RDDs: Transformation and Actions

Loading...

En provenance du cours de École Polytechnique Fédérale de Lausanne

Big Data Analysis with Scala and Spark

1623 notes

École Polytechnique Fédérale de Lausanne

Big Data Analysis with Scala and Spark

1623 notes

Cours 4 sur 5 dans Specialization Functional Programming in Scala

Manipulating big data distributed over a cluster using functional concepts is rampant in industry, and is arguably one of the first widespread industrial uses of functional ideas. This is evidenced by the popularity of MapReduce and Hadoop, and most recently Apache Spark, a fast, in-memory distributed collections framework written in Scala. In this course, we'll see how the data parallel paradigm can be extended to the distributed case, using Spark throughout. We'll cover Spark's programming model in detail, being careful to understand how and when it differs from familiar programming models, like shared-memory parallel collections or sequential Scala collections. Through hands-on examples in Spark and Scala, we'll learn when important issues related to distribution like latency and network communication should be considered and how they can be addressed effectively for improved performance. Learning Outcomes. By the end of this course you will be able to: - read data from persistent storage and load it into Apache Spark, - manipulate data with Spark and Scala, - express algorithms for data analysis in a functional style, - recognize how to avoid shuffles and recomputation in Spark, Recommended background: You should have at least one year programming experience. Proficiency with Java or C# is ideal, but experience with other languages such as C/C++, Python, Javascript or Ruby is also sufficient. You should have some familiarity using the command line. This course is intended to be taken after Parallel Programming: https://www.coursera.org/learn/parprog1.

À partir de la leçon

Getting Started + Spark Basics

Get up and running with Scala on your computer. Complete an example assignment to familiarize yourself with our unique way of submitting assignments. In this week, we'll bridge the gap between data parallelism in the shared memory scenario (learned in the Parallel Programming course, prerequisite) and the distributed scenario. We'll look at important concerns that arise in distributed systems, like latency and failure. We'll go on to cover the basics of Spark, a functionally-oriented framework for big data processing in Scala. We'll end the first week by exercising what we learned about Spark by immediately getting our hands dirty analyzing a real-world data set.

RDDs, Spark's Distributed Collection9:18

RDDs: Transformation and Actions16:10

Evaluation in Spark: Unlike Scala Collections!20:36

Cluster Topology Matters!8:21

Rencontrer les enseignants

Dr. Heather Miller
Research Scientist
EPFL

In this session, we're going to focus on transformations and actions.

Transformations and actions are the different kinds of operations on RDDs.

To understand transformations and actions and its work, first recall transformers

and accessors from Scala's sequential and parallel collections.

If you don't remember what these terms mean, I will briefly remind you.

Transformers are operations that return new collections as a result.

So, not a single value.

If you, for example, call a method on a list and

you get back another list, typically this is a transformer.

So in this case, if we call a map on list, we get back another list.

An accessor is a method that returns a single value as a result and

not a collection.

So examples of accessors are things like reduce, fold, or aggregate.

So just to remind you what reduce does, it goes through and according to some kind

of function you would give it, it combines all of the elements together and

it will give you the single combined result back.

This is an accessor.

Now, where we had transformers, transformers and

accessors in regular Scala collections, we have in Spark

transformations instead of transformers and actions instead of accessors.

So the definition of a transformation, very similar to a transformer,

is an operation that returns not a collection but an RDD as a result.

And likewise, an action is something that computes a the result based on an RDD and

it returns or saves that result to an external storage system, but

it doesn't return an RDD.

It returns something like a value, or something that's not like an RDD.

But this is the most important thing, if you remember anything from this lecture,

you need to remember that transformations are lazy and actions are eager.

This is extremely important to remember.

This is an enormous difference from Scala collections.

So while you might have a method with a very similar signature like map,

similar to the signature of map on Scala's lists,

the fact that it's lazy and not eager is an enormous

semantic difference that you need keep in mind when you are writing Spark programs.

And the reason why this is so

important is because this is the way that Spark deals with the network.

The fact that transformation operations are lazy and

actions are eager lets us aggressively reduce the amount of

network communication that's required to undertake Spark jobs.

So if you recall, in an earlier session,

we were talking about how important latency was, and I

mentioned at some point that this is going to sneak out into the programming model.

This is how it gets into the programming model.

Okay, so let's make that more concrete by looking at an example.

So let's assume that we have a large list.

And you remember, we're going to use this parallelize method.

So I didn't define it here, but sc is a Spark context.

We'll see later more concretely how to create an instance of these.

