I have a simple program:
import scalaz._
import stream._
object Play extends App {
val in1 = io.linesR("C:/tmp/as.txt")
val in2 = io.linesR("C:/tmp/bs.txt")
val p = (in1 merge in2) to io.stdOutLines
p.run.run
}
The file as.txt contains five as and the file bs.txt contain 3 bs. I see this sort of output:
a
b
b
a
a
b
a
a
a
However, when I change the declaration of in2 as follows:
val in2 = io.stdInLines
Then I get what I think is unexpected behaviour. According to the documentation 1, the program should pull data non-deterministically from each stream according to whichever stream is quicker to supply stuff. This should mean that I see a bunch of as immediately printed to the console but this is not what happens at all.
Indeed, until I press ENTER, nothing happens. It's quite clear that the behaviour looks a lot like what I would expect if I was choosing a stream at random to get the next element from and then, if that stream was blocking, the merged process blocks too (even if the other stream contains data).
What is going on?
1 - well, OK, there is very little documentation, but Dan Spiewak said very clearly in his talk that it would grab whoever was the first to supply data
The problem is in the implementation of stdInLines. It is blocking, it never Task.forks a thread.
Try changing the implentation of stdInLines to this one:
def stdInLines: Process[Task,String] =
Process.repeatEval(Task.apply {
Option(scala.Console.readLine())
.getOrElse(throw Cause.Terminated(Cause.End))
})
The original io.stdInLines is running the readLine() in the same thread, so it always waits there until you type something.
Related
I am using the broadcast pattern to connect two streams and read data from one to another. The code looks like this
case class Broadcast extends BroadCastProcessFunction[MyObject,(String,Double), MyObject]{
override def processBroadcastElement(in2: (String, Double),
context: BroadcastProcessFunction[MyObject, (String, Double), MyObject]#Context,
collector:Collector[MyObject]):Unit={
context.getBroadcastState(broadcastStateDescriptor).put(in2._1,in2._2)
}
override def processElement(obj: MyObject,
readOnlyContext:BroadCastProcessFunction[MyObject, (String,Double),
MyObject]#ReadOnlyContext, collector: Collector[MyObject]):Unit={
val theValue = readOnlyContext.getBroadccastState(broadcastStateDesriptor).get(obj.prop)
//If I print the context of the state here sometimes it is empty.
out.collect(MyObject(new, properties, go, here))
}
}
The state descriptor:
val broadcastStateDescriptor: MapStateDescriptor[String, Double) = new MapStateDescriptor[String, Double]("name_for_this", classOf[String], classOf[Double])
My execution code looks like this.
val streamA :DataStream[MyObject] = ...
val streamB :DataStream[(String,Double)] = ...
val broadcastedStream = streamB.broadcast(broadcastStateDescriptor)
streamA.connect(streamB).process(new Broadcast)
The problem is in the processElement function the state sometimes is empty and sometimes not. The state should always contain data since I am constantly streaming from a file that I know it has data. I do not understand why it is flushing the state and I cannot get the data.
I tried adding some printing in the processBroadcastElement before and after putting the data to the state and the result is the following
0 - 1
1 - 2
2 - 3
.. all the way to 48 where it resets back to 0
UPDATE:
something that I noticed is when I decrease the value of the timeout of the streaming execution context, the results are a bit better. when I increase it then the map is always empty.
env.setBufferTimeout(1) //better results
env.setBufferTimeout(200) //worse result (default is 100)
Whenever two streams are connected in Flink, you have no control over the timing with which Flink will deliver events from the two streams to your user function. So, for example, if there is an event available to process from streamA, and an event available to process from streamB, either one might be processed next. You cannot expect the broadcastedStream to somehow take precedence over the other stream.
There are various strategies you might employ to cope with this race between the two streams, depending on your requirements. For example, you could use a KeyedBroadcastProcessFunction and use its applyToKeyedState method to iterate over all existing keyed state whenever a new broadcast event arrives.
As David mentioned the job could be restarting. I disabled the checkpoints so I could see any possible exception thrown instead of flink silently failing and restarting the job.
It turned out that there was an error while trying to parse the file. So the job kept restarting thus the state was empty and flink kept consuming the stream over and over again.
