Clone or download
sasbury Pushed version to 2.0.2.
Fixed missing param in javadoc.
Latest commit 2aac130 Aug 12, 2018


NATS - Java Client

A Java client for the NATS messaging system.

License Apache 2 Build Status Coverage Status Maven Central Javadoc

A Note on Versions

This is version 2.0 of the java-nats library. This version is a ground up rewrite of the original library. Part of the goal of this re-write was to address the excessive use of threads, we created a Dispatcher construct to allow applications to control thread creation more intentionally. This version also removes all non-JDK runtime dependencies.

The API is simple to use and highly performant.

Version 2.0 uses a simplified versioning scheme. Any issues will be fixed in the incremental version number. As a major release, the major version has been updated to 2.0 to allow clients to limit there use of this new API.

Previous versions are still available in the repo.


The java-nats client is provided in a single jar file, with no external dependencies. See Building From Source for details on building the library.

Downloading the Jar

You can download the latest jar at

Using Gradle

The NATS client is available in the Maven central repository, and can be imported as a standard dependency in your build.gradle file:

dependencies {
    implementation 'io.nats:jnats:2.0.0'

If you need the latest and greatest before Maven central updates, you can use:

repositories {
    maven {
        url ""
    maven {
        url ""

Using Maven

The NATS client is available on the Maven central repository, and can be imported as a normal dependency in your pom.xml file:


If you need the absolute latest, before it propagates to maven central, you can use the repository:


If you are using the 1.x version of java-nats and don't want to upgrade to 2.0.0 please use ranges in your POM file, java-nats-streaming 1.x is using [1.1, 1.9.9) for this.

Linux Platform Note

NATS uses RNG to generate unique inbox names. A peculiarity of the JDK on Linux (see JDK-6202721 and JDK-6521844) causes Java to use /dev/random even when /dev/urandom is called for. The net effect is that successive calls to newInbox(), either directly or through calling request() will become very slow, on the order of seconds, making many applications unusable if the issue is not addressed. A simple workaround would be to use the following jvm args.

Basic Usage

Sending and receiving with NATS is as simple as connecting to the gnatsd and publishing or subscribing for messages. A number of examples are provided in this repo as described in


There are four different ways to connect using the Java library:

  1. Connect to a local server on the default port:

    Connection nc = Nats.connect();
  2. Connect to a server using a URL:

    Connection nc = Nats.connect("nats://myhost:4222");
  3. Connect to one or more servers with a custom configuration:

    Options o = new Options.Builder().server("nats://serverone:4222").server("nats://servertwo:4222").maxReconnects(-1).build();
    Connection nc = Nats.connect(o);

    See the javadoc for a complete list of configuration options.

  4. Connect asynchronously, this requires a callback to tell the application when the client is connected:

     Options options = new Options.Builder().server(Options.DEFAULT_URL).connectionListener(handler).build();
     Nats.connectAsynchronously(options, true);

    This feature is experimental, please let us know if you like it.


Once connected, publishing is accomplished via one of three methods:

  1. With a subject and message body:

    nc.publish("subject", "hello world".getBytes(StandardCharsets.UTF_8));
  2. With a subject and message body, as well as a subject for the receiver to reply to:

    nc.publish("subject", "replyto", "hello world".getBytes(StandardCharsets.UTF_8));
  3. As a request that expects a reply. This method uses a Future to allow the application code to wait for the response. Under the covers a request/reply pair is the same as a publish/subscribe only the library manages the subscription for you.

    Future<Message> incoming = nc.request("subject", "hello world".getBytes(StandardCharsets.UTF_8));
    Message msg = incoming.get(500, TimeUnit.MILLISECONDS);
    String response = new String(msg.getData(), StandardCharsets.UTF_8);

All of these methods, as well as the incoming message code use byte arrays for maximum flexibility. Applications can send JSON, Strings, YAML, Protocol Buffers, or any other format through NATS to applications written in a wide range of languages.

Listening for Incoming Messages

The Java NATS library provides two mechanisms to listen for messages, three if you include the request/reply discussed above.

  1. Synchronous subscriptions where the application code manually asks for messages and blocks until they arrive. Each subscription is associated with a single subject, although that subject can be a wildcard.

