Spark: aggregate versus map and reduce - apache-spark

I'm learning Spark and start understanding how Spark distributes the data and combines the results.
I came to the conclusion that using the operation map followed by reduce has an advantage on using just the operation aggregate. This is (at least I believe so) because aggregate uses a sequential operation, which hurts parallelism, while map and reduce can benefit from full parallelism.
So when having a choice, isn't it better to use map and reduce than aggregate ? Are there cases where aggregate is preferred ? Or maybe when aggregate can't be replaced by the combination map and reduce ?
As an example - I want to find the string with the max length:
val z = sc.parallelize(List("123","12","345","4567"))
// instead of this aggregate ....
z.aggregate(0)((x, y) => math.max(x, y.length), (x, y) => math.max(x, y))
// .... shouldn't I rather use this map - reduce combination ?
z.map(_.length).reduce((x, y) => math.max(x, y))

A little example will can be better than long explanations.
Imagine you have a class Toto with an age field. You have many Toto and you desire to compute sum of ages of every Toto.
final case class Toto(val age: Int)
val rdd = sc.parallelize(0 until n).map(Toto(_))
// map/reduce style
val sum1 = rdd
// O(n) operations to go througth every Toto's age
.map(_.age)
// another O(n) to access data then O(n) operations to sum the n values
.reduce(_ + _)
// You get the result with 2 pass over your data plus O(n) additions
// aggregate style
val sum2 = rdd.aggregate(0)((agg, e) => agg + e.age, _ + _)
// With one pass over the data, and O(n) additions you obtain the same result
It's a bit more complicate if you take into account access and each operations.
Because aggregate still access then sum the age into the aggregate wich represent O(2.n) operations, O(n) access plus O(n) additions, plus negligeable merged operation between aggregates.
On the other side with map/reduce style, first the map represent O(n) access, then again O(n) access to data to reduce them with an overhead of O(n) addition operations for a total of O(3.n) operations.
Without forgetting the fact that Spark is lazy and all of your transformation will be leverage by a final action.
I presume that using aggregate will save some operations and then will improve application running time. But depending on what you're doing it could be more usefull to express successive map followed by a reduce for readability compare to an aggregate or combineByKey (generalization of aggregateByKey). So i will suppose that it depends on which goals you desire to reach depending the use case.

I believe I can partially answer my own question. I was wrongly assuming that, because a sequential operation is used, aggregate might be hurt in its parallelism. The data can still be parallelized and the sequential op will be executed on each chunk. This doesn't seem less performing than the map operation. So then the question that remains is: why would you use aggregate as opposed to the map-reduce combination ?

Aggregate operation allows to specify a combiner function (to reduce the amount of data sent through the shuffle), which is different to reducer, with map-reduce combination the same function is used to combine and reduce. I know used old Map Reduce terminology but conceptually all shared nothing shuffle based frameworks do this and if you google for mapreduce combiner you will find a lot of explanations of the concept.

Related

reduce, reduceByKey, reduceGroups in Spark or Flink

reduce: function takes accumulated value and next value to find some aggregation.
reduceByKey: is also the same operation with specified key.
reduceGroups: is apply specified operation to the grouped data.
I don't know how memory managed for these operations. For example, how data is taken while using reduce function(e.g all data loaded to the memory?)? I want to know how data is managed for reduce operations. I also want to know what is the difference between these operations according to the data management.
Reduce is one of the cheapest operations in Spark,since that the only thing it does is actually grouping similar data to the same node.The only cost of a reduce operation is the reading of the tuple and a decision of where it should be grouped.
This means that the simple reduce,in contrast to the reduceByKey or reduceGroups is more expensive because Spark does not know how to make the grouping and searches for correlations among tuples.
Reduce can also ignore a tuple if it does not meet any requirement.

