Flink Streaming is a system for high-throughput, low-latency data stream processing. Flink Streaming natively supports stateful computation, data-driven windowing semantics and iterative stream processing. The system can connect to and process data streams from different data sources like file sources, web sockets, message queues (Apache Kafka, RabbitMQ, Twitter Streaming API …), and also from any user defined data sources. Data streams can be transformed and modified to create new data streams using high-level functions similar to the ones provided by the batch processing API.
The Streaming API is currently part of the flink-staging Maven project. All relevant classes are located in the org.apache.flink.streaming package.
Add the following dependency to your pom.xml to use the Flink Streaming.
In order to create your own Flink Streaming program, we encourage you to start with the skeleton and gradually add your own transformations. The remaining sections act as references for additional transformations and advanced features.
The following program is a complete, working example of streaming WordCount, that incrementally counts the words coming from a web socket. You can copy & paste the code to run it locally.
{% highlight java %} public class StreamingWordCount {
public static void main(String[] args) {
StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
DataStream<Tuple2<String, Integer>> dataStream = env
.socketTextStream("localhost", 9999)
.flatMap(new Splitter())
.groupBy(0)
.sum(1);
dataStream.print();
env.execute("Socket Stream WordCount");
}
public static class Splitter implements FlatMapFunction<String, Tuple2<String, Integer>> {
@Override
public void flatMap(String sentence, Collector<Tuple2<String, Integer>> out) throws Exception {
for (String word: sentence.split(" ")) {
out.collect(new Tuple2<String, Integer>(word, 1));
}
}
}
} {% endhighlight %}
object WordCount { def main(args: Array[String]) {
val env = StreamExecutionEnvironment.getExecutionEnvironment
val text = env.socketTextStream("localhost", 9999)
val counts = text.flatMap { _.toLowerCase.split("\\W+") filter { _.nonEmpty } }
.map { (_, 1) }
.groupBy(0)
.sum(1)
counts.print
env.execute("Scala Socket Stream WordCount")
} } {% endhighlight %}
To run the example program, start the input stream with netcat first from a terminal:
nc -lk 9999
The lines typed to this terminal will be the source data stream for your streaming job.
As presented in the example, a Flink Streaming program looks almost identical to a regular Flink program. Each stream processing program consists of the following parts:
StreamExecutionEnvironment,As these steps are basically the same as in the batch API, we will only note the important differences. For stream processing jobs, the user needs to obtain a StreamExecutionEnvironment in contrast with the batch API where one would need an ExecutionEnvironment. Otherwise, the process is essentially the same:
{% highlight java %} StreamExecutionEnvironment.getExecutionEnvironment(); StreamExecutionEnvironment.createLocalEnvironment(parallelism); StreamExecutionEnvironment.createRemoteEnvironment(String host, int port, int parallelism, String... jarFiles); {% endhighlight %}
For connecting to data streams the StreamExecutionEnvironment has many different methods, from basic file sources to completely general user defined data sources. We will go into details in the basics section.
For example:
{% highlight java %} env.socketTextStream(host, port); env.fromElements(elements…); env.addSource(sourceFunction) {% endhighlight %}
After defining the data stream sources the user can specify transformations on the data streams to create a new data stream. Different data streams can be also combined together for joint transformations which are being showcased in the transformations section.
For example:
{% highlight java %} dataStream.map(mapFunction).reduce(reduceFunction); {% endhighlight %}
The processed data can be pushed to different outputs called sinks. The user can define their own sinks or use any predefined filesystem, message queue or database sink.
For example:
{% highlight java %} dataStream.writeAsCsv(path); dataStream.print(); dataStream.addSink(sinkFunction) {% endhighlight %}
Once the complete program is specified execute(programName) is to be called on the StreamExecutionEnvironment. This will either execute on the local machine or submit the program for execution on a cluster, depending on the chosen execution environment.
{% highlight java %} env.execute(programName); {% endhighlight %}
As presented in the example a Flink Streaming program looks almost identical to a regular Flink program. Each stream processing program consists of the following parts:
StreamExecutionEnvironment,As these steps are basically the same as in the batch API we will only note the important differences. For stream processing jobs, the user needs to obtain a StreamExecutionEnvironment in contrast with the batch API where one would need an ExecutionEnvironment. The process otherwise is essentially the same:
{% highlight scala %} StreamExecutionEnvironment.getExecutionEnvironment StreamExecutionEnvironment.createLocalEnvironment(parallelism) StreamExecutionEnvironment.createRemoteEnvironment(host: String, port: String, parallelism: Int, jarFiles: String*) {% endhighlight %}
For connecting to data streams the StreamExecutionEnvironment has many different methods, from basic file sources to completely general user defined data sources. We will go into details in the basics section.
For example:
{% highlight scala %} env.socketTextStream(host, port) env.fromElements(elements…) env.addSource(sourceFunction) {% endhighlight %}
After defining the data stream sources the user can specify transformations on the data streams to create a new data stream. Different data streams can be also combined together for joint transformations which are being showcased in the transformations section.
For example:
{% highlight scala %} dataStream.map(mapFunction).reduce(reduceFunction) {% endhighlight %}
The processed data can be pushed to different outputs called sinks. The user can define their own sinks or use any predefined filesystem, message queue or database sink.
For example:
{% highlight scala %} dataStream.writeAsCsv(path) dataStream.print dataStream.addSink(sinkFunction) {% endhighlight %}
Once the complete program is specified execute(programName) is to be called on the StreamExecutionEnvironment. This will either execute on the local machine or submit the program for execution on a cluster, depending on the chosen execution environment.