So, I was trying to learn about Continuation. I came across with the following saying (link):
Say you're in the kitchen in front of the refrigerator, thinking about a sandwich. You take a continuation right there and stick it in your pocket. Then you get some turkey and bread out of the refrigerator and make yourself a sandwich, which is now sitting on the counter. You invoke the continuation in your pocket, and you find yourself standing in front of the refrigerator again, thinking about a sandwich. But fortunately, there's a sandwich on the counter, and all the materials used to make it are gone. So you eat it. :-) — Luke Palmer
Also, I saw a program in Scala:
var k1 : (Unit => Sandwich) = null
reset {
shift { k : Unit => Sandwich) => k1 = k }
makeSandwich
}
val x = k1()
I don't really know the syntax of Scala (looks similar to Java and C mixed together) but I would like to understand the concept of Continuation.
Firstly, I tried to run this program (by adding it into main). But it fails, I think that it has a syntax error due to the ) near Sandwich but I'm not sure. I removed it but it still does not compile.
How to create a fully compiled example that shows the concept of the story above?
How this example shows the concept of Continuation.
In the link above there was the following saying: "Not a perfect analogy in Scala because makeSandwich is not executed the first time through (unlike in Scheme)". What does it mean?
Since you seem to be more interested in the concept of the "continuation" rather than specific code, let's forget about that code for a moment (especially because it is quite old and I don't really like those examples because IMHO you can't understand them correctly unless you already know what a continuation is).
Note: this is a very long answer with some attempts to describe what a continuations is and why it is useful. There are some examples in Scala-like pseudo-code none of which can actually be compiled and run (there is just one compilable example at the very end and it references another example from the middle of the answer). Expect to spend a significant amount of time just reading this answer.
Intro to continuations
Probably the first thing you should do to understand a continuation is to forget about how modern compilers for most of the imperative languages work and how most of the modern CPUs work and particularly the idea of the call stack. This is actually implementation details (although quite popular and quite useful in practice).
Assume you have a CPU that can execute some sequence of instructions. Now you want to have a high level languages that support the idea of methods that can call each other. The obvious problem you face is that the CPU needs some "forward only" sequence of commands but you want some way to "return" results from a sub-program to the caller. Conceptually it means that you need to have some way to store somewhere before the call all the state of the caller method that is required for it to continue to run after the result of the sub-program is computed, pass it to the sub-program and then ask the sub-program at the end to continue execution from that stored state. This stored state is exactly a continuation. In most of the modern environments those continuations are stored on the call stack and often there are some assembly instructions specifically designed to help handling it (like call and return). But again this is just implementation details. Potentially they might be stored in an arbitrary way and it will still work.
So now let's re-iterate this idea: a continuation is a state of the program at some point that is enough to continue its execution from that point, typically with no additional input or some small known input (like a return value of the called method). Running a continuation is different from a method call in that usually continuation never explicitly returns execution control back to the caller, it can only pass it to another continuation. Potentially you can create such a state yourself, but in practice for the feature to be useful you need some support from the compiler to build continuations automatically or emulate it in some other way (this is why the Scala code you see requires a compiler plugin).
Asynchronous calls
Now there is an obvious question: why continuations are useful at all? Actually there are a few more scenarios besides the simple "return" case. One such scenario is asynchronous programming. Actually if you do some asynchronous call and provide a callback to handle the result, this can be seen as passing a continuation. Unfortunately most of the modern languages do not support automatic continuations so you have to grab all the relevant state yourself. Another problem appears if you have some logic that needs a sequence of many async calls. And if some of the calls are conditional, you easily get to the callbacks hell. The way continuations help you avoid it is by allowing you build a method with effectively inverted control flow. With typical call it is the caller that knows the callee and expects to get a result back in a synchronous way. With continuations you can write a method with several "entry points" (or "return to points") for different stages of the processing logic that you can just pass to some other method and that method can still return to exactly that position.