But typically you always create a Spark context once in your program and

then you keep reusing it, you create RDDs from it.

And in this case, we can use that parallelize method we learned in the last

session to transform this large list of strings into an RDD now of strings.

Now this has type RDD of string.

And so now that we have an RDD of String, we can do a map on it.

We can go to each element and call length on each individual element,

then we'll get back an RDD now of lengths.

So now the type of lengthsRdd will be an RDD of integers.

Okay, cool.

So what's happening on the cluster now at this point?

Well, the answer is nothing.

Because remember, a map is a transformation method,

it's deferred, nothing happens.

All we get back is a reference to an RDD that doesn't yet exist.

So how de we ensure that this computation is done on the cluster?

What do we do?

Well, to kick off the computation and to wait for its result to be completed and

returned, we can add an action.

So in this case,

we get the total number of characters in the entire RDD of strings.

And we can do that using this reduce operation,

which we saw on a earlier slide is an action.

Reduce returns not another RDD, but a single value like an integer, for example.

In this case, we're going to get back an integer.

So this is important to remember.

This is probably one of the most common problems that

people have when they're just learning Spark.

They erroneously assume that after a map or a filter operation has been done,

that their computation has started or been completed.

When in reality, nothing happens on the cluster.

Nothing happens till invoking some type of action to return a result.

So this is going to be very important for

you remember in your first programming assignment.

So just to give you a flavor of some of the more

commonly used transformations that you'll see in Spark programs.

So there are the usual suspects here, map,

which, again, note that it's a transformation because of its return type.

In of all these cases, an RDD is being returned.

So these are transformations because the return type isn't RDD.

Which means these things are lazy, right?

I'm just going to remind you again, lazy operations.

So map we already know pretty well, apply a function to each element in the RDD and

then return another RDD of the result of all of these applications per element.

flatMap, same idea, but flatten the result.

So return an RDD of the contents of the iterators returned.

In the case of filter, we use this predicate function.

And we return an RDD of elements that have passed this predicate

condition in this case.

It's a function from some type to a Boolean.

And finally, distinct.

This is also another rather common operation.

And basically, this is kind of like distinct on a set operation.

So it will remove duplicated elements and return a new RDD without duplicates.

Now that we've seen a few common transformations that you're going stumble

across in your assignments and

in the wild, let's look at common actions that you might find in the wild.

And again, I'm going to remind you with big letters that these are eager.

These are how you kick off staged up computations.

So some of the most commonly used operations are collect.

So collect is very popular, because it's a way,

after you've reduced your data structure down, let's say you did a bunch

of filter operations and you've gone from some RDD that was 50 gigabytes, and

now perhaps you've only got a handful of elements in it.

This is how you get the result back.

After you've reduced the size of your dataset, you can use collect to

get back the subset or the smaller data set that you filtered down.

So you get all elements out of the RDD.

But typically you use collect after you've done a few transformations.

You know your data set's going to be smaller,

then you do collect to get all of those results collected on one machine.

The next operator which we've used so far actually is count.

So the idea behind count is very similar to the idea

behind counts in collections API.

Just return the number of elements in the RDD.

Take is another very important action, okay?

It's an action because we go from an RDD to an array.

So you maybe have a very large RDD of 50 gigabytes worth of elements.

And you say okay, take 100.

And you get back then an array of those 100 elements of type T.

So you basically are converting what was once an RDD into an Array.

That's the same with collect, I didn't mention that, the return type of collect

is an Array of T, because you're taking things out of an RDD and you're putting

it into an Array on one machine instead of spread out on many machines in an RDD.

Finally, we have reduce and foreach.

Reduce I think we know pretty well, combine all of the elements

in RDD together using this operation function and then return the result.

So the return type of this is A instead of an RDD, that's how we know it's an action.

And finally, foreach, because foreach returns type unit.

So this applies this function to each element in the RDD.

But since it's not returning an RDD, it's an action.

So, again, how do you determine whether something is an action or

a transformation?

You look at the return type.

If it's not an RDD, it's an action, which means it's eager.

Never forget this.

Let's look at another example.

So let's assume we have an RDD of type String.

So it doesn't matter where it came from.

We have this RDD of type String now, which contains gigabytes and

gigabytes of collected logs over the previous year.

And each element of this RDD represents one line of logging.

Okay, so this is a very realistic scenario which you might want to use Spark for.

Perhaps you have many devices or machines constantly logging to some persistent

storage somewhere like S3, okay.

And now you want to analyze maybe how many errors you have in the last month.

So assuming that we have these dates in a typical

year-month-day-hour-minute-second format, and errors, when there is an error,

the errors are logged with a prefix that includes the word error in it.

