Chapter 16. Streams

You may want to process a subset or all data in the cache to produce a result. This may bring thoughts of Map Reduce. Data Grid allows the user to do something very similar but utilizes the standard JRE APIs to do so. Java 8 introduced the concept of a Stream which allows functional-style operations on collections rather than having to procedurally iterate over the data yourself. Stream operations can be implemented in a fashion very similar to MapReduce. Streams, just like MapReduce allow you to perform processing upon the entirety of your cache, possibly a very large data set, but in an efficient way.


Streams are the preferred method when dealing with data that exists in the cache because streams automatically adjust to cluster topology changes.

Also since we can control how the entries are iterated upon we can more efficiently perform the operations in a cache that is distributed if you want it to perform all of the operations across the cluster concurrently.

A stream is retrieved from the entrySet, keySet or values collections returned from the Cache by invoking the stream or parallelStream methods.

16.1. Common stream operations

This section highlights various options that are present irrespective of what type of underlying cache you are using.

16.2. Key filtering

It is possible to filter the stream so that it only operates upon a given subset of keys. This can be done by invoking the filterKeys method on the CacheStream. This should always be used over a Predicate filter and will be faster if the predicate was holding all keys.

If you are familiar with the AdvancedCache interface you may be wondering why you even use getAll over this keyFilter. There are some small benefits (mostly smaller payloads) to using getAll if you need the entries as is and need them all in memory in the local node. However if you need to do processing on these elements a stream is recommended since you will get both distributed and threaded parallelism for free.

16.3. Segment based filtering


This is an advanced feature and should only be used with deep knowledge of Data Grid segment and hashing techniques. These segments based filtering can be useful if you need to segment data into separate invocations. This can be useful when integrating with other tools such as Apache Spark.

This option is only supported for replicated and distributed caches. This allows the user to operate upon a subset of data at a time as determined by the KeyPartitioner. The segments can be filtered by invoking filterKeySegments method on the CacheStream. This is applied after the key filter but before any intermediate operations are performed.

16.4. Local/Invalidation

A stream used with a local or invalidation cache can be used just the same way you would use a stream on a regular collection. Data Grid handles all of the translations if necessary behind the scenes and works with all of the more interesting options (ie. storeAsBinary and a cache loader). Only data local to the node where the stream operation is performed will be used, for example invalidation only uses local entries.

16.5. Example

The code below takes a cache and returns a map with all the cache entries whose values contain the string "JBoss"

Map<Object, String> jbossValues = cache.entrySet().stream()
              .filter(e -> e.getValue().contains("JBoss"))
              .collect(Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue));

16.6. Distribution/Replication/Scattered

This is where streams come into their stride. When a stream operation is performed it will send the various intermediate and terminal operations to each node that has pertinent data. This allows processing the intermediate values on the nodes owning the data, and only sending the final results back to the originating nodes, improving performance.

16.6.1. Rehash Aware

Internally the data is segmented and each node only performs the operations upon the data it owns as a primary owner. This allows for data to be processed evenly, assuming segments are granular enough to provide for equal amounts of data on each node.

When you are utilizing a distributed cache, the data can be reshuffled between nodes when a new node joins or leaves. Distributed Streams handle this reshuffling of data automatically so you don’t have to worry about monitoring when nodes leave or join the cluster. Reshuffled entries may be processed a second time, and we keep track of the processed entries at the key level or at the segment level (depending on the terminal operation) to limit the amount of duplicate processing.

It is possible but highly discouraged to disable rehash awareness on the stream. This should only be considered if your request can handle only seeing a subset of data if a rehash occurs. This can be done by invoking CacheStream.disableRehashAware() The performance gain for most operations when a rehash doesn’t occur is completely negligible. The only exceptions are for iterator and forEach, which will use less memory, since they do not have to keep track of processed keys.


Please rethink disabling rehash awareness unless you really know what you are doing.