Consider following example (in pseudo-code that is Scala-like but is actually far from the real Scala in many details):
def someBusinessLogic() = {
val userInput = getIntFromUser()
val firstServiceRes = requestService1(userInput)
val secondServiceRes = if (firstServiceRes % 2 == 0) requestService2v1(userInput) else requestService2v2(userInput)
showToUser(combineUserInputAndResults(userInput,secondServiceRes))
}
If all those calls a synchronous blocking calls, this code is easy. But assume all those get and request calls are asynchronous. How to re-write the code? The moment you put the logic in callbacks you loose the clarity of the sequential code. And here is where continuations might help you:
def someBusinessLogicCont() = {
// the method entry point
val userInput
getIntFromUserAsync(cont1, captureContinuationExpecting(entry1, userInput))
// entry/return point after user input
entry1:
val firstServiceRes
requestService1Async(userInput, captureContinuationExpecting(entry2, firstServiceRes))
// entry/return point after the first request to the service
entry2:
val secondServiceRes
if (firstServiceRes % 2 == 0) {
requestService2v1Async(userInput, captureContinuationExpecting(entry3, secondServiceRes))
// entry/return point after the second request to the service v1
entry3:
} else {
requestService2v2Async(userInput, captureContinuationExpecting(entry4, secondServiceRes))
// entry/return point after the second request to the service v2
entry4:
}
showToUser(combineUserInputAndResults(userInput, secondServiceRes))
}
It is hard to capture the idea in a pseudo-code. What I mean is that all those Async method never return. The only way to continue execution of the someBusinessLogicCont is to call the continuation passed into the "async" method. The captureContinuationExpecting(label, variable) call is supposed to create a continuation of the current method at the label with the input (return) value bound to the variable. With such a re-write you still has a sequential-looking business logic even with all those asynchronous calls. So now for a getIntFromUserAsync the second argument looks like just another asynchronous (i.e. never-returning) method that just requires one integer argument. Let's call this type Continuation[T]
trait Continuation[T] {
def continue(value: T):Nothing
}
Logically Continuation[T] looks like a function T => Unit or rather T => Nothing where Nothing as the return type signifies that the call actually never returns (note, in actual Scala implementation such calls do return, so no Nothing there, but I think conceptually it is easy to think about no-return continuations).
Internal vs external iteration
Another example is a problem of iteration. Iteration can be internal or external. Internal iteration API looks like this:
trait CollectionI[T] {
def forEachInternal(handler: T => Unit): Unit
}
External iteration looks like this:
trait Iterator[T] {
def nextValue(): Option[T]
}
trait CollectionE[T] {
def forEachExternal(): Iterator[T]
}
Note: often Iterator has two method like hasNext and nextValue returning T but it will just make the story a bit more complicated. Here I use a merged nextValue returning Option[T] where the value None means the end of the iteration and Some(value) means the next value.
Assuming the Collection is implemented by something more complicated than an array or a simple list, for example some kind of a tree, there is a conflict here between the implementer of the API and the API user if you use typical imperative language. And the conflict is over the simple question: who controls the stack (i.e. the easy to use state of the program)? The internal iteration is easier for the implementer because he controls the stack and can easily store whatever state is needed to move to the next item but for the API user the things become tricky if she wants to do some aggregation of the stored data because now she has to save the state between the calls to the handler somewhere. Also you need some additional tricks to let the user stop the iteration at some arbitrary place before the end of the data (consider you are trying to implement find via forEach). Conversely the external iteration is easy for the user: she can store all the state necessary to process data in any way in local variables but the API implementer now has to store his state between calls to the nextValue somewhere else. So fundamentally the problem arises because there is only one place to easily store the state of "your" part of the program (the call stack) and two conflicting users for that place. It would be nice if you could just have two different independent places for the state: one for the implementer and another for the user. And continuations provide exactly that. The idea is that we can pass execution control between two methods back and forth using two continuations (one for each part of the program). Let's change the signatures to:
// internal iteration
// continuation of the iterator
type ContIterI[T] = Continuation[(ContCallerI[T], ContCallerLastI)]
// continuation of the caller
type ContCallerI[T] = Continuation[(T, ContIterI[T])]
// last continuation of the caller
type ContCallerLastI = Continuation[Unit]
// external iteration
// continuation of the iterator
type ContIterE[T] = Continuation[ContCallerE[T]]
// continuation of the caller
type ContCallerE[T] = Continuation[(Option[T], ContIterE[T])]
trait Iterator[T] {
def nextValue(cont : ContCallerE[T]): Nothing
}
trait CollectionE[T] {
def forEachExternal(): Iterator[T]
}
trait CollectionI[T] {
def forEachInternal(cont : ContCallerI[T]): Nothing
}
Here ContCallerI[T] type, for example, means that this is a continuation (i.e. a state of the program) the expects two input parameters to continue running: one of type T (the next element) and another of type ContIterI[T] (the continuation to switch back). Now you can see that the new forEachInternal and the new forEachExternal+Iterator have almost the same signatures. The only difference in how the end of the iteration is signaled: in one case it is done by returning None and in other by passing and calling another continuation (ContCallerLastI).