Why does collect_list in Spark not use partial aggregation

I recently played around with UDAFs and looked into the sourcecode of the built-in aggregation function collect_list, I was suprised to see that collect_list does not have a merge method implemented, although I think this is really straight-farward (just concatenate two Arrays). Code taken from org.apache.spark.sql.catalyst.expressions.aggregate.collect.Collect
override def merge(buffer: InternalRow, input: InternalRow): Unit = {
sys.error("Collect cannot be used in partial aggregations.")
}
It is no longer the case, as SPARK-1893 but I'd assume that the initial design had mostly collect_list in mind.
Because collect_list is logically equivalent to groupByKey the motivation would be exactly the same to avoid long GC pauses. In particular map side combine in groupByKey has been disabled with Spark SPARK-772:
Map side combine in group by key case does not reduce the amount of data shuffled. Instead, it forces a lot more objects to go into old gen, and leads to worse GC.
So to address you comment
I think this is really straight-farward (just concatenate two Arrays).
It might be simple but it doesn't add much value (unless there is another reducing operation on top of it) and sequence concatenation is expensive.

Difference between reduce and reduceByKey in Apache Spark

What is the difference between reduce and reduceByKey in Apache Spark in terms of their functionalities?
Why reduceByKey is a transformation and reduce is an action?
This is close to a duplicate of my answer explaining reduceByKey, but I will elaborate to the specific part that makes the two different. However refer to my answer for a bit more specifics on the internals of reduceByKey.
Basically, reduce must pull the entire dataset down into a single location because it is reducing to one final value. reduceByKey on the other hand is one value for each key. And since this action can be run on each machine locally first then it can remain an RDD and have further transformations done on its dataset.
Note, however that there is a reduceByKeyLocally you can use to automatically pull down the Map to a single location also.
Please go through this official documentation link .
reduce is an action which Aggregate the elements of the dataset using a function func (which takes two arguments and returns one),also we can use reduce for single RDDs (for more info Please click HERE).
reduceByKey When called on a dataset of (K, V) pairs, returns a dataset of (K, V) pairs where the values for each key are aggregated using the given reduce function func, which must be of type (V,V) => V. (for more info Please click HERE)
this is the qt assistant :
reduce(f): Reduces the elements of this RDD using the specified
commutative and associative binary operator. Currently reduces
partitions locally.
reduceByKey(func, numPartitions=None, partitionFunc=) :
Merge the values for each key using an associative and commutative reduce
function.

