.silex.sample.split.SplitSampleRDDFunctions
Methods are available from open source silex project:
https://github.com/willb/silex
A blog post explaining how they work:
http://erikerlandson.github.io/blog/2016/02/08/efficient-multiplexing-for-spark-rdds/
def muxPartitions[U :ClassTag](n: Int, f: (Int, Iterator[T]) => Seq[U],
persist: StorageLevel): Seq[RDD[U]] = {
val mux = self.mapPartitionsWithIndex { case (id, itr) =>
Iterator.single(f(id, itr))
}.persist(persist)
Vector.tabulate(n) { j => mux.mapPartitions { itr => Iterator.single(itr.next()(j)) } }
}
def flatMuxPartitions[U :ClassTag](n: Int, f: (Int, Iterator[T]) => Seq[TraversableOnce[U]],
persist: StorageLevel): Seq[RDD[U]] = {
val mux = self.mapPartitionsWithIndex { case (id, itr) =>
Iterator.single(f(id, itr))
}.persist(persist)
Vector.tabulate(n) { j => mux.mapPartitions { itr => itr.next()(j).toIterator } }
}
As mentioned elsewhere, these methods do involve a trade-off of memory for speed, because they operate by computing entire partition results "eagerly" instead of "lazily." Therefore, it is possible for these methods to run into memory problems on large partitions, where more traditional lazy transforms will not.
One way is to use a custom partitioner to partition the data depending upon your filter condition. This can be achieved by extending Partitioner and implementing something similar to the RangePartitioner.
A map partitions can then be used to construct multiple RDDs from the partitioned RDD without reading all the data.
val filtered = partitioned.mapPartitions { iter => {
new Iterator[Int](){
override def hasNext: Boolean = {
if(rangeOfPartitionsToKeep.contains(TaskContext.get().partitionId)) {
false
} else {
iter.hasNext
}
}
override def next():Int = iter.next()
}
Just be aware that the number of partitions in the filtered RDDs will be the same as the number in the partitioned RDD so a coalesce should be used to reduce this down and remove the empty partitions.
If you split an RDD using the randomSplit API call, you get back an array of RDDs.
If you want 5 RDDs returned, pass in 5 weight values.
e.g.
val sourceRDD = val sourceRDD = sc.parallelize(1 to 100, 4)
val seedValue = 5
val splitRDD = sourceRDD.randomSplit(Array(1.0,1.0,1.0,1.0,1.0), seedValue)
splitRDD(1).collect()
res7: Array[Int] = Array(1, 6, 11, 12, 20, 29, 40, 62, 64, 75, 77, 83, 94, 96, 100)

Collection.partition in spark [duplicate]

I'm looking for a way to split an RDD into two or more RDDs. The closest I've seen is Scala Spark: Split collection into several RDD? which is still a single RDD.
If you're familiar with SAS, something like this:
data work.split1, work.split2;
set work.preSplit;
if (condition1)
output work.split1
else if (condition2)
output work.split2
run;
which resulted in two distinct data sets. It would have to be immediately persisted to get the results I intend...
It is not possible to yield multiple RDDs from a single transformation*. If you want to split a RDD you have to apply a filter for each split condition. For example:
def even(x): return x % 2 == 0
def odd(x): return not even(x)
rdd = sc.parallelize(range(20))
rdd_odd, rdd_even = (rdd.filter(f) for f in (odd, even))
If you have only a binary condition and computation is expensive you may prefer something like this:
kv_rdd = rdd.map(lambda x: (x, odd(x)))
kv_rdd.cache()
rdd_odd = kv_rdd.filter(lambda kv: kv[1]).keys()
rdd_even = kv_rdd.filter(lambda kv: not kv[1]).keys()
It means only a single predicate computation but requires additional pass over all data.
It is important to note that as long as an input RDD is properly cached and there no additional assumptions regarding data distribution there is no significant difference when it comes to time complexity between repeated filter and for-loop with nested if-else.
With N elements and M conditions number of operations you have to perform is clearly proportional to N times M. In case of for-loop it should be closer to (N + MN) / 2 and repeated filter is exactly NM but at the end of the day it is nothing else than O(NM). You can see my discussion** with Jason Lenderman to read about some pros-and-cons.
At the very high level you should consider two things:
Spark transformations are lazy, until you execute an action your RDD is not materialized
Why does it matter? Going back to my example:
rdd_odd, rdd_even = (rdd.filter(f) for f in (odd, even))
If later I decide that I need only rdd_odd then there is no reason to materialize rdd_even.
If you take a look at your SAS example to compute work.split2 you need to materialize both input data and work.split1.
RDDs provide a declarative API. When you use filter or map it is completely up to Spark engine how this operation is performed. As long as the functions passed to transformations are side effects free it creates multiple possibilities to optimize a whole pipeline.
At the end of the day this case is not special enough to justify its own transformation.
This map with filter pattern is actually used in a core Spark. See my answer to How does Sparks RDD.randomSplit actually split the RDD and a relevant part of the randomSplit method.