{% highlight scala %} env.execute(programName) {% endhighlight %}
The DataStream is the basic data abstraction provided by Flink Streaming. It represents a continuous, parallel, immutable stream of data of a certain type. By applying transformations the user can create new data streams or output the results of the computations. For instance the map transformation creates a new DataStream by applying a user defined function on each element of a given DataStream
The transformations may return different data stream types allowing more elaborate transformations, for example the groupBy(…) method returns a GroupedDataStream which can be used for grouped transformations such as aggregating by key. We will discover more elaborate data stream types in the upcoming sections.
Apache Flink is trying to reduce the number of object allocations for better performance.
By default, user defined functions (like map() or reduce()) are getting new objects on each call (or through an iterator). So it is possible to keep references to the objects inside the function (for example in a List).
There is a switch at the ExectionConfig which allows users to enable the object reuse mode:
env.getExecutionConfig().enableObjectReuse()
For mutable types, Flink will reuse object instances. In practice that means that a map() function will always receive the same object instance (with its fields set to new values). The object reuse mode will lead to better performance because fewer objects are created, but the user has to manually take care of what they are doing with the object references.
Partitioning controls how individual data points of a stream are distributed among the parallel instances of the transformation operators. This also controls the ordering of the records in the DataStream. There is partial ordering guarantee for the outputs with respect to the partitioning scheme (outputs produced from each partition are guaranteed to arrive in the order they were produced).
There are several partitioning types supported in Flink Streaming:
dataStream.forward()dataStream.shuffle()dataStream.rebalance()dataStream.partitionByHash(fields…)dataStream.groupBy(fields…)dataStream.broadcast()dataStream.global()By default Forward partitioning is used.
Partitioning does not remain in effect after a transformation, so it needs to be set again for subsequent operations.
The user is expected to connect to the outside world through the source and the sink interfaces.
Sources can by created by using StreamExecutionEnvironment.addSource(sourceFunction). Either use one of the source functions that come with Flink or write a custom source by implementing the SourceFunction interface. By default, sources run with parallelism of 1. To create parallel sources the user's source function needs to implement ParallelSourceFunction or extend RichParallelSourceFunction in which cases the source will have the parallelism of the environment. The parallelism for ParallelSourceFunctions can be changed after creation by using source.setParallelism(parallelism).
The SourceFunction interface has two methods: run(SourceContext) and cancel(). The run() method is not expected to return until the source has either finished by itself or received a cancel request. The source can communicate with the outside world using the source context. For example, the emit(element) method is used to emit one element from the source. Most sources will have an infinite while loop inside the run() method to read from the input and emit elements. Upon invocation of the cancel() method the source is required to break out of its internal loop and return from the run() method. A common implementation for this is the following:
{% highlight java %} public static class MySource implements SourceFunction {
// utility for job cancellation
private volatile boolean isRunning = false;
@Override
public void run(SourceContext<Long> ctx) throws Exception {
isRunning = true;
while (isRunning) {
// the source runs, isRunning flag should be checked frequently
}
}
}
// invoked by the framework in case of job cancellation
@Override
public void cancel() {
isRunning = false;
}
} {% endhighlight %}
In addition to the bounded data sources (with similar method signatures as the batch API) there are several predefined stream sources accessible from the StreamExecutionEnvironment:
Socket text stream: Creates a new DataStream that contains the strings received from the given socket. Strings are decoded by the system's default character set. The user can optionally set the delimiters or the number of connection retries in case of errors. Usage: env.socketTextStream(hostname, port,…)
Text file stream: Creates a new DataStream that contains the lines of the files created (or modified) in a given directory. The system continuously monitors the given path, and processes any new files or modifications based on the settings. The file will be read with the system's default character set. Usage: env.readFileStream(String path, long checkFrequencyMillis, WatchType watchType)
Message queue connectors: There are pre-implemented connectors for a number of popular message queue services, please refer to the section on connectors for more details.
Custom source: Creates a new DataStream by using a user defined SourceFunction implementation. Usage: env.addSource(sourceFunction)
DataStreamSink represents the different outputs of Flink Streaming programs. The user can either define his own SinkFunction implementation or chose one of the available implementations (methods of DataStream).
For example:
dataStream.print() – Writes the DataStream to the standard output, practical for testing purposesdataStream.writeAsText(parameters) – Writes the DataStream to a text filedataStream.writeAsCsv(parameters) – Writes the DataStream to CSV formatdataStream.addSink(sinkFunction) – Custom sink implementationThere are pre-implemented connectors for a number of the most popular message queue services, please refer to the section on connectors for more detail.
Transformations, also called operators, represent the users' business logic on the data stream. Operators consume data streams and produce new data streams. The user can chain and combine multiple operators on the data stream to produce the desired processing steps. Most of the operators work very similar to the batch Flink API allowing developers to reason about DataStream the same way as they would about DataSet. At the same time there are operators that exploit the streaming nature of the data to allow advanced functionality.
Basic transformations can be seen as functions that operate on records of the data stream.
<tr>
<td><strong>FlatMap</strong></td>
<td>
<p>Takes one element and produces zero, one, or more elements. A flatmap that splits sentences to words:</p>
{% highlight java %} dataStream.flatMap(new FlatMapFunction<String, String>() { @Override public void flatMap(String value, Collector out) throws Exception { for(String word: value.split(" ")){ out.collect(word); } } }); {% endhighlight %}
<tr>
<td><strong>Filter</strong></td>
<td>
<p>Evaluates a boolean function for each element and retains those for which the function returns true.
<br/>
<br/>
A filter that filters out zero values:
</p>
{% highlight java %} dataStream.filter(new FilterFunction() { @Override public boolean filter(Integer value) throws Exception { return value != 0; } }); {% endhighlight %}
<tr>
<td><strong>Reduce</strong></td>
<td>
<p>Combines a stream of elements into another stream by repeatedly combining two elements
into one and emits the current state after every reduction. Reduce may be applied on a full, windowed or grouped data stream.