16.6.2. Serialization

Since the operations are sent across to other nodes they must be serializable by Data Grid marshalling. This allows the operations to be sent to the other nodes.

The simplest way is to use a CacheStream instance and use a lambda just as you would normally. Data Grid overrides all of the various Stream intermediate and terminal methods to take Serializable versions of the arguments (ie. SerializableFunction, SerializablePredicate…​) You can find these methods at CacheStream. This relies on the spec to pick the most specific method as defined here.

In our previous example we used a Collector to collect all the results into a Map. Unfortunately the Collectors class doesn’t produce Serializable instances. Thus if you need to use these, there are two ways to do so:

One option would be to use the CacheCollectors class which allows for a Supplier<Collector> to be provided. This instance could then use the Collectors to supply a Collector which is not serialized.

Map<Object, String> jbossValues = cache.entrySet().stream()
              .filter(e -> e.getValue().contains("Jboss"))
              .collect(CacheCollectors.serializableCollector(() -> Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue)));

Alternatively, you can avoid the use of CacheCollectors and instead use the overloaded collect methods that take Supplier<Collector>. These overloaded collect methods are only available via CacheStream interface.

Map<Object, String> jbossValues = cache.entrySet().stream()
              .filter(e -> e.getValue().contains("Jboss"))
              .collect(() -> Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue));

If however you are not able to use the Cache and CacheStream interfaces you cannot utilize Serializable arguments and you must instead cast the lambdas to be Serializable manually by casting the lambda to multiple interfaces. It is not a pretty sight but it gets the job done.

Map<Object, String> jbossValues = map.entrySet().stream()
              .filter((Serializable & Predicate<Map.Entry<Object, String>>) e -> e.getValue().contains("Jboss"))
              .collect(CacheCollectors.serializableCollector(() -> Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue)));

The recommended and most performant way is to use an AdvancedExternalizer as this provides the smallest payload. Unfortunately this means you cannot use lamdbas as advanced externalizers require defining the class before hand.

You can use an advanced externalizer as shown below:

   Map<Object, String> jbossValues = cache.entrySet().stream()
              .filter(new ContainsFilter("Jboss"))
              .collect(() -> Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue));

   class ContainsFilter implements Predicate<Map.Entry<Object, String>> {
      private final String target;

      ContainsFilter(String target) { = target;

      public boolean test(Map.Entry<Object, String> e) {
         return e.getValue().contains(target);

   class JbossFilterExternalizer implements AdvancedExternalizer<ContainsFilter> {

      public Set<Class<? extends ContainsFilter>> getTypeClasses() {
         return Util.asSet(ContainsFilter.class);

      public Integer getId() {
         return CUSTOM_ID;

      public void writeObject(ObjectOutput output, ContainsFilter object) throws IOException {

      public ContainsFilter readObject(ObjectInput input) throws IOException, ClassNotFoundException {
         return new ContainsFilter(input.readUTF());

You could also use an advanced externalizer for the collector supplier to reduce the payload size even further.

Map<Object, String> map = (Map<Object, String>) cache.entrySet().stream()
              .filter(new ContainsFilter("Jboss"))
              .collect(Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue));

 class ToMapCollectorSupplier<K, U> implements Supplier<Collector<Map.Entry<K, U>, ?, Map<K, U>>> {
      static final ToMapCollectorSupplier INSTANCE = new ToMapCollectorSupplier();

      private ToMapCollectorSupplier() { }

      public Collector<Map.Entry<K, U>, ?, Map<K, U>> get() {
         return Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue);

   class ToMapCollectorSupplierExternalizer implements AdvancedExternalizer<ToMapCollectorSupplier> {

      public Set<Class<? extends ToMapCollectorSupplier>> getTypeClasses() {
         return Util.asSet(ToMapCollectorSupplier.class);

      public Integer getId() {
         return CUSTOM_ID;

      public void writeObject(ObjectOutput output, ToMapCollectorSupplier object) throws IOException {

      public ToMapCollectorSupplier readObject(ObjectInput input) throws IOException, ClassNotFoundException {
         return ToMapCollectorSupplier.INSTANCE;

16.7. Parallel Computation

Distributed streams by default try to parallelize as much as possible. It is possible for the end user to control this and actually they always have to control one of the options. There are 2 ways these streams are parallelized.

