Csdrhrt

I was reading about statelessness and came across this in doc:

Stream pipeline results may be nondeterministic or incorrect if the
behavioral parameters to the stream operations are stateful. A
stateful lambda (or other object implementing the appropriate
functional interface) is one whose result depends on any state which
might change during the execution of the stream pipeline.

Now if I have the a list of string (strList say) and then trying to remove duplicate strings from it using parallel streams in the following way:

List<String> resultOne = strList.parallelStream().distinct().collect(Collectors.toList());

or in case we want case insensitive:

List<String> result2 = strList.parallelStream().map(String::toLowerCase)

                       .distinct().collect(Collectors.toList());

Can this code have any problem as parallel streams will split the input and distinct in one chunk does not necessarily mean distinct in the whole input?

EDIT (Quick summary of the answers below)

The distinct is a stateful operation and in case of stateful intermediate operations parallel streams may require multiple passes or substantial buffering overheads. Also distinct can be implemented more efficiently if ordering of elements is not relevant.
Also as per doc:

For ordered streams, the selection of distinct elements is stable (for
duplicated elements, the element appearing first in the encounter
order is preserved.) For unordered streams, no stability guarantees
are made.

But in case of ordered stream running in parallel distinct may be unstable - means it will keep an arbitrary element in case of duplicates and not necessarily the first one as expected from distinct otherwise.

From the link:

Internally, the distinct() operation keeps a Set that contains
elements that have been seen previously, but it’s buried inside the
operation and we can’t get to it from application code.

So in case of parallel streams it would probably consume the entire stream or may use CHM (sth like ConcurrentHashMap.newKeySet()). And for ordered ones most likely it would be using LinkedHashSet or similar contruct.

Roughly pointing out the relevant parts of the doc (Emphasis mine):

Intermediate operations are further divided into stateless and
stateful operations. Stateless operations, such as filter and map,
retain no state from previously seen element when processing a new
element -- each element can be processed independently of operations
on other elements. Stateful operations, such as distinct and sorted,
may incorporate state from previously seen elements when processing
new elements

Stateful operations may need to process the entire input before
producing a result. For example, one cannot produce any results from
sorting a stream until one has seen all elements of the stream. As a
result, under parallel computation, some pipelines containing stateful
intermediate operations may require multiple passes on the data or may
need to buffer significant data. Pipelines containing exclusively
stateless intermediate operations can be processed in a single pass,
whether sequential or parallel, with minimal data buffering

If you read further down (section on ordering):

Streams may or may not have a defined encounter order. Whether or not
a stream has an encounter order depends on the source and the
intermediate operations. Certain stream sources (such as List or
arrays) are intrinsically ordered, whereas others (such as HashSet)
are not. Some intermediate operations, such as sorted(), may impose an
encounter order on an otherwise unordered stream, and others may
render an ordered stream unordered, such as BaseStream.unordered().
Further, some terminal operations may ignore encounter order, such as
forEach().

...

For parallel streams, relaxing the ordering constraint can sometimes
enable more efficient execution. Certain aggregate operations, such as
filtering duplicates (distinct()) or grouped reductions
(Collectors.groupingBy()) can be implemented more efficiently if
ordering of elements is not relevant. Similarly, operations that are
intrinsically tied to encounter order, such as limit(), may require
buffering to ensure proper ordering, undermining the benefit of
parallelism. In cases where the stream has an encounter order, but the
user does not particularly care about that encounter order, explicitly
de-ordering the stream with unordered() may improve parallel
performance for some stateful or terminal operations. However, most
stream pipelines, such as the "sum of weight of blocks" example above,
still parallelize efficiently even under ordering constraints.

In conclusion,

• distinct will work fine with parallel streams, but as you may already know, it has to consume the entire stream before continuing and this may use a lot of memory.


• If the source of the items is an unordered collection (such as hashset) or the stream is unordered(), then distinct is not worried about ordering the output and thus will be efficient

Solution is to add .unordered() to the stream pipeline if you are not worried about order and would like to see more performance.

List<String> result2 = strList.parallelStream()

                              .unordered()

                              .map(String::toLowerCase)

                              .distinct()

                              .collect(Collectors.toList());

Alas there is no (available builtin) concurrent hashset in Java (unless they got clever with ConcurrentHashMap), so I can only leave you with the unfortunate possibility that distinct is implemented in a blocking fashion using a regular Java set. In which case, I don't see any benefit of doing a parallel distinct.

Edit: I spoke too soon. There might be some benefit with using parallel streams with distinct. It looks like distinct is implemented with more cleverness than I initially thought. See @Eugene's answer.

You seem to miss quite a few things from the documentation you provide and the actual example.