Performance benefits of DataSet over RDD

After reading few great articles (this, this and this) about Spark's DataSets, I finishing with next DataSet's performance benefits over RDD:
Logical and physical plan optimization;
Strict typization;
Vectorized operations;
Low level memory management.
Questions:
Spark's RDD also builds physical plan and can combine/optimize multiple transformations at the same stage. Then what is the benefit of DataSet over RDD?
From the first link you can see an example of RDD[Person]. Does DataSet have advanced typization?
What do they mean by "vectorized operations"?
As I understand, DataSet's low memory management = advanced serialization. That means off-heap storage of serializable objects, where you can read only one field of an object without deserialization. But how about the situation when you have IN_MEMORY_ONLY persistence strategy? Will DataSet serialize everything any case? Will it have any performance benefit over RDD?
Spark's RDD also builds physical plan and can combine/optimize multiple transformations at the same stage. Than what is the benefit of DataSet over RDD?
When working with RDD what you write is what you get. While certain transformations are optimized by chaining, the execution plan is direct translation of the DAG. For example:
rdd.mapPartitions(f).mapPartitions(g).mapPartitions(h).shuffle()
where shuffle is an arbitrary shuffling transformation (*byKey, repartition, etc.) all three mapPartitions (map, flatMap, filter) will be chained without creating intermediate objects but cannot be rearranged.
Compared to that Datasets use significantly more restrictive programming model but can optimize execution using a number of techniques including:
Selection (filter) pushdown. For example if you have:
df.withColumn("foo", col("bar") + 1).where(col("bar").isNotNull())
can be executed as:
df.where(col("bar").isNotNull()).withColumn("foo", col("bar") + 1)
Early projections (select) and eliminations. For example:
df.withColumn("foo", col("bar") + 1).select("foo", "bar")
can be rewritten as:
df.select("foo", "bar").withColumn("foo", col("bar") + 1)
to avoid fetching and passing obsolete data. In the extreme case it can eliminate particular transformation completely:
df.withColumn("foo", col("bar") + 1).select("bar")
can be optimized to
df.select("bar")
These optimizations are possible for two reasons:
Restrictive data model which enables dependency analysis without complex and unreliable static code analysis.
Clear operator semantics. Operators are side effects free and we clearly distinguish between deterministic and nondeterministic ones.
To make it clear let's say we have a following data model:
case class Person(name: String, surname: String, age: Int)
val people: RDD[Person] = ???
And we want to retrieve surnames of all people older than 21. With RDD it can be expressed as:
people
.map(p => (p.surname, p.age)) // f
.filter { case (_, age) => age > 21 } // g
Now let's ask ourselves a few questions:
What is the relationship between the input age in f and age variable with g?
Is f and then g the same as g and then f?
Are f and g side effects free?
While the answer is obvious for a human reader it is not for a hypothetical optimizer. Compared to that with Dataframe version:
people.toDF
.select(col("surname"), col("age")) // f'
.where(col("age") > 21) // g'
the answers are clear for both optimizer and human reader.
This has some further consequences when using statically typed Datasets (Spark 2.0 Dataset vs DataFrame).
Have DataSet got more advanced typization?
No - if you care about optimizations. The most advanced optimizations are limited to Dataset[Row] and at this moment it is not possible to encode complex type hierarchy.
Maybe - if you accept overhead of the Kryo or Java encoders.
What does they mean by "vectorized operations"?
In context of optimization we usually mean loop vectorization / loop unrolling. Spark SQL uses code generation to create compiler friendly version of the high level transformations which can be further optimized to take advantage of the vectorized instruction sets.
As I understand, DataSet's low memory management = advanced serialization.
Not exactly. The biggest advantage of using native allocation is escaping garbage collector loop. Since garbage collections is quite often a limiting factor in Spark this is a huge improvement, especially in contexts which require large data structures (like preparing shuffles).
Another important aspect is columnar storage which enables effective compression (potentially lower memory footprint) and optimized operations on compressed data.
In general you can apply exactly the same types of optimizations using hand crafted code on plain RDDs. After all Datasets are backed by RDDs. The difference is only how much effort it takes.
Hand crafted execution plan optimizations are relatively simple to achieve.
Making code compiler friendly requires some deeper knowledge and is error prone and verbose.
Using sun.misc.Unsafe with native memory allocation is not for the faint-hearted.
Despite all its merits Dataset API is not universal. While certain types of common tasks can benefit from its optimizations in many contexts you may so no improvement whatsoever or even performance degradation compared to RDD equivalent.

Time Complexity for index and drop of first item in Data.Sequence

I was recently working on an implementation of calculating moving average from a stream of input, using Data.Sequence. I figured I could get the whole operation to be O(n) by using a deque.
My first attempt was (in my opinion) a bit more straightforward to read, but not a true a deque. It looked like:
let newsequence = (|>) sequence n
...
let dropFrontTotal = fromIntegral (newtotal - index newsequence 0)
let newsequence' = drop 1 newsequence.
...
According to the hackage docs for Data.Sequence, index should take O(log(min(i,n-i))) while drop should also take O(log(min(i,n-i))).
Here's my question:
If I do drop 1 someSequence, doesn't this mean a time complexity of O(log(min(1, (length someSequence)))), which in this case means: O(log(1))?
If so, isn't O(log(1)) effectively constant?
I had the same question for index someSequence 0: shouldn't that operation end up being O(log(0))?
Ultimately, I had enough doubts about my understanding that I resorted to using Criterion to benchmark the two implementations to prove that the index/drop version is slower (and the amount it's slower by grows with the input). The informal results on my machine can be seen at the linked gist.
I still don't really understand how to calculate time complexity for these operations, though, and I would appreciate any clarification anyone can provide.
What you suggest looks correct to me.
As a minor caveat remember that these are amortized complexity bounds, so a single operation could require more than constant time, but a long chain of operations will only require a constant times the number of the chain.
If you use criterion to benchmark and "reset" the state at every computation, you might see non-constant time costs, because the "reset" is preventing the amortization. It really depends on how you perform the test. If you start from a sequence an perform a long chain of operations on that, it should be OK. If you repeat many times a single operation using the same operands, then it could be not OK.
Further, I guess bounds such as O(log(...)) should actually be read as O(log(1 + ...)) -- you can't realistically have O(log(1)) = O(0) or, worse O(log(0))= O(-inf) as a complexity bound.

Resources