Local to each node When a stream is created from the cache collection the end user can choose between invoking stream or parallelStream method. Depending on if the parallel stream was picked will enable multiple threading for each node locally. Note that some operations like a rehash aware iterator and forEach operations will always use a sequential stream locally. This could be enhanced at some point to allow for parallel streams locally.

Users should be careful when using local parallelism as it requires having a large number of entries or operations that are computationally expensive to be faster. Also it should be noted that if a user uses a parallel stream with forEach that the action should not block as this would be executed on the common pool, which is normally reserved for computation operations.

Remote requests When there are multiple nodes it may be desirable to control whether the remote requests are all processed at the same time concurrently or one at a time. By default all terminal operations except the iterator perform concurrent requests. The iterator, method to reduce overall memory pressure on the local node, only performs sequential requests which actually performs slightly better.

If a user wishes to change this default however they can do so by invoking the sequentialDistribution or parallelDistribution methods on the CacheStream.

16.8. Task timeout

It is possible to set a timeout value for the operation requests. This timeout is used only for remote requests timing out and it is on a per request basis. The former means the local execution will not timeout and the latter means if you have a failover scenario as described above the subsequent requests each have a new timeout. If no timeout is specified it uses the replication timeout as a default timeout. You can set the timeout in your task by doing the following:

CacheStream<Map.Entry<Object, String>> stream = cache.entrySet().stream();
stream.timeout(1, TimeUnit.MINUTES);

For more information about this, please check the java doc in timeout javadoc.

16.9. Injection

The Stream has a terminal operation called forEach which allows for running some sort of side effect operation on the data. In this case it may be desirable to get a reference to the Cache that is backing this Stream. If your Consumer implements the CacheAware interface the injectCache method be invoked before the accept method from the Consumer interface.

16.10. Distributed Stream execution

Distributed streams execution works in a fashion very similiar to map reduce. Except in this case we are sending zero to many intermediate operations (map, filter etc.) and a single terminal operation to the various nodes. The operation basically comes down to the following:

  1. The desired segments are grouped by which node is the primary owner of the given segment
  2. A request is generated to send to each remote node that contains the intermediate and terminal operations including which segments it should process

    1. The terminal operation will be performed locally if necessary
    2. Each remote node will receive this request and run the operations and subsequently send the response back
  3. The local node will then gather the local response and remote responses together performing any kind of reduction required by the operations themselves.
  4. Final reduced response is then returned to the user

In most cases all operations are fully distributed, as in the operations are all fully applied on each remote node and usually only the last operation or something related may be reapplied to reduce the results from multiple nodes. One important note is that intermediate values do not actually have to be serializable, it is the last value sent back that is the part desired (exceptions for various operations will be highlighted below).

Terminal operator distributed result reductions The following paragraphs describe how the distributed reductions work for the various terminal operators. Some of these are special in that an intermediate value may be required to be serializable instead of the final result.

allMatch noneMatch anyMatch
The allMatch operation is ran on each node and then all the results are logically anded together locally to get the appropriate value. The noneMatch and anyMatch operations use a logical or instead. These methods also have early termination support, stopping remote and local operations once the final result is known.
The collect method is interesting in that it can do a few extra steps. The remote node performs everything as normal except it doesn’t perform the final finisher upon the result and instead sends back the fully combined results. The local thread then combines the remote and local result into a value which is then finally finished. The key here to remember is that the final value doesn’t have to be serializable but rather the values produced from the supplier and combiner methods.
The count method just adds the numbers together from each node.
findAny findFirst
The findAny operation returns just the first value they find, whether it was from a remote node or locally. Note this supports early termination in that once a value is found it will not process others. Note the findFirst method is special since it requires a sorted intermediate operation, which is detailed in the exceptions section.
max min
The max and min methods find the respective min or max value on each node then a final reduction is performed locally to ensure only the min or max across all nodes is returned.
The various reduce methods 1 , 2 , 3 will end up serializing the result as much as the accumulator can do. Then it will accumulate the local and remote results together locally, before combining if you have provided that. Note this means a value coming from the combiner doesn’t have to be Serializable.