Stream pipeline results may be nondeterministic or incorrect if the behavioral parameters to the stream operations are stateful.

In your example, you don't have any stateful operations defined by you. Stateful in the doc means the ones the you define, not the ones that are implemented by jdk itself - like distinct in your example. But either way you could define a stateful operation that would be correct, even Stuart Marks - working at Oracle/Java, provides such an example.

So you are more than OK in the examples that you provide, be it parallel or not.

The expensive part of distinct (in parallel) come from the fact that internally there has to be a thread-safe data structure that would keep distinct elements; in jdk case it is a ConcurrentHashMap used in case the order does not matter, or a reduction using a LinkedHashSet when order matters.

distinct btw is a pretty smart implementation, it looks if your source of the stream is already distinct (in such a case it is a no-op), or looks if your data is sorted, in which case it will do a little smarter traversal of the source (since it knows that if you have seen one element, the next to come is either the same you just seen or a different one), or using a ConcurrentHashMap internally, etc.

There won't be a problem (problem as in a wrong result) but as the API note says

Preserving stability for distinct() in parallel pipelines is relatively expensive

But if performance is of concern and if stability is not a problem (i.e the result having a different order of elements with respect to the collection it processed ) then you follow the API's note

removing the ordering constraint with BaseStream.unordered() may
result in significantly more efficient execution for distinct() in
parallel pipelines,

I thought why not benchmark performance of parallel and sequential streams for distinct

public static void main(String args) {

        List<String> strList = Arrays.asList("cat", "nat", "hat", "tat", "heart", "fat", "bat", "lad", "crab", "snob");



        List<String> words = new Vector<>();





        int wordCount = 1_000_000; // no. of words in the list words

        int avgIter = 10; // iterations to run to find average running time



        //populate a list randomly with the strings in `strList`

        for (int i = 0; i < wordCount; i++) 

            words.add(strList.get((int) Math.round(Math.random() * (strList.size() - 1))));











        //find out average running times

        long starttime, pod = 0, pud = 0, sod = 0;

        for (int i = 0; i < avgIter; i++) {

            starttime = System.currentTimeMillis();

            List<String> parallelOrderedDistinct = words.parallelStream().distinct().collect(Collectors.toList());

            pod += System.currentTimeMillis() - starttime;



            starttime = System.currentTimeMillis();

            List<String> parallelUnorderedDistinct =

                    words.parallelStream().unordered().distinct().collect(Collectors.toList());

            pud += System.currentTimeMillis() - starttime;



            starttime = System.currentTimeMillis();

            List<String> sequentialOrderedDistinct = words.stream().distinct().collect(Collectors.toList());

            sod += System.currentTimeMillis() - starttime;

        }



        System.out.println("Parallel ordered time in ms: " + pod / avgIter);

        System.out.println("Parallel unordered time in ms: " + pud / avgIter);

        System.out.println("Sequential implicitly ordered time in ms: " + sod / avgIter);

    }

The above was compiled by open-jdk 8 and run on openjdk's jre 8 (no jvm specific arguments) on an i3 6th gen (4 logical cores) and I got these results

Seemed like after a certain no. of elements, ordered parallel was faster and ironically parallel unordered was the slowest. The reason behind this (thanks to @Hulk) is the because of the way its implemented (using a HashSet).So a general rule would be that if you a few elements and a lot of duplication several magnitudes greater you might benefit from the parallel().

1)

Parallel ordered time in ms: 52

Parallel unordered time in ms: 81

Sequential implicitly ordered time in ms: 35

2)

Parallel ordered time in ms: 48

Parallel unordered time in ms: 83

Sequential implicitly ordered time in ms: 34

3)

Parallel ordered time in ms: 36

Parallel unordered time in ms: 70

Sequential implicitly ordered time in ms: 32

The unordered parallel was twice slower than both.

Then I upped wordCount to 5_000_000 and these were the results

1)

Parallel ordered time in ms: 93

Parallel unordered time in ms: 363

Sequential implicitly ordered time in ms: 123

2)

Parallel ordered time in ms: 100

Parallel unordered time in ms: 363

Sequential implicitly ordered time in ms: 124

3)

Parallel ordered time in ms: 89

Parallel unordered time in ms: 365

Sequential implicitly ordered time in ms: 118

and then to 10_000_000

1)

Parallel ordered time in ms: 148

Parallel unordered time in ms: 725

Sequential implicitly ordered time in ms: 218

2)

Parallel ordered time in ms: 150

Parallel unordered time in ms: 749

Sequential implicitly ordered time in ms: 224

3)

Parallel ordered time in ms: 143

Parallel unordered time in ms: 743

Sequential implicitly ordered time in ms: 222