Here is a naive pseudo-code implementation of a sum of elements in an array of Int using these signatures (an array is used instead of something more complicated to simplify the example):
class ArrayCollection[T](val data:T[]) : CollectionI[T] {
def forEachInternal(cont0 : ContCallerI[T], lastCont: ContCallerLastI): Nothing = {
var contCaller = cont0
for(i <- 0 to data.length) {
val contIter = captureContinuationExpecting(label, contCaller)
contCaller.continue(data(i), contIter)
label:
}
}
}
def sum(arr: ArrayCollection[Int]): Int = {
var sum = 0
val elem:Int
val iterCont:ContIterI[Int]
val contAdd0 = captureContinuationExpecting(labelAdd, elem, iterCont)
val contLast = captureContinuation(labelReturn)
arr.forEachInternal(contAdd0, contLast)
labelAdd:
sum += elem
val contAdd = captureContinuationExpecting(labelAdd, elem, iterCont)
iterCont.continue(contAdd)
// note that the code never execute this line, the only way to jump out of labelAdd is to call contLast
labelReturn:
return sum
}
Note how both implementations of the forEachInternal and of the sum methods look fairly sequential.
Multi-tasking
Cooperative multitasking also known as coroutines is actually very similar to the iterations example. Cooperative multitasking is an idea that the program can voluntarily give up ("yield") its execution control either to the global scheduler or to another known coroutine. Actually the last (re-written) example of sum can be seen as two coroutines working together: one doing iteration and another doing summation. But more generally your code might yield its execution to some scheduler that then will select which other coroutine to run next. And what the scheduler does is manages a bunch of continuations deciding which to continue next.
Preemptive multitasking can be seen as a similar thing but the scheduler is run by some hardware interruption and then the scheduler needs a way to create a continuation of the program being executed just before the interruption from the outside of that program rather than from the inside.
Scala examples
What you see is a really old article that is referring to Scala 2.8 (while current versions are 2.11, 2.12, and soon 2.13). As #igorpcholkin correctly pointed out, you need to use a Scala continuations compiler plugin and library. The sbt compiler plugin page has an example how to enable exactly that plugin (for Scala 2.12 and #igorpcholkin's answer has the magic strings for Scala 2.11):
val continuationsVersion = "1.0.3"
autoCompilerPlugins := true
addCompilerPlugin("org.scala-lang.plugins" % "scala-continuations-plugin_2.12.2" % continuationsVersion)
libraryDependencies += "org.scala-lang.plugins" %% "scala-continuations-library" % continuationsVersion
scalacOptions += "-P:continuations:enable"
The problem is that plugin is semi-abandoned and is not widely used in practice. Also the syntax has changed since the Scala 2.8 times so it is hard to get those examples running even if you fix the obvious syntax bugs like missing ( here and there. The reason of that state is stated on the GitHub as:
You may also be interested in https://github.com/scala/async, which covers the most common use case for the continuations plugin.
What that plugin does is emulates continuations using code-rewriting (I suppose it is really hard to implement true continuations over the JVM execution model). And under such re-writings a natural thing to represent a continuation is some function (typically called k and k1 in those examples).
So now if you managed to read the wall of text above, you can probably interpret the sandwich example correctly. AFAIU that example is an example of using continuation as means to emulate "return". If we re-sate it with more details, it could go like this:
You (your brain) are inside some function that at some points decides that it wants a sandwich. Luckily you have a sub-routine that knows how to make a sandwich. You store your current brain state as a continuation into the pocket and call the sub-routine saying to it that when the job is done, it should continue the continuation from the pocket. Then you make a sandwich according to that sub-routine messing up with your previous brain state. At the end of the sub-routine it runs the continuation from the pocket and you return to the state just before the call of the sub-routine, forget all your state during that sub-routine (i.e. how you made the sandwich) but you can see the changes in the outside world i.e. that the sandwich is made now.