<br/>
<strong>IMPORTANT:</strong> The streaming and the batch reduce functions have different semantics. A streaming reduce on a data stream emits the current reduced value for every new element on a data stream. On a windowed data stream it works as a batch reduce: it produces at most one value per window.
<br/>
<br/>
A reducer that sums up the incoming stream, the result is a stream of intermediate sums:</p>
{% highlight java %} dataStream.reduce(new ReduceFunction() { @Override public Integer reduce(Integer value1, Integer value2) throws Exception { return value1 + value2; } }); {% endhighlight %}
<tr>
<td><strong>Fold</strong></td>
<td>
<p>Combines a stream element by element with an initial aggregator value. Fold may be applied on a full, windowed or grouped data stream.
<br/>
A folder that appends strings one by one to the empty sting:</p>
{% highlight java %} dataStream.fold("", new FoldFunction<String, String>() { @Override public String fold(String accumulator, String value) throws Exception { return accumulator + value; } }); {% endhighlight %}
<tr>
<td><strong>Union</strong></td>
<td>
<p>Union of two or more data streams creating a new stream containing all the elements from all the streams.</p>
{% highlight java %} dataStream.union(otherStream1, otherStream2, …) {% endhighlight %}
The following transformations are available on data streams of Tuples:
<tr>
<td><strong>Map</strong></td>
<td>
<p>Takes one element and produces one element. A map that doubles the values of the input stream:</p>
{% highlight scala %} dataStream.map{ x => x * 2 } {% endhighlight %}
<tr>
<td><strong>FlatMap</strong></td>
<td>
<p>Takes one element and produces zero, one, or more elements. A flatmap that splits sentences to words:</p>
{% highlight scala %} data.flatMap { str => str.split(" ") } {% endhighlight %}
<tr>
<td><strong>Filter</strong></td>
<td>
<p>Evaluates a boolean function for each element and retains those for which the function returns true.
<br/>
<br/>
A filter that filters out zero values:
</p>
{% highlight scala %} dataStream.filter{ _ != 0 } {% endhighlight %}
<tr>
<td><strong>Reduce</strong></td>
<td>
<p>Combines a stream of elements into another stream by repeatedly combining two elements
into one and emits the current state after every reduction. Reduce may be applied on a full, windowed or grouped data stream.
<br/>
<strong>IMPORTANT:</strong> The streaming and the batch reduce functions have different semantics. A streaming reduce on a data stream emits the current reduced value for every new element on a data stream. On a windowed data stream it works as a batch reduce: it produces at most one value per window.
<br/>
<br/>
A reducer that sums up the incoming stream, the result is a stream of intermediate sums:</p>
{% highlight scala %} dataStream.reduce{ _ + _} {% endhighlight %}
<tr>
<td><strong>Fold</strong></td>
<td>
<p>Combines a stream element by element with an initial aggregator value. Fold may be applied on a full, windowed or grouped data stream.
<br/>
A folder that appends strings one by one to the empty sting:</p>
{% highlight scala %} dataStream.fold{"", _ + _ } {% endhighlight %}
<tr>
<td><strong>Union</strong></td>
<td>
<p>Union of two or more data streams creating a new stream containing all the elements from all the streams.</p>
{% highlight scala %} dataStream.union(otherStream1, otherStream2, …) {% endhighlight %}
Some transformations require that the elements of a DataStream are grouped on some key. The user can create a GroupedDataStream by calling the groupBy(key) method of a non-grouped DataStream. Keys can be of three types: field positions (applicable for tuple/array types), field expressions (applicable for pojo types), KeySelector instances.
Aggregation or reduce operators called on GroupedDataStreams produce elements on a per group basis.
The Flink Streaming API supports different types of pre-defined aggregations of DataStreams. A common property of these operators, is that they produce the stream of intermediate aggregate values (just like reduce on streams).
Types of aggregations: sum(field), min(field), max(field), minBy(field, first), maxBy(field, first).
With sum, min, and max for every incoming tuple the selected field is replaced with the current aggregated value. Fields can be selected using either field positions or field expressions (similarly to grouping).
With minBy and maxBy the output of the operator is the element with the current minimal or maximal value at the given field. If more components share the minimum or maximum value, the user can decide if the operator should return the first or last element. This can be set by the first boolean parameter.
Flink streaming provides very flexible data-driven windowing semantics to create arbitrary windows (also referred to as discretizations or slices) of the data streams and apply reduce, map or aggregation transformations on the windows acquired. Windowing can be used for instance to create rolling aggregations of the most recent N elements, where N could be defined by Time, Count or any arbitrary user defined measure.
The user can control the size (eviction) of the windows and the frequency of transformation or aggregation calls (trigger) on them in an intuitive API. We will describe the exact semantics of these operators in the policy based windowing section.
Some examples:
dataStream.window(eviction).every(trigger).reduceWindow(…)dataStream.window(…).every(…).mapWindow(…).flatten()dataStream.window(…).every(…).groupBy(…).aggregate(…).getDiscretizedStream()The core abstraction of the Windowing semantics is the WindowedDataStream and the StreamWindow. The WindowedDataStream is created when we first call the window(…) method of the DataStream and represents the windowed discretisation of the underlying stream. The user can think about it simply as a DataStream<StreamWindow<T>> where additional API functions are supplied to provide efficient transformations of individual windows.
Please note at this point that the .every(…) call belongs together with the preceding .window(…) call and does not define a new transformation in itself.
The result of a window transformation is again a WindowedDataStream which can also be used to further apply other windowed computations. In this sense, window transformations define mapping from stream windows to stream windows.