16.11. Key based rehash aware operators

The iterator, spliterator and forEach are unlike the other terminal operators in that the rehash awareness has to keep track of what keys per segment have been processed instead of just segments. This is to guarantee an exactly once (iterator & spliterator) or at least once behavior (forEach) even under cluster membership changes.

The iterator and spliterator operators when invoked on a remote node will return back batches of entries, where the next batch is only sent back after the last has been fully consumed. This batching is done to limit how many entries are in memory at a given time. The user node will hold onto which keys it has processed and when a given segment is completed it will release those keys from memory. This is why sequential processing is preferred for the iterator method, so only a subset of segment keys are held in memory at once, instead of from all nodes.

The forEach() method also returns batches, but it returns a batch of keys after it has finished processing at least a batch worth of keys. This way the originating node can know what keys have been processed already to reduce chances of processing the same entry again. Unfortunately this means it is possible to have an at least once behavior when a node goes down unexpectedly. In this case that node could have been processing a batch and not yet completed one and those entries that were processed but not in a completed batch will be ran again when the rehash failure operation occurs. Note that adding a node will not cause this issue as the rehash failover doesn’t occur until all responses are received.

These operations batch sizes are both controlled by the same value which can be configured by invoking distributedBatchSize method on the CacheStream. This value will default to the chunkSize configured in state transfer. Unfortunately this value is a tradeoff with memory usage vs performance vs at least once and your mileage may vary.

Using iterator with replicated and distributed caches

When a node is the primary or backup owner of all requested segments for a distributed stream, Data Grid performs the iterator or spliterator terminal operations locally, which optimizes performance as remote iterations are more resource intensive.

This optimization applies to both replicated and distributed caches. However, Data Grid performs iterations remotely when using cache stores that are both shared and have write-behind enabled. In this case performing the iterations remotely ensures consistency.

16.12. Intermediate operation exceptions

There are some intermediate operations that have special exceptions, these are skip, peek, sorted 12. & distinct. All of these methods have some sort of artificial iterator implanted in the stream processing to guarantee correctness, they are documented as below. Note this means these operations may cause possibly severe performance degradation.

An artificial iterator is implanted up to the intermediate skip operation. Then results are brought locally so it can skip the appropriate amount of elements.
WARNING: This operation requires having all entries in memory on the local node. An artificial iterator is implanted up to the intermediate sorted operation. All results are sorted locally. There are possible plans to have a distributed sort which returns batches of elements, but this is not yet implemented.
WARNING: This operation requires having all or nearly all entries in memory on the local node. Distinct is performed on each remote node and then an artificial iterator returns those distinct values. Then finally all of those results have a distinct operation performed upon them.

The rest of the intermediate operations are fully distributed as one would expect.

