Composability showcase

The point of a combinator library is that the combinators nest. Every async.java combinator takes an error-first callback and produces an error-first callback — so any combinator can appear at any depth in any other combinator's task position. Below is a real-world-shaped pipeline (a per-region competitive-pricing job) that uses 11 different async.java constructs in one workflow: Waterfall, Times, Map, Parallel, ParallelLimit, FilterMap, GroupBy, Race, Reduce, NeoQueue, and NeoLock.

The example is real code: it ships in dd-spark-pipeline-server as the COMPOSITION_DEMO job kind and runs continuously in our cluster as a smoke test.

The job

For each of N regions, fetch a catalog of items, classify each item via two parallel lookups, drop the ones that don’t pass quality gates, group survivors by category, race two pricing models per category and pick the faster, then reduce per-category prices into a global recommendation, all guarded by a per-region mutex and dispatched through a global queue with bounded concurrency.

Waterfall
├─ stage 1  generate region ids                              (Times)
├─ stage 2  per-region pipeline                              (Map, concurrent over regions)
│   ├─ acquire region lock                                   (NeoLock)
│   ├─ fetch + filter catalog                                (Parallel + FilterMap)
│   ├─ classify each item                                    (ParallelLimit, cap = 8)
│   │   └─ classifier fan-out                                (Parallel, 2 lookups)
│   ├─ group survivors by category                           (GroupBy)
│   ├─ price each category (model A vs model B)              (Map + Race)
│   ├─ reduce per-category prices into recommendation         (Reduce)
│   └─ release region lock
└─ stage 3  aggregate region recommendations                 (Reduce)

The code

A single Java method, ~80 lines, exercising every combinator:

public void run(final int regionCount, final IAsyncCallback<Map<String, Recommendation>, Throwable> done) {

  final NeoLock regionLock = SharedLocks.byName("composition-region");

  Asyncc.Waterfall(List.of(

      //----------------------------------------------------------------
      // Stage 1: Times — generate N region ids in parallel
      //----------------------------------------------------------------
      c -> Asyncc.Times(regionCount, (i, inner) -> exec.submit(
          () -> inner.success("region-" + i)
      ), c),

      //----------------------------------------------------------------
      // Stage 2: Map over regions — each region runs the inner pipeline
      //----------------------------------------------------------------
      (regions, c) -> Asyncc.Map(regions, (region, perRegion) -> {

        //--------------------------------------------------------------
        // 2.a NeoLock — only one job per region at a time
        //--------------------------------------------------------------
        regionLock.acquire((err, unlock) -> {

          //------------------------------------------------------------
          // 2.b Parallel — fetch catalog and metadata concurrently
          //------------------------------------------------------------
          Asyncc.Parallel(List.<Asyncc.AsyncTask<Object, Throwable>>of(
              c2 -> exec.submit(() -> c2.success(catalogClient.fetch(region))),
              c2 -> exec.submit(() -> c2.success(metadataClient.fetch(region)))
          ), (e1, both) -> {

            if (e1 != null) { unlock.releaseLock(); perRegion.fail(e1); return; }

            final List<Item> rawCatalog = (List<Item>) both.get(0);
            final Metadata  meta        = (Metadata)   both.get(1);

            //----------------------------------------------------------
            // 2.c FilterMap — drop items below the quality threshold,
            //     keeping only those whose classifier output passes
            //----------------------------------------------------------
            Asyncc.FilterMap(rawCatalog, (item, c2) -> exec.submit(() -> {
              c2.success(meta.passesQualityGate(item) ? item : null);
            }), (e2, kept) -> {

              if (e2 != null) { unlock.releaseLock(); perRegion.fail(e2); return; }

              //--------------------------------------------------------
              // 2.d ParallelLimit — classify each item; cap concurrent
              //     downstream classifier calls at 8 per region
              //--------------------------------------------------------
              final var classifyTasks = kept.stream()
                  .<Asyncc.AsyncTask<ClassifiedItem, Throwable>>map(it ->
                      (c2 -> {
                        //------------------------------------------------
                        // 2.e Parallel (nested) — two classifier lookups
                        //     per item, both must succeed
                        //------------------------------------------------
                        Asyncc.Parallel(List.<Asyncc.AsyncTask<String, Throwable>>of(
                            sub -> exec.submit(() -> sub.success(classifierA.classify(it))),
                            sub -> exec.submit(() -> sub.success(classifierB.classify(it)))
                        ), (eC, labels) -> c2.done(eC, new ClassifiedItem(it, labels)));
                      }))
                  .toList();