The user has different ways of using the result of a window operation:
windowedDataStream.flatten() - streams the results element wise and returns a DataStream<T> where T is the type of the underlying windowed streamwindowedDataStream.getDiscretizedStream() - returns a DataStream<StreamWindow<T>> for applying some advanced logic on the stream windows itself. Be careful here, as at this point, we need to materialise the full windowsThe next example would create windows that hold elements of the last 5 seconds, and the user defined transformation would be executed on the windows every second (sliding the window by 1 second):
This approach is often referred to as policy based windowing. Different policies (count, time, etc.) can be mixed as well, for example to downsample our stream, a window that takes the latest 100 elements of the stream every minute is created as follows:
The user can also omit the every(…) call which results in a tumbling window emptying the window after every transformation call.
Several predefined policies are provided in the API, including delta-based, count-based and time-based policies. These can be accessed through the static methods provided by the PolicyHelper classes:
Time.of(…)Count.of(…)Delta.of(…)FullStream.window()For detailed description of these policies please refer to the Javadocs.
The policy based windowing is a highly flexible way to specify stream discretisation also called windowing semantics. Two types of policies are used for such a specification:
TriggerPolicy defines when to trigger the reduce or transformation UDF on the current window and emit the result. In the API it completes a window statement such as: window(…).every(…), while the triggering policy is passed within every. In case the user wants to use tumbling eviction policy (the window is emptied after the transformation) he can omit the .every(…) call and pass the trigger policy directly to the .window(…) call.
EvictionPolicy defines the length of a window as a means of a predicate for evicting tuples when they are no longer needed. In the API this can be defined by the window(…) operation on a stream. There are mostly the same predefined policy types provided as for trigger policies.
Trigger and eviction policies work totally independently of each other. The eviction policy continuously maintains a window, into which it adds new elements and based on the eviction logic removes older elements in the order of arrival. The trigger policy on the other hand only decided at each new incoming element, whether it should trigger computation (and output results) on the currently maintained window.
Several predefined policies are provided in the API, including delta-based, punctuation based, count-based and time-based policies. Policies are in general UDFs and can implement any custom behaviour.
In addition to the dataStream.window(…).every(…) style, users can specifically pass the trigger and eviction policies during the window call:
By default triggers can only trigger when a new element arrives. This might not be suitable for all the use-cases with low data rates. To also provide triggering between elements, so called active policies can be used (the two interfaces controlling this special behaviour is ActiveTriggerPolicy and CentralActiveTrigger). The predefined time-based policies are already implemented in such a way and can hold as an example for user defined active policy implementations.
Time-based trigger and eviction policies can work with user defined TimeStamp implementations, these policies already cover most use cases.
The WindowedDataStream<T>.reduceWindow(ReduceFunction<T>) transformation calls the user-defined ReduceFunction at every trigger on the records currently in the window. The user can also use the different pre-implemented streaming aggregations such as sum, min, max, minBy and maxBy.
The following is an example for a window reduce that sums the elements in the last minute with 10 seconds slide interval:
The WindowedDataStream<T>.mapWindow(WindowMapFunction<T,O>) transformation calls mapWindow(…) for each StreamWindow in the discretised stream, providing access to all elements in the window through the iterable interface. At each function call the output StreamWindow<O> will consist of all the elements collected to the collector. This allows a straightforward way of mapping one stream window to another.
Calling the groupBy(…) method on a windowed stream groups the elements by the given fields inside the stream windows. The window sizes (evictions) and slide sizes (triggers) will be calculated on the whole stream (in a global fashion), but the user defined functions will be applied on a per group basis inside the window. This means that for a call windowedStream.groupBy(…).reduceWindow(…) will transform each window into another window consisting of as many elements as keys in the original window, with the reduced values per key. Similarly the mapWindow transformation is applied per group as well.
The user can also create discretisation on a per group basis calling window(…).every(…) on an already grouped data stream. This will apply the discretisation logic independently for each key.
To highlight the differences let us look at two examples.
To get the maximal value for each key on the last 100 elements (global) we use the first approach:
Using this approach we took the last 100 elements, divided it into groups by key, and then applied the aggregation. To create fixed size windows for every key, we need to bring the groupBy call before the window call. So to take the max for the last 100 elements in each group:
This will create separate windows for different keys and apply the trigger and eviction policies on a per group basis.
Using the WindowedDataStream abstraction we can apply several transformations one after another on the discretised streams without having to re-discretise it:
The above call would create global windows of 1000 elements, group them by the first key, and then apply a mapWindow transformation. The resulting windowed stream will then be grouped by the second key and further reduced. The results of the reduce transformation are then flattened.
Notice that here we only defined the window size once at the beginning of the transformation. This means that anything that happens afterwards (groupBy(firstKey).mapWindow(…).groupBy(secondKey).reduceWindow(…)) happens inside the 1000 element windows. Of course the mapWindow might reduce the number of elements, but the key idea is that each transformation still corresponds to the same 1000 elements in the original stream.
Sometimes it is necessary to aggregate over all the previously seen data in the stream. For this purpose either use the dataStream.window(FullStream.window()).every(trigger) or equivalently dataStream.every(trigger).
By default all window discretisation calls (dataStream.window(…)) define global windows meaning that a global window of count 100 will contain the last 100 elements arrived at the discretisation operator in order. In most cases (except for Time) this means that the operator doing the actual discretisation needs to have a parallelism of 1 to be able to correctly execute the discretisation logic.
Sometimes it is sufficient to create local discretisations, which allows the discretiser to run in parallel and apply the given discretisation logic at every discretiser instance. To allow local discretisation use the local() method of the windowed data stream.