How would we determine the number of errors that were logged in the month of

December, 2016?

Easy, we can start by calling filter.

And on each line in the log, we can check to see

if the string 2016-12 exists.

And if so, only pass this predicate if also this is satisfie.

So if error exists also in that line of text.

So if you have both 2016-12 and error in that line of text,

then filter it down into a new RDD with just these elements.

So of course, nothing happens at that point.

This is just a staged-up computation.

And again, by staged, I mean it's a computation that we know we're going to

eventually do, but we haven't started it yet.

You could imagine that we could stage up many more computations as well.

We could do another filter on this.

We could do a map on it.

We could convert everything in the logs to uppercase.

We can do many more transformations that don't actually get executed yet.

We can just stage them up, we can queue them up, okay.

And finally, in this case, we'll call count.

That actually gives the order to Spark to send this function over the network

to all of the little individual machines to do their computations,

and then to add them up and send back the results, the count call.

And to aggregate it, combine it all up, so that you have one integer or

one long with the number of errors in the logs.

I bring up this example because it illustrates why laziness is useful.

So even though this is very simple, it stills illustrates something that is

really useful about doing this in a way where transformations are lazy and

actions are eager.

By staging up computations in this way, we can optimize them.

So Spark can be very smart and decide when it can stop doing work to save time.

So in a very similar example,

it's basically the same example, where we have an RDD of strings, we do a filter,

it contains an error, and then we do an action, in this case take 10.

You could imagine that it's exactly the same code as here.

What's really happening is that the execution of filter is

being deferred until this take function is applied.

And this means, Spark can be very smart.

Spark can stop filtering elements once we've gotten ten of these.

So Spark does not have to first go through everything,

make a new RDD all over the place, all over the whole network with all

of the instances of error, it can just stop when we've gotten to 10.

So this is beneficial because it saves time and work, and

it can do smart things like not computing intermediate RDDs.

This is why it's advantageous that transformations are lazy,

because then we can do all kinds of optimizations on them.

Finally, I just want to run through a few other kinds of transformations that you

might encounter in this course.

There are a special group of transformations that combine two RDDs.

These are typically set-like operations like union or intersection.

So if I have an rdd1 and an rdd2,

I can create the union of those

two RDDs by saying val rdd3 =

rdd1.union(rdd2).

In this case, rdd3 will be a new RDD containing

the elements from both rdd1 and rdd2.

So these are regular set-like operations.

And the same goes for intersection.

So you can do the same thing with two different RDDs, or

you can return a new RDD containing the elements only found in both RDDs.

Subtract is also quite useful, it's the same idea.

But you can subtract the elements out of another RDD.

This is kind of like when you're using a vector editing program and

you can also subtract shapes from one another, right.

And then finally, there's the Cartesian product with other RDDs.

And again, you know that these things are transformations because the return type

is always an RDD.

So it's lazy.

Finally, there are a handful of actions that you're going to come across.

These are actions that don't exist in Scala collections, surely there are some

things that are different between Scala collections and Spark.

These are operations which are useful when dealing with distributed data or

very large data sets.

So just to repeat, we know that these are actions because the return type is not

an RDD, which means that they're eager.

Okay, so operations like takeSample.

If you want to sample down or decrease the size of your data set, takeSample will

give you a random sample of a certain number of elements that you want, but

it returns an array to you.

So you take a very large data set, you sample different pieces of it, and

then collect the subsample data set to one array on one machine rather than

have it be distributed in RDDs.

Likewise, there's another operation called takeOrdered.

So you can, in this case, you can do a take, but have them be ordered.

So you can return the first n elements of an RDD using either their natural order or

a custom comparator.

So that's this implicit ordering parameter here.

And finally, there are two different kinds of save as functions.

So this is of course very important when you're doing distributed jobs.

So these are side effecting operations.

You can call saveAsTextFile on an RDD.

It's an action.

You had queued up many transformations, and then now you say,

okay, save as a text file.

It causes all these of transformations to be computed, and

then this thing can be written to file at this path.

This lets your write to a file either in the local filesystem or in HDFS.

And saveAsSequenceFile is very similar.

So you can save the elements of your dataset as a Hadoop SequenceFile in

the local filesystem or HDFS.

So again, these are actions.

And these are things that are very practically useful and

that you might come across in this course or in Spark codebases elsewhere.

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera propose un accès universel à la meilleure formation au monde, en partenariat avec des universités et des organisations du plus haut niveau, pour proposer des cours en ligne.

© 2018 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play