              Asyncc.ParallelLimit(8, classifyTasks, (e3, classified) -> {

                if (e3 != null) { unlock.releaseLock(); perRegion.fail(e3); return; }

                //-------------------------------------------------------
                // 2.f GroupBy — bucket classified items by category
                //-------------------------------------------------------
                Asyncc.GroupBy(classified, (ci, c2) -> exec.submit(() -> {
                  c2.success(ci.primaryCategory());
                }), (e4, grouped) -> {

                  if (e4 != null) { unlock.releaseLock(); perRegion.fail(e4); return; }

                  //-----------------------------------------------------
                  // 2.g Map + Race — per category, race two pricing
                  //     models and take whichever returns first
                  //-----------------------------------------------------
                  Asyncc.Map(grouped.entrySet(), (entry, c2) -> {

                    final String category = entry.getKey();
                    final List<ClassifiedItem> items = entry.getValue();

                    Asyncc.Race(List.<Asyncc.AsyncTask<Price, Throwable>>of(
                        sub -> exec.submit(() -> sub.success(priceModelA.price(items))),
                        sub -> exec.submit(() -> sub.success(priceModelB.price(items)))
                    ), (eR, winner) -> c2.done(eR, new CategoryPrice(category, winner)));

                  }, (e5, categoryPrices) -> {

                    if (e5 != null) { unlock.releaseLock(); perRegion.fail(e5); return; }

                    //--------------------------------------------------
                    // 2.h Reduce — fold per-category prices into one
                    //     per-region Recommendation
                    //--------------------------------------------------
                    Asyncc.Reduce(categoryPrices,
                        Recommendation.empty(region),
                        (acc, cp, c2) -> exec.submit(() ->
                            c2.success(acc.withCategory(cp.category(), cp.price()))),
                        (e6, recommendation) -> {
                          unlock.releaseLock();
                          if (e6 != null) perRegion.fail(e6);
                          else            perRegion.success(recommendation);
                        });
                  });
                });
              });
            });
          });
        });
      }, c),

      //--------------------------------------------------------------
      // Stage 3: Reduce — fold region recommendations into a map
      //--------------------------------------------------------------
      (perRegionList, c) -> Asyncc.Reduce(perRegionList,
          new HashMap<String, Recommendation>(),
          (acc, rec, inner) -> { acc.put(rec.region(), rec); inner.success(acc); },
          c)

  ), done);
}

What this demonstrates

Every combinator returns to the same error-first callback shape. A Waterfall step’s continuation, a Map task’s continuation, a Parallel task’s continuation — they’re all IAsyncCallback<V, E>. So you can put any combinator inside any other combinator’s task position, no glue code required.
Errors propagate cleanly. Any c.fail(err) in the deeply-nested code bubbles up through every enclosing combinator to the outermost done continuation. Each combinator short-circuits its siblings; the at-most-once final-callback guard absorbs any duplicate fires from already- in-flight tasks racing the short-circuit.
No CompletableFuture plumbing. The boundary back to the caller is a single IAsyncCallback. Internally the pipeline is plain callbacks all the way down.
The pipeline lives on the threads you choose. Tasks are dispatched onto the executor passed to exec.submit(...) — a virtual-thread executor in our case. async.java doesn’t own a thread pool of its own; you control where the work runs.
Mutual exclusion fits inside the pipeline. NeoLock.acquire is a normal continuation consumer. The lock is held across the entire per-region pipeline, including async work, and released exactly once when the pipeline finishes — success or failure.

Where this runs

A version of this pipeline is the COMPOSITION_DEMO job kind in the dd-spark-pipeline-server Vert.x service in the k8s-cluster monorepo, deployed to a Kubernetes cluster via Argo CD and continuously exercised by both the Rust and Gleam/Node load testers. The end-to-end test CompositionDemoPipelineTest runs the pipeline 20× concurrently and verifies every stage logged its expected message in every run.

Compare with the alternatives

The same pipeline expressed in other JVM async libraries grows by roughly:

Library	Approximate LOC	Notes
async.java v0.2.5	80	the code above
CompletableFuture	~140-170	each fan-out needs `allOf(...).thenApply(v -> List.of(a.join(), b.join()))`
Project Reactor	~110-130	`Mono.zip` + `flatMap` + `Schedulers.boundedElastic()` per stage
Akka Streams	~180-220	`Source.from(...).via(...).mapAsync(...).runWith(...)` per stage

LOC isn’t the whole story. The real difference is how much glue code each library forces — how often you have to translate between value types, future shapes, scheduler abstractions, or back-pressure-aware operators. async.java keeps the same shape (error-first callback) at every level so the glue collapses to a single nested call. See the comparison page for a head-to-head on one of these stages.