For example, dataStream.window(Count.of(100)).maxBy(field) would create global windows of 100 elements (Count discretises with parallelism of 1) and return the record with the max value by the selected field; alternatively the dataStream.window(Count.of(100)).local().maxBy(field) would create several count discretisers (as defined by the environment parallelism) and compute the max values accordingly.
While database style operators like joins (on key) and crosses are hard to define properly on data streams, a straightforward interpretation is to apply these operators on windows of the data streams.
Currently join and cross operators are supported only on time windows. We are working on alleviating this limitation in the next release.
Temporal operators take the current windows of both streams and apply the join/cross logic on these window pairs.
The Join transformation produces a new Tuple DataStream with two fields. Each tuple holds a joined element of the first input DataStream in the first tuple field and a matching element of the second input DataStream in the second field for the current window. The user can also supply a custom join function to control the produced elements.
The following code shows a default Join transformation using field position keys:
The Cross transformation combines two DataStreams into one DataStream. It builds all pairwise combinations of the elements of both input DataStreams in the current window, i.e., it builds a temporal Cartesian product. The user can also supply a custom cross function to control the produced elements
Co operators allow the users to jointly transform two DataStreams of different types, providing a simple way to jointly manipulate streams with a shared state. It is designed to support joint stream transformations where union is not appropriate due to different data types, or in case the user needs explicit tracking of the origin of individual elements. Co operators can be applied to ConnectedDataStreams which represent two DataStreams of possibly different types. A ConnectedDataStream can be created by calling the connect(otherDataStream) method of a DataStream.
Applies a CoMap transformation on two separate DataStreams, mapping them to a common output type. The transformation calls a CoMapFunction.map1() for each element of the first input and CoMapFunction.map2() for each element of the second input. Each CoMapFunction call returns exactly one element. A CoMap operator that outputs true if an Integer value is received and false if a String value is received:
dataStream1.connect(dataStream2) .map(new CoMapFunction<Integer, String, Boolean>() {
@Override
public Boolean map1(Integer value) {
return true;
}
@Override
public Boolean map2(String value) {
return false;
}
})
{% endhighlight %}
(dataStream1 connect dataStream2) .map( (_ : Int) => true, (_ : String) => false ) {% endhighlight %}
The FlatMap operator for the ConnectedDataStream works similarly to CoMap, but instead of returning exactly one element after each map call the user can output arbitrarily many values using the Collector interface.
dataStream1.connect(dataStream2) .flatMap(new CoFlatMapFunction<Integer, String, String>() {
@Override
public void flatMap1(Integer value, Collector<String> out) {
out.collect(value.toString());
}
@Override
public void flatMap2(String value, Collector<String> out) {
for (String word: value.split(" ")) {
out.collect(word);
}
}
})
{% endhighlight %}
(dataStream1 connect dataStream2) .flatMap( (num : Int) => List(num.toString), (str : String) => str.split(" ") ) {% endhighlight %}
The windowReduce operator applies a user defined CoWindowFunction to time aligned windows of the two data streams and return zero or more elements of an arbitrary type. The user can define the window and slide intervals and can also implement custom timestamps to be used for calculating windows.
The Reduce operator for the ConnectedDataStream applies a simple reduce transformation on the joined data streams and then maps the reduced elements to a common output type.
Most data stream operators support directed outputs (output splitting), meaning that different output elements are sent only to specific outputs. The outputs are referenced by their name given at the point of receiving:
{% highlight java %} SplitDataStream split = someDataStream.split(outputSelector); DataStream even = split.select(“even”); DataStream odd = split.select(“odd”); {% endhighlight %} In the above example the data stream named “even” will only contain elements that are directed to the output named “even”. The user can of course further transform these new streams by for example squaring only the even elements.
Data streams only receive the elements directed to selected output names. The user can also select multiple output names by splitStream.select(“output1”, “output2”, …). It is common that a stream listens to all the outputs, so split.selectAll() provides this functionality without having to select all names.
The outputs of an operator are directed by implementing a selector function (implementing the OutputSelector interface):
{% highlight java %} Iterable select(OUT value); {% endhighlight %}
The data is sent to all the outputs returned in the iterable (referenced by their name). This way the direction of the outputs can be determined by the value of the data sent.
For example to split even and odd numbers:
{% highlight java %} @Override Iterable select(Integer value) {
List<String> outputs = new ArrayList<String>();
if (value % 2 == 0) {
outputs.add("even");
} else {
outputs.add("odd");
}
return outputs;
} {% endhighlight %}
Every output will be emitted to the selected outputs exactly once, even if you add the same output names more than once.
The functionality provided by output splitting can also be achieved efficiently (due to operator chaining) by multiple filter operators.
Most data stream operators support directed outputs (output splitting), meaning that different output elements are sent only to specific outputs. The outputs are referenced by their name given at the point of receiving:
{% highlight scala %} val split = someDataStream.split( (num: Int) => (num % 2) match { case 0 => List(“even”) case 1 => List(“odd”) } )
val even = split select “even” val odd = split select “odd” {% endhighlight %}
In the above example the data stream named “even” will only contain elements that are directed to the output named “even”. The user can of course further transform these new stream by for example squaring only the even elements.
Data streams only receive the elements directed to selected output names. The user can also select multiple output names by splitStream.select(“output1”, “output2”, …). It is common that a stream listens to all the outputs, so split.selectAll provides this functionality without having to select all names.
The outputs of an operator are directed by implementing a function that returns the output names for the value. The data is sent to all the outputs returned by the function (referenced by their name). This way the direction of the outputs can be determined by the value of the data sent.
Every output will be emitted to the selected outputs exactly once, even if you add the same output names more than once.