To provide at least one compilable example with the current version of the scala-continuations, here is a simplified version of my asynchronous example:
case class RemoteService(private val readData: Array[Int]) {
private var readPos = -1
def asyncRead(callback: Int => Unit): Unit = {
readPos += 1
callback(readData(readPos))
}
}
def readAsyncUsage(rs1: RemoteService, rs2: RemoteService): Unit = {
import scala.util.continuations._
reset {
val read1 = shift(rs1.asyncRead)
val read2 = if (read1 % 2 == 0) shift(rs1.asyncRead) else shift(rs2.asyncRead)
println(s"read1 = $read1, read2 = $read2")
}
}
def readExample(): Unit = {
// this prints 1-42
readAsyncUsage(RemoteService(Array(1, 2)), RemoteService(Array(42)))
// this prints 2-1
readAsyncUsage(RemoteService(Array(2, 1)), RemoteService(Array(42)))
}
Here remote calls are emulated (mocked) with a fixed data provided in arrays. Note how readAsyncUsage looks like a totally sequential code despite the non-trivial logic of which remote service to call in the second read depending on the result of the first read.
For full example you need prepare Scala compiler to use continuations and also use a special compiler plugin and library.
The simplest way is a create a new sbt.project in IntellijIDEA with the following files: build.sbt - in the root of the project, CTest.scala - inside main/src.
Here is contents of both files:
build.sbt:
name := "ContinuationSandwich"
version := "0.1"
scalaVersion := "2.11.6"
autoCompilerPlugins := true
addCompilerPlugin(
"org.scala-lang.plugins" % "scala-continuations-plugin_2.11.6" % "1.0.2")
libraryDependencies +=
"org.scala-lang.plugins" %% "scala-continuations-library" % "1.0.2"
scalacOptions += "-P:continuations:enable"
CTest.scala:
import scala.util.continuations._
object CTest extends App {
case class Sandwich()
def makeSandwich = {
println("Making sandwich")
new Sandwich
}
var k1 : (Unit => Sandwich) = null
reset {
shift { k : (Unit => Sandwich) => k1 = k }
makeSandwich
}
val x = k1()
}
What the code above essentially does is calling makeSandwich function (in a convoluted manner). So execution result would be just printing "Making sandwich" into console. The same result would be achieved without continuations:
object CTest extends App {
case class Sandwich()
def makeSandwich = {
println("Making sandwich")
new Sandwich
}
val x = makeSandwich
}
So what's the point? My understanding is that we want to "prepare a sandwich", ignoring the fact that we may be not ready for that. We mark a point of time where we want to return to after all necessary conditions are met (i.e. we have all necessary ingredients ready). After we fetch all ingredients we can return to the mark and "prepare a sandwich", "forgetting that we were unable to do that in past". Continuations allow us to "mark point of time in past" and return to that point.
Now step by step. k1 is a variable to hold a pointer to a function which should allow to "create sandwich". We know it because k1 is declared so: (Unit => Sandwich).
However initially the variable is not initialized (k1 = null, "there are no ingredients to make a sandwich yet"). So we can't call the function preparing sandwich using that variable yet.
So we mark a point of execution where we want to return to (or point of time in past we want to return to) using "reset" statement.
makeSandwich is another pointer to a function which actually allows to make a sandwich. It's the last statement of "reset block" and hence it is passed to "shift" (function) as argument (shift { k : (Unit => Sandwich).... Inside shift we assign that argument to k1 variable k1 = k thus making k1 ready to be called as a function. After that we return to execution point marked by reset. The next statement is execution of function pointed to by k1 variable which is now properly initialized so finally we call makeSandwich which prints "Making sandwich" to a console. It also returns an instance of sandwich class which is assigned to x variable.
Not sure, probably it means that makeSandwich is not called inside reset block but just afterwards when we call it as k1().
Let be a job that contains two phases that (for convenience) cannot be merged. Let's call A the first step and B the second step. Hence, we always need to do A then B.