If the only goal is to achieve a split on input it is possible to use partitionBy clause for DataFrameWriter which text output format:
def makePairs(row: T): (String, String) = ???
data
.map(makePairs).toDF("key", "value")
.write.partitionBy($"key").format("text").save(...)
* There are only 3 basic types of transformations in Spark:
RDD[T] => RDD[T]
RDD[T] => RDD[U]
(RDD[T], RDD[U]) => RDD[W]
where T, U, W can be either atomic types or products / tuples (K, V). Any other operation has to be expressed using some combination of the above. You can check the original RDD paper for more details.
** https://chat.stackoverflow.com/rooms/91928/discussion-between-zero323-and-jason-lenderman
*** See also Scala Spark: Split collection into several RDD?
As other posters mentioned above, there is no single, native RDD transform that splits RDDs, but here are some "multiplex" operations that can efficiently emulate a wide variety of "splitting" on RDDs, without reading multiple times:
http://silex.freevariable.com/latest/api/#com.redhat.et.silex.rdd.multiplex.MuxRDDFunctions
Some methods specific to random splitting:
http://silex.freevariable.com/latest/api/#com.redhat.et.silex.sample.split.SplitSampleRDDFunctions
Methods are available from open source silex project:
https://github.com/willb/silex
A blog post explaining how they work:
http://erikerlandson.github.io/blog/2016/02/08/efficient-multiplexing-for-spark-rdds/
def muxPartitions[U :ClassTag](n: Int, f: (Int, Iterator[T]) => Seq[U],
persist: StorageLevel): Seq[RDD[U]] = {
val mux = self.mapPartitionsWithIndex { case (id, itr) =>
Iterator.single(f(id, itr))
}.persist(persist)
Vector.tabulate(n) { j => mux.mapPartitions { itr => Iterator.single(itr.next()(j)) } }
}
def flatMuxPartitions[U :ClassTag](n: Int, f: (Int, Iterator[T]) => Seq[TraversableOnce[U]],
persist: StorageLevel): Seq[RDD[U]] = {
val mux = self.mapPartitionsWithIndex { case (id, itr) =>
Iterator.single(f(id, itr))
}.persist(persist)
Vector.tabulate(n) { j => mux.mapPartitions { itr => itr.next()(j).toIterator } }
}
As mentioned elsewhere, these methods do involve a trade-off of memory for speed, because they operate by computing entire partition results "eagerly" instead of "lazily." Therefore, it is possible for these methods to run into memory problems on large partitions, where more traditional lazy transforms will not.
One way is to use a custom partitioner to partition the data depending upon your filter condition. This can be achieved by extending Partitioner and implementing something similar to the RangePartitioner.
A map partitions can then be used to construct multiple RDDs from the partitioned RDD without reading all the data.
val filtered = partitioned.mapPartitions { iter => {
new Iterator[Int](){
override def hasNext: Boolean = {
if(rangeOfPartitionsToKeep.contains(TaskContext.get().partitionId)) {
false
} else {
iter.hasNext
}
}
override def next():Int = iter.next()
}
Just be aware that the number of partitions in the filtered RDDs will be the same as the number in the partitioned RDD so a coalesce should be used to reduce this down and remove the empty partitions.
If you split an RDD using the randomSplit API call, you get back an array of RDDs.
If you want 5 RDDs returned, pass in 5 weight values.
e.g.
val sourceRDD = val sourceRDD = sc.parallelize(1 to 100, 4)
val seedValue = 5
val splitRDD = sourceRDD.randomSplit(Array(1.0,1.0,1.0,1.0,1.0), seedValue)
splitRDD(1).collect()
res7: Array[Int] = Array(1, 6, 11, 12, 20, 29, 40, 62, 64, 75, 77, 83, 94, 96, 100)

Creating multiple RDDs out of one RDD [duplicate]

I'm looking for a way to split an RDD into two or more RDDs. The closest I've seen is Scala Spark: Split collection into several RDD? which is still a single RDD.
If you're familiar with SAS, something like this:
data work.split1, work.split2;
set work.preSplit;
if (condition1)
output work.split1
else if (condition2)
output work.split2
run;
which resulted in two distinct data sets. It would have to be immediately persisted to get the results I intend...
It is not possible to yield multiple RDDs from a single transformation*. If you want to split a RDD you have to apply a filter for each split condition. For example:
def even(x): return x % 2 == 0
def odd(x): return not even(x)
rdd = sc.parallelize(range(20))
rdd_odd, rdd_even = (rdd.filter(f) for f in (odd, even))
If you have only a binary condition and computation is expensive you may prefer something like this:
kv_rdd = rdd.map(lambda x: (x, odd(x)))
kv_rdd.cache()
rdd_odd = kv_rdd.filter(lambda kv: kv[1]).keys()
rdd_even = kv_rdd.filter(lambda kv: not kv[1]).keys()
It means only a single predicate computation but requires additional pass over all data.