The functionality provided by output splitting can also be achieved efficiently (due to operator chaining) by multiple filter operators.
{% highlight java %} IterativeDataStream iteration = source.iterate(maxWaitTimeMillis); {% endhighlight %}
The operator applied on the iteration starting point is the head of the iteration, where data is fed back from the iteration tail.
{% highlight java %} DataStream head = iteration.map(new IterationHead()); {% endhighlight %}
To close an iteration and define the iteration tail, the user calls closeWith(iterationTail) method of the IterativeDataStream. This iteration tail (the DataStream given to the closeWith function) will be fed back to the iteration head. A common pattern is to use filters to separate the output of the iteration from the feedback-stream.
{% highlight java %} DataStream tail = head.map(new IterationTail());
iteration.closeWith(tail.filter(isFeedback));
DataStream output = tail.filter(isOutput);
output.map(…).project(…); {% endhighlight %}
In this case, all values passing the isFeedback filter will be fed back to the iteration head, and the values passing the isOutput filter will produce the output of the iteration that can be transformed further (here with a map and a projection) outside the iteration.
Because iterative streaming programs do not have a set number of iterations for each data element, the streaming program has no information on the end of its input. As a consequence iterative streaming programs run until the user manually stops the program. While this is acceptable under normal circumstances, a method is provided to allow iterative programs to shut down automatically if no input is received by the iteration head for a predefined number of milliseconds. To use this functionality the user needs to add the maxWaitTimeMillis parameter to the dataStream.iterate(…) call to control the max wait time.
By default the partitioning of the feedback stream will be automatically set to be the same as the input of the iteration head. To override this the user can set an optional boolean flag in the closeWith method.
A common pattern is to use filters to separate the output from the feedback-stream. In this case all values passing the isFeedback filter will be fed back to the iteration head, and the values passing the isOutput filter will produce the output of the iteration that can be transformed further (here with a map and a projection) outside the iteration.
{% highlight scala %} val iteratedStream = someDataStream.iterate(maxWaitTime) { iteration => { val head = iteration.map(iterationHead) val tail = head.map(iterationTail) (tail.filter(isFeedback), tail.filter(isOutput)) } }.map(…).project(…) {% endhighlight %}
Because iterative streaming programs do not have a set number of iterations for each data element, the streaming program has no information on the end of its input. As a consequence iterative streaming programs run until the user manually stops the program. While this is acceptable under normal circumstances a method is provided to allow iterative programs to shut down automatically if no input received by the iteration head for a predefined number of milliseconds. To use this functionality the user needs to add the maxWaitTimeMillis parameter to the dataStream.iterate(…) call to control the max wait time.
By default the partitioning of the feedback stream will be automatically set to be the same as the input of the iteration head. To override this the user can set an optional boolean flag in the iterate method.
The usage of rich functions are essentially the same as in the batch Flink API. All transformations that take as argument a user-defined function can instead take a rich function as argument:
@Override
public String map(Integer value) { return value.toString(); }
}); {% endhighlight %}
Rich functions provide, in addition to the user-defined function (map(), reduce(), etc), the open() and close() methods for initialization and finalization.
Flink supports the checkpointing and persistence of user defined state, so in case of a failure this state can be restored to the latest checkpoint and the processing can continue from there. This gives exactly once semantics for anything that is stored in the state when the sources are stateful as well and checkpoint their current offset. The PersistentKafkaSource provides this stateful functionality for example.
For example when implementing a rolling count over the stream Flink gives you the possibility to safely store the counter. Another common usecase is when reading from a Kafka source to save the latest committed offset to catch up from. To mark a function for checkpointing it has to implement the flink.streaming.api.checkpoint.Checkpointed interface or preferably its special case where the checkpointing can be done asynchronously, CheckpointedAsynchronously.
Checkpointing can be enabled from the StreamExecutionEnvironment using the enableCheckpointing(…) where additional parameters can be passed to modify the default 5 second checkpoint interval.
By default state checkpoints will be stored in-memory at the JobManager. Flink also supports storing the checkpoints on any flink-supported file system (such as HDFS or Tachyon) which can be set in the flink-conf.yaml. Note that the state backend must be accessible from the JobManager, use file:// only for local setups.
For example let us write a reduce function that besides summing the data it also counts have many elements it has seen.
{% highlight java %} public class CounterSum implements ReduceFunction, CheckpointedAsynchronously {
//persistent counter
private long counter = 0;
@Override
public Long reduce(Long value1, Long value2) throws Exception {
counter++;
return value1 + value2;
}
// regularly persists state during normal operation
@Override
public Serializable snapshotState(long checkpointId, long checkpointTimestamp) throws Exception {
return new Long(counter);
}
// restores state on recovery from failure
@Override
public void restoreState(Serializable state) {
counter = (Long) state;
}
} {% endhighlight %}
Stateful sources require a bit more care as opposed to other operators they are not data driven, but their run(SourceContext) methods potentially run infinitely. In order to make the updates to the state and output collection atomic the user is required to get a lock from the source's context.
{% highlight java %} public static class CounterSource implements SourceFunction, CheckpointedAsynchronously {
// utility for job cancellation
private volatile boolean isRunning = false;
private long counter;
@Override
public void run(SourceContext<Long> ctx) throws Exception {
isRunning = true;
Object lock = ctx.getCheckpointLock();
while (isRunning) {
// output and state update are atomic
synchronized (lock){
ctx.collect(counter);
counter++;
}
}
}
@Override
public void cancel() {
isRunning = false;
}
@Override
public Serializable snapshotState(long checkpointId, long checkpointTimestamp) throws Exception {
return new Long(counter);
}
@Override
public void restoreState(Serializable state) {
counter = (Long) state;
}
} {% endhighlight %}
Some operators might need the information when a checkpoint is fully acknowledged by Flink to communicate that with the outside world. In this case see the flink.streaming.api.checkpoint.CheckpointComitter interface.