16.13. Examples

Word Count

Word count is a classic, if overused, example of map/reduce paradigm. Assume we have a mapping of key → sentence stored on Data Grid nodes. Key is a String, each sentence is also a String, and we have to count occurrence of all words in all sentences available. The implementation of such a distributed task could be defined as follows:

public class WordCountExample {

    * In this example replace c1 and c2 with
    * real Cache references
    * @param args
   public static void main(String[] args) {
      Cache<String, String> c1 = ...;
      Cache<String, String> c2 = ...;

      c1.put("1", "Hello world here I am");
      c2.put("2", "Infinispan rules the world");
      c1.put("3", "JUDCon is in Boston");
      c2.put("4", "JBoss World is in Boston as well");
      c1.put("12","JBoss Application Server");
      c2.put("15", "Hello world");
      c1.put("14", "Infinispan community");
      c2.put("15", "Hello world");

      c1.put("111", "Infinispan open source");
      c2.put("112", "Boston is close to Toronto");
      c1.put("113", "Toronto is a capital of Ontario");
      c2.put("114", "JUDCon is cool");
      c1.put("211", "JBoss World is awesome");
      c2.put("212", "JBoss rules");
      c1.put("213", "JBoss division of RedHat ");
      c2.put("214", "RedHat community");

      Map<String, Long> wordCountMap = c1.entrySet().parallelStream()
         .map(e -> e.getValue().split("\\s"))
         .collect(() -> Collectors.groupingBy(Function.identity(), Collectors.counting()));

In this case it is pretty simple to do the word count from the previous example.

However what if we want to find the most frequent word in the example? If you take a second to think about this case you will realize you need to have all words counted and available locally first. Thus we actually have a few options.

We could use a finisher on the collector, which is invoked on the user thread after all the results have been collected. Some redundant lines have been removed from the previous example.

public class WordCountExample {
   public static void main(String[] args) {
      // Lines removed

      String mostFrequentWord = c1.entrySet().parallelStream()
         .map(e -> e.getValue().split("\\s"))
         .collect(() -> Collectors.collectingAndThen(
            Collectors.groupingBy(Function.identity(), Collectors.counting()),
               wordCountMap -> {
                  String mostFrequent = null;
                  long maxCount = 0;
                     for (Map.Entry<String, Long> e : wordCountMap.entrySet()) {
                        int count = e.getValue().intValue();
                        if (count > maxCount) {
                           maxCount = count;
                           mostFrequent = e.getKey();
                     return mostFrequent;


Unfortunately the last step is only going to be ran in a single thread, which if we have a lot of words could be quite slow. Maybe there is another way to parallelize this with Streams.

We mentioned before we are in the local node after processing, so we could actually use a stream on the map results. We can therefore use a parallel stream on the results.

public class WordFrequencyExample {
   public static void main(String[] args) {
      // Lines removed

      Map<String, Long> wordCount = c1.entrySet().parallelStream()
              .map(e -> e.getValue().split("\\s"))
              .collect(() -> Collectors.groupingBy(Function.identity(), Collectors.counting()));
      Optional<Map.Entry<String, Long>> mostFrequent = wordCount.entrySet().parallelStream().reduce(
              (e1, e2) -> e1.getValue() > e2.getValue() ? e1 : e2);

This way you can still utilize all of the cores locally when calculating the most frequent element.

Remove specific entries

Distributed streams can also be used as a way to modify data where it lives. For example you may want to remove all entries in your cache that contain a specific word.

public class RemoveBadWords {
   public static void main(String[] args) {
      // Lines removed
      String word = ..

         .filter(e -> e.getValue().contains(word))
         .forEach((c, e) -> c.remove(e.getKey()));

If we carefully note what is serialized and what is not, we notice that only the word along with the operations are serialized across to other nods as it is captured by the lambda. However the real saving piece is that the cache operation is performed on the primary owner thus reducing the amount of network traffic required to remove these values from the cache. The cache is not captured by the lambda as we provide a special BiConsumer method override that when invoked on each node passes the cache to the BiConsumer

One thing to keep in mind using the forEach command in this manner is that the underlying stream obtains no locks. The cache remove operation will still obtain locks naturally, but the value could have changed from what the stream saw. That means that the entry could have been changed after the stream read it but the remove actually removed it.

We have specifically added a new variant which is called LockedStream.

Plenty of other examples

The Streams API is a JRE tool and there are lots of examples for using it. Just remember that your operations need to be Serializable in some way.