It is important to note that as long as an input RDD is properly cached and there no additional assumptions regarding data distribution there is no significant difference when it comes to time complexity between repeated filter and for-loop with nested if-else.
With N elements and M conditions number of operations you have to perform is clearly proportional to N times M. In case of for-loop it should be closer to (N + MN) / 2 and repeated filter is exactly NM but at the end of the day it is nothing else than O(NM). You can see my discussion** with Jason Lenderman to read about some pros-and-cons.
At the very high level you should consider two things:
Spark transformations are lazy, until you execute an action your RDD is not materialized
Why does it matter? Going back to my example:
rdd_odd, rdd_even = (rdd.filter(f) for f in (odd, even))
If later I decide that I need only rdd_odd then there is no reason to materialize rdd_even.
If you take a look at your SAS example to compute work.split2 you need to materialize both input data and work.split1.
RDDs provide a declarative API. When you use filter or map it is completely up to Spark engine how this operation is performed. As long as the functions passed to transformations are side effects free it creates multiple possibilities to optimize a whole pipeline.
At the end of the day this case is not special enough to justify its own transformation.
This map with filter pattern is actually used in a core Spark. See my answer to How does Sparks RDD.randomSplit actually split the RDD and a relevant part of the randomSplit method.
If the only goal is to achieve a split on input it is possible to use partitionBy clause for DataFrameWriter which text output format:
def makePairs(row: T): (String, String) = ???
data
.map(makePairs).toDF("key", "value")
.write.partitionBy($"key").format("text").save(...)
* There are only 3 basic types of transformations in Spark:
RDD[T] => RDD[T]
RDD[T] => RDD[U]
(RDD[T], RDD[U]) => RDD[W]
where T, U, W can be either atomic types or products / tuples (K, V). Any other operation has to be expressed using some combination of the above. You can check the original RDD paper for more details.
** https://chat.stackoverflow.com/rooms/91928/discussion-between-zero323-and-jason-lenderman
*** See also Scala Spark: Split collection into several RDD?
As other posters mentioned above, there is no single, native RDD transform that splits RDDs, but here are some "multiplex" operations that can efficiently emulate a wide variety of "splitting" on RDDs, without reading multiple times:
http://silex.freevariable.com/latest/api/#com.redhat.et.silex.rdd.multiplex.MuxRDDFunctions
Some methods specific to random splitting:
http://silex.freevariable.com/latest/api/#com.redhat.et.silex.sample.split.SplitSampleRDDFunctions
Methods are available from open source silex project:
https://github.com/willb/silex
A blog post explaining how they work:
http://erikerlandson.github.io/blog/2016/02/08/efficient-multiplexing-for-spark-rdds/
def muxPartitions[U :ClassTag](n: Int, f: (Int, Iterator[T]) => Seq[U],
persist: StorageLevel): Seq[RDD[U]] = {
val mux = self.mapPartitionsWithIndex { case (id, itr) =>
Iterator.single(f(id, itr))
}.persist(persist)
Vector.tabulate(n) { j => mux.mapPartitions { itr => Iterator.single(itr.next()(j)) } }
}
def flatMuxPartitions[U :ClassTag](n: Int, f: (Int, Iterator[T]) => Seq[TraversableOnce[U]],
persist: StorageLevel): Seq[RDD[U]] = {
val mux = self.mapPartitionsWithIndex { case (id, itr) =>
Iterator.single(f(id, itr))
}.persist(persist)
Vector.tabulate(n) { j => mux.mapPartitions { itr => itr.next()(j).toIterator } }
}
As mentioned elsewhere, these methods do involve a trade-off of memory for speed, because they operate by computing entire partition results "eagerly" instead of "lazily." Therefore, it is possible for these methods to run into memory problems on large partitions, where more traditional lazy transforms will not.
One way is to use a custom partitioner to partition the data depending upon your filter condition. This can be achieved by extending Partitioner and implementing something similar to the RangePartitioner.
A map partitions can then be used to construct multiple RDDs from the partitioned RDD without reading all the data.
val filtered = partitioned.mapPartitions { iter => {
new Iterator[Int](){
override def hasNext: Boolean = {
if(rangeOfPartitionsToKeep.contains(TaskContext.get().partitionId)) {
false
} else {
iter.hasNext
}
}
override def next():Int = iter.next()
}
Just be aware that the number of partitions in the filtered RDDs will be the same as the number in the partitioned RDD so a coalesce should be used to reduce this down and remove the empty partitions.
If you split an RDD using the randomSplit API call, you get back an array of RDDs.
If you want 5 RDDs returned, pass in 5 weight values.
e.g.
val sourceRDD = val sourceRDD = sc.parallelize(1 to 100, 4)
val seedValue = 5
val splitRDD = sourceRDD.randomSplit(Array(1.0,1.0,1.0,1.0,1.0), seedValue)
splitRDD(1).collect()
res7: Array[Int] = Array(1, 6, 11, 12, 20, 29, 40, 62, 64, 75, 77, 83, 94, 96, 100)