Fink currently only provides processing guarantees for jobs without iterations. Enabling checkpointing on an iterative job causes an exception. In order to force checkpointing on an iterative program the user needs to set a special flag when enabling checkpointing: env.enableCheckpointing(interval, force = true).
For a more concise code one can rely on one of the main features of Java 8: lambda expressions. The following program has similar functionality to the one provided in the example section, while showcasing the usage of lambda expressions.
DataStream<String> text = env.fromElements(
"Who's there?",
"I think I hear them. Stand, ho! Who's there?");
DataStream<Tuple2<String, Integer>> counts =
// normalize and split each line
text.map(line -> line.toLowerCase().split("\\W+"))
// convert splitted line in pairs (2-tuples) containing: (word,1)
.flatMap((String[] tokens, Collector<Tuple2<String, Integer>> out) -> {
// emit the pairs with non-zero-length words
Arrays.stream(tokens)
.filter(t -> t.length() > 0)
.forEach(t -> out.collect(new Tuple2<>(t, 1)));
})
// group by the tuple field "0" and sum up tuple field "1"
.groupBy(0)
.sum(1);
counts.print();
env.execute("Streaming WordCount");
}
} {% endhighlight %}
For a detailed Java 8 Guide please refer to the Java 8 Programming Guide. Operators specific to streaming, such as Operator splitting also support this usage. Output splitting can be rewritten as follows:
Setting parallelism for operators works exactly the same way as in the batch Flink API. The user can control the number of parallel instances created for each operator by calling the operator.setParallelism(parallelism) method.
By default, data points are not transferred on the network one-by-one, which would cause unnecessary network traffic, but are buffered in the output buffers. The size of the output buffers can be set in the Flink config files. While this method is good for optimizing throughput, it can cause latency issues when the incoming stream is not fast enough. To tackle this issue the user can call env.setBufferTimeout(timeoutMillis) on the execution environment (or on individual operators) to set a maximum wait time for the buffers to fill up. After this time, the buffers are flushed automatically even if they are not full. The default value for this timeout is 100 ms, which should be appropriate for most use-cases.
Usage:
env.genereateSequence(1,10).map(new MyMapper()).setBufferTimeout(timeoutMillis); {% endhighlight %}
env.genereateSequence(1,10).map(myMap).setBufferTimeout(timeoutMillis) {% endhighlight %}
To maximise the throughput the user can call setBufferTimeout(-1) which will remove the timeout and buffers will only be flushed when they are full. To minimise latency, set the timeout to a value close to 0 (for example 5 or 10 ms). Theoretically, a buffer timeout of 0 will cause all output to be flushed when produced, but this setting should be avoided, because it can cause severe performance degradation.
Connectors provide an interface for accessing data from various third party sources (message queues). Currently three connectors are natively supported, namely Apache Kafka, RabbitMQ and the Twitter Streaming API.
Typically the connector packages consist of a source and sink class (with the exception of Twitter where only a source is provided). To use these sources the user needs to pass Serialization/Deserialization schemas for the connectors for the desired types. (Or use some predefined ones)
To run an application using one of these connectors, additional third party components are usually required to be installed and launched, e.g. the servers for the message queues. Further instructions for these can be found in the corresponding subsections. Docker containers are also provided encapsulating these services to aid users getting started with connectors.
This connector provides access to data streams from Apache Kafka. To use this connector, add the following dependency to your project:
{% highlight xml %} org.apache.flink flink-connector-kafka {{site.version }} {% endhighlight %}
Note that the streaming connectors are currently not part of the binary distribution. See linking with them for cluster execution here.
advertised.host.name setting in the config/server.properties file the must be set to the machine's IP address.The standard KafkaSource is a Kafka consumer providing an access to one topic.
The following parameters have to be provided for the KafkaSource(...) constructor:
Example:
As Kafka persists all the data, a fault tolerant Kafka source can be provided.
The PersistentKafkaSource can read a topic, and if the job fails for some reason, the source will continue on reading from where it left off after a restart. For example if there are 3 partitions in the topic with offsets 31, 122, 110 read at the time of job failure, then at the time of restart it will continue on reading from those offsets, no matter whether these partitions have new messages.
To use fault tolerant Kafka Sources, monitoring of the topology needs to be enabled at the execution environment:
Also note that Flink can only restart the topology if enough processing slots are available to restart the topology. So if the topology fails due to loss of a TaskManager, there must still be enough slots available afterwards. Flink on YARN supports automatic restart of lost YARN containers.
The following arguments have to be provided for the PersistentKafkaSource(...) constructor:
Example:
A class providing an interface for sending data to Kafka.
The followings have to be provided for the KafkaSink(…) constructor in order:
Example:
The user can also define custom Kafka producer configuration for the KafkaSink with the constructor:
If this constructor is used, the user needs to make sure to set the broker with the “metadata.broker.list” property. Also the serializer configuration should be left default, the serialization should be set via SerializationSchema.
More about Kafka can be found here.
This connector provides access to data streams from RabbitMQ. To use this connector, add the following dependency to your project:
{% highlight xml %} org.apache.flink flink-connector-rabbitmq {{site.version }} {% endhighlight %}
Note that the streaming connectors are currently not part of the binary distribution. See linking with them for cluster execution here.
Follow the instructions from the RabbitMQ download page. After the installation the server automatically starts, and the application connecting to RabbitMQ can be launched.
A class providing an interface for receiving data from RabbitMQ.