Workflow
start new cluster with a job A
build MyOutput
count(MyOutput) = 2000 (*)
write(MyOutput)
start new cluster with a job B
read(MyOutput)
count(MyOutput) = 1788 (**).
Precisions
A provides an output which is an RDD[MyObject], namely MyOutput. To write MyOutput, here is what I do: MyOutput.saveAsObjectFile("...")
Then B uses MyOutput as an input, reading the previously written file. Here is what I do: val MyOutput: RDD[MyObject] = sc.objectFile("...")
A and B happen on two separate clusters.
Problem
The main point is that the problem does not ALWAYS happen but when this problem appears, (*) < (**) -- it seems we lost some data whereas we should not have any differences between these data. Obvisously something is wrong.
Do you know what happens here? How to solve that?
A tricky problem happened related to the sharp increase of executing time.
I run my scala code in local spark, part of which is to build a n*n matrix.
When running a small dataset, it just takes 5s to finish. The most time-consuming part is to build 2000*2000 matrix. And this part is executed within map, which just deals with array data structure.
However, just out of curiosity, I add "println" within the matrix-building code to see the number of iterations. Suddenly, the whole running time increases to 1min23s.
And the final results are same.
I am new to Spark and have no idea what really causes this situation.
The codes are simply:
val x = someRDD.map(buildMatrix)
def buildMatrix(stringVect:Array[String]): Array[Array[Double]] = {
//var count = 0
val num = stringVect.length
var simi_matrix = Array[Array[Double]]()
for (i<- 0 until num-1){
for (j<- (i+1) until num){
"build the matrix with some computation"
//println(count)
//count += 1
}
}
}
TL;DR
This does not have to do anything with Spark. I/O access to the console is synchronized and costly. It will slow down any program on the JVM (Scala/Java/Clojure/...).
println defaults to java.lang.System.out which is a PrintStream. println delegates to PrintStream#println, hence entering the synchronized block of the println implementation to output to the console: There are two expenses:
Getting a synchronized lock
I/O to the console OutputStream
The slowdown observed is to be expected. Just don't use println in hot parts of the code (like a tight loop in this case).
I'm failing to get a simple scalaz-stream example running, reading from TCP and writing to std out.
val src = tcp.reads(1024)
val addr = new InetSocketAddress(12345)
val p = tcp.server(addr, concurrentRequests = 1) {
src ++ tcp.lift(io.stdOutLines)
}
p.run.run
It just sits there, not printing anything.
I've also tried various arrangements using to, always with the tcp.lift incantation to get a Process[Connection, A], including
tcp.server(addr, concurrentRequests = 1)(src) map (_ to tcp.lift(io.stdOutLines))
which doesn't even compile.
Do I need to wye the source and print streams together? An example I found on the original pull request for tcp replacing nio seemed to indicate this, but wye no longer appears to exist on Process, so confusion reigns unfortunately.
Edit it turns out that in addition to the type problems explained by Paul, you also need to run the inner processes "manually", for example by doing p.map(_.run.run).run.run. I don't suppose that's the idiomatic way to do this, but it does work.
You need to pass src through the sink to actually write anything. I think this should do it:
import scalaz.stream.{io,tcp,text}
import scalaz.stream.tcp.syntax._
val p = tcp.server(addr, concurrentRequests = 1) {
tcp.reads(1024).pipe(text.utf8Decode) through tcp.lift(io.stdOutLines)
}
p.run.run
The expression src ++ tcp.lift(io.stdOutLines) should really be a type error. The type of tcp.reads(1024) is Process[Connection,ByteVector], and the type of tcp.lift(io.stdOutLines) is Process[Connection, String => Task[Unit]]. Appending those two processes does not make sense, and the only reason it typechecks is due to the covariance of Process[+F[_],+O]. Scala is "helpfully" inferring Any when you append two processes with unrelated output types.
A future release of scalaz-stream may add a constraint on ++ and other functions that exploit covariance to make sure the least upper bound that gets computed isn't something useless like Any or Serializable. This would go a long way to preventing mistakes like this. In the meantime, make sure you understand the types of all the functions you are working with, what they do, and how you are sticking them together.