Mapping Data in a Pipeline Using a a gRPC Service

When processing data in a pipeline, you can use an external service to execute a function on input data, and pass the output to the next stage of pipeline. A popular way of exposing external services is to use the gRPC protocol.

gRPC is an open-source universal RPC framework available for many platforms and languages.

The hazelcast-jet-grpc module makes it easy to perform calls to a gRPC service. Hazelcast supports two gRPC service methods:

  • unary RPC

  • bidirectional streaming RPC


Add this dependency to your Java project:

  • Gradle

  • Maven

compile 'com.hazelcast.jet:hazelcast-jet-grpc:5.4.1'

The Hazelcast cluster must also have this module on the classpath. Make sure it is present in the lib directory, if not add it there and restart the cluster.

Unary RPC

The classical request-response RPC pattern is what gRPC calls "unary RPC". You send a single request message and get a single response message back, just like a plain function call.

Let’s use this protobuf definition as an example:

service ProductService {
  rpc ProductInfo (ProductInfoRequest)
      returns (ProductInfoReply) {}

message ProductInfoRequest {
  int32 id = 1;

message ProductInfoReply {
  string productName = 1;

To call this service, use GrpcServices.unaryService():

ServiceFactory<?, ? extends GrpcService<ProductInfoRequest, ProductInfoReply>>
productService = unaryService(
    () -> ManagedChannelBuilder.forAddress("localhost", PORT) .usePlaintext(),
    channel -> ProductServiceGrpc.newStub(channel)::productInfo

The first parameter is a factory function that returns a channel builder. Modify the builder settings as required, as an example we could have enabled TLS via io.grpc.ManagedChannelBuilder.useTransportSecurity().

The second parameter is a function which, given a gRPC network channel, creates a client-side stub and returns a reference to its method that invokes the service. The stub code is auto-generated by the protobuf compiler.

Returning a method reference isn’t a requirement, though, you can also modify the stub, the input item or anything else you need. The functional type you must comply with is as follows (wildcards omitted for clarity):

FunctionEx<ManagedChannel, BiConsumerEx<T, StreamObserver<R>>> callStubFn

Here, BiConsumerEx<T, StreamObserver<R>> corresponds to the signature of the generated gRPC method: it is a function that takes an input item and the gRPC result observer, and ensures that the observer eventually receives the gRPC invocation result.

Now you can use the service factory in any of the`mapUsingService*Async` methods:

StreamStage<Trade> trades = ...
(service, trade) -> {
    ProductInfoRequest request = ProductInfoRequest.newBuilder()
    return ->
            tuple2(trade, productReply.getProductName()));

Bidirectional Streaming RPC

Bidirectional streaming RPC is an extension of the classic RPC paradigm to streaming. You make a single RPC call, within which you can send any number of messages to the server, and it can send any number of messages to you. Since we are using gRPC to transform pipeline items, this general protocol is constrained: the response stream must match one-for-one with the request stream.

This constraint makes the streaming communication very similar to unary RPC, but it eliminates some of the overheads. Sending messages within a single established call is cheaper than creating a new call from scratch. Our benchmarks show 1.5x to 3x improvement, depending on various factors.

This is an example of a bidirectional streaming RPC definition:

service BrokerService {
  rpc BrokerInfo (stream BrokerInfoRequest)
      returns (stream BrokerInfoReply) {}

message BrokerInfoRequest {
  int32 id = 1;

message BrokerInfoReply {
  string brokerName = 1;

Note the stream keyword appearing both in the request and the response.

We can create the following service factory using GrpcServices.bidirectionalStreamingService() method:

ServiceFactory<?, ? extends GrpcService<BrokerInfoRequest, BrokerInfoReply>>
brokerService = bidirectionalStreamingService(
    () -> ManagedChannelBuilder.forAddress("localhost", PORT).usePlaintext(),
    channel -> BrokerServiceGrpc.newStub(channel)::brokerInfo

As with the unary service, the first parameter is a supplier returning a channel builder.

The full type of the second parameter is as follows (with wildcards omitted):

FunctionEx<ManagedChannel, FunctionEx<StreamObserver<R>, StreamObserver<T>>> callStubFn

The service method is now a function with the signature

StreamObserver<R> -> StreamObserver<T>

Hazelcast provides its observer of the output (same as in the unary RPC), and this function returns an object where Hazelcast will push the input items.

Now the service factory can be used in any of the mapUsingService* methods, preferably the mapUsingServiceAsync.

StreamStage<Tuple2<Trade, String>> tradeAndProducts = ...
    (service, t) -> {
        BrokerInfoRequest request = BrokerInfoRequest
        return service
            .thenApply(brokerReply ->
                tuple3(t.f0(), t.f1(), brokerReply.getBrokerName()));

Improving Throughput with Batching

If your gRPC service’s throughput capacity is very high, and the gRPC link is the bottleneck, you can significantly improve the throughput by applying batching. For example, you can use a protobuf definition like this one (note the repeated keyword):

service Greeter {
  rpc SayHelloListBidirectional (stream HelloRequestList)
      returns (stream HelloReplyList) {}
message HelloRequestList {
  repeated string name = 1;
message HelloReplyList {
  repeated string message = 1;

Create the service in a way similar to previous example:

ServiceFactory<?, ? extends GrpcService<HelloRequestList, HelloReplyList>> bidiService =
    () -> ManagedChannelBuilder.forAddress(host, port).usePlaintext(),
    channel -> GreeterGrpc.newStub(channel)::sayHelloListBidirectional

In the pipeline, use the specialized mapUsingServiceAsyncBatched transform:

StreamStage<String> stage = ...
    (service, itemList) -> {
        CompletableFuture<HelloReplyList> future =
        return future.thenApply(HelloReplyList::getMessageList);

If your batch takes more than ~0.8 seconds (including the network overhead), you should increase the value of the following properties so that the clean shutdown succeeds:


The GrpcProperties JavaDoc provides more details about these properties.

See the grpc example module for a complete code example.