The followings have to be provided for the RMQSource(…) constructor in order:
Example:
A class providing an interface for sending data to RabbitMQ.
The followings have to be provided for the RMQSink(…) constructor in order:
Example:
More about RabbitMQ can be found here.
Twitter Streaming API provides opportunity to connect to the stream of tweets made available by Twitter. Flink Streaming comes with a built-in TwitterSource class for establishing a connection to this stream. To use this connector, add the following dependency to your project:
{% highlight xml %} org.apache.flink flink-connector-twitter {{site.version }} {% endhighlight %}
Note that the streaming connectors are currently not part of the binary distribution. See linking with them for cluster execution here.
In order to connect to Twitter stream the user has to register their program and acquire the necessary information for the authentication. The process is described below.
First of all, a Twitter account is needed. Sign up for free at twitter.com/signup or sign in at Twitter's Application Management and register the application by clicking on the “Create New App” button. Fill out a form about your program and accept the Terms and Conditions. After selecting the application, the API key and API secret (called consumerKey and sonsumerSecret in TwitterSource respectively) is located on the “API Keys” tab. The necessary access token data (token and secret) can be acquired here. Remember to keep these pieces of information a secret and do not push them to public repositories.
Create a properties file, and pass its path in the constructor of TwitterSource. The content of the file should be similar to this:
#properties file for my app secret=*** consumerSecret=*** token=***-*** consumerKey=***
The TwitterSource class has two constructors.
public TwitterSource(String authPath, int numberOfTweets); to emit finite number of tweetspublic TwitterSource(String authPath); for streamingBoth constructors expect a String authPath argument determining the location of the properties file containing the authentication information. In the first case, numberOfTweets determine how many tweet the source emits.
In contrast to other connectors, the TwitterSource depends on no additional services. For example the following code should run gracefully:
The TwitterSource emits strings containing a JSON code. To retrieve information from the JSON code you can add a FlatMap or a Map function handling JSON code. For example, there is an implementation JSONParseFlatMap abstract class among the examples. JSONParseFlatMap is an extension of the FlatMapFunction and has a
function which can be use to acquire the value of a given field.
There are two basic types of tweets. The usual tweets contain information such as date and time of creation, id, user, language and many more details. The other type is the delete information.
TwitterLocal is an example how to use TwitterSource. It implements a language frequency counter program.
A Docker container is provided with all the required configurations for test running the connectors of Apache Flink. The servers for the message queues will be running on the docker container while the example topology can be run on the user's computer.
The official Docker installation guide can be found here. After installing Docker an image can be pulled for each connector. Containers can be started from these images where all the required configurations are set.
For the easiest setup, create a jar with all the dependencies of the flink-streaming-connectors project.
cd /PATH/TO/GIT/flink/flink-staging/flink-streaming-connectors mvn assembly:assembly ~~~bash This creates an assembly jar under *flink-streaming-connectors/target*. #### RabbitMQ Pull the docker image: ~~~bash sudo docker pull flinkstreaming/flink-connectors-rabbitmq
To run the container, type:
sudo docker run -p 127.0.0.1:5672:5672 -t -i flinkstreaming/flink-connectors-rabbitmq
Now a terminal has started running from the image with all the necessary configurations to test run the RabbitMQ connector. The -p flag binds the localhost‘s and the Docker container’s ports so RabbitMQ can communicate with the application through these.
To start the RabbitMQ server:
sudo /etc/init.d/rabbitmq-server start
To launch the example on the host computer, execute:
java -cp /PATH/TO/JAR-WITH-DEPENDENCIES org.apache.flink.streaming.connectors.rabbitmq.RMQTopology \ > log.txt 2> errorlog.txt
There are two connectors in the example. One that sends messages to RabbitMQ, and one that receives messages from the same queue. In the logger messages, the arriving messages can be observed in the following format:
<DATE> INFO rabbitmq.RMQTopology: String: <one> arrived from RMQ <DATE> INFO rabbitmq.RMQTopology: String: <two> arrived from RMQ <DATE> INFO rabbitmq.RMQTopology: String: <three> arrived from RMQ <DATE> INFO rabbitmq.RMQTopology: String: <four> arrived from RMQ <DATE> INFO rabbitmq.RMQTopology: String: <five> arrived from RMQ
Pull the image:
sudo docker pull flinkstreaming/flink-connectors-kafka
To run the container type:
sudo docker run -p 127.0.0.1:2181:2181 -p 127.0.0.1:9092:9092 -t -i \ flinkstreaming/flink-connectors-kafka
Now a terminal has started running from the image with all the necessary configurations to test run the Kafka connector. The -p flag binds the localhost‘s and the Docker container’s ports so Kafka can communicate with the application through these. First start a zookeeper in the background:
/kafka_2.9.2-0.8.1.1/bin/zookeeper-server-start.sh /kafka_2.9.2-0.8.1.1/config/zookeeper.properties \ > zookeeperlog.txt &
Then start the kafka server in the background:
/kafka_2.9.2-0.8.1.1/bin/kafka-server-start.sh /kafka_2.9.2-0.8.1.1/config/server.properties \ > serverlog.txt 2> servererr.txt &
To launch the example on the host computer execute:
java -cp /PATH/TO/JAR-WITH-DEPENDENCIES org.apache.flink.streaming.connectors.kafka.KafkaTopology \ > log.txt 2> errorlog.txt
In the example there are two connectors. One that sends messages to Kafka, and one that receives messages from the same queue. In the logger messages, the arriving messages can be observed in the following format:
<DATE> INFO kafka.KafkaTopology: String: (0) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (1) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (2) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (3) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (4) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (5) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (6) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (7) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (8) arrived from Kafka <DATE> INFO kafka.KafkaTopology: String: (9) arrived from Kafka