    Subscription sub = nc.subscribe("subject");
    Message msg = sub.nextMessage(Duration.ofMillis(500));
    String response = new String(msg.getData(), StandardCharsets.UTF_8);
  2. A Dispatcher that will call application code in a background thread. Dispatchers can manage multiple subjects with a single thread and single callback.

    Dispatcher d = nc.createDispatcher((msg) -> {
        String response = new String(msg.getData(), StandardCharsets.UTF_8);

Advanced Usage


NATS supports TLS 1.2. The server can be configured to verify client certificates or not. Depending on this setting the client has several options.

  1. The Java library allows the use of the tls:// protocol in its urls. This setting expects a default SSLContext to be set. You can set this default context using System properties, or in code. For example, you could run the publish example using:

    java io.nats.examples.NatsPub tls://localhost:4443 test "hello world"

    where the following properties are being set:

    This method can be used with or without client verification.

  2. During development, or behind a firewall where the client can trust the server, the library supports the opentls:// protocol which will use a special SSLContext that trusts all server certificates, but provides no client certificates.

    java io.nats.examples.NatsSub opentls://localhost:4443 test 3

    This method requires that client verification is off.

  3. Your code can build an SSLContext to work with or without client verification.

    SSLContext ctx = createContext();
    Options options = new Options.Builder().server(ts.getURI()).sslContext(ctx).build();
    Connection nc = Nats.connect(options);

If you want to try out these techniques, take a look at the for instructions.

Clusters & Reconnecting

The Java client will automatically reconnect if it loses its connection the gnatsd. If given a single server, the client will keep trying that one. If given a list of servers, the client will rotate between them. When the gnatsd servers are in a cluster, they will tell the client about the other servers, so that in the simplest case a client could connect to one server, learn about the cluster and reconnect to another server if its initial one goes down.

To tell the connection about multiple servers for the initial connection, use the servers() method on the options builder, or call server() multiple times.

String[] serverUrls = {"nats://serverOne:4222", "nats://serverTwo:4222"};
Options o = new Options.Builder().servers(serverUrls).build();

Reconnection behavior is controlled via a few options, see the javadoc for the Options.Builder class for specifics on reconnect limits, delays and buffers.


The io.nats.examples package contains two benchmarking tools, modeled after tools in other NATS clients. Both examples run against an existing gnatsd. The first called io.nats.examples.benchmark.NatsBench runs two simple tests, the first simply publishes messages, the second also receives messages. Tests are run with 1 thread/connection per publisher or subscriber. Running on an iMac (2017), with 4.2 GHz Intel Core i7 and 64GB of memory produced results like:

Starting benchmark(s) [msgs=5000000, msgsize=256, pubs=2, subs=2]
Current memory usage is 966.14 mb / 981.50 mb / 14.22 gb free/total/max
Use ctrl-C to cancel.
Pub Only stats: 9,584,263 msgs/sec ~ 2.29 gb/sec
 [ 1] 4,831,495 msgs/sec ~ 1.15 gb/sec (2500000 msgs)
 [ 2] 4,792,145 msgs/sec ~ 1.14 gb/sec (2500000 msgs)
  min 4,792,145 | avg 4,811,820 | max 4,831,495 | stddev 19,675.00 msgs
Pub/Sub stats: 3,735,744 msgs/sec ~ 912.05 mb/sec
 Pub stats: 1,245,680 msgs/sec ~ 304.12 mb/sec
  [ 1] 624,385 msgs/sec ~ 152.44 mb/sec (2500000 msgs)
  [ 2] 622,840 msgs/sec ~ 152.06 mb/sec (2500000 msgs)
   min 622,840 | avg 623,612 | max 624,385 | stddev 772.50 msgs
 Sub stats: 2,490,461 msgs/sec ~ 608.02 mb/sec
  [ 1] 1,245,230 msgs/sec ~ 304.01 mb/sec (5000000 msgs)
  [ 2] 1,245,231 msgs/sec ~ 304.01 mb/sec (5000000 msgs)
   min 1,245,230 | avg 1,245,230 | max 1,245,231 | stddev .71 msgs
Final memory usage is 2.02 gb / 2.94 gb / 14.22 gb free/total/max

The second, called io.nats.examples.autobench.NatsAutoBench runs a series of tests with various message sizes. Running this test on the same iMac, resulted in:

PubOnly 0b           10,000,000          8,464,850 msg/s       0.00 b/s
PubOnly 8b           10,000,000         10,065,263 msg/s     76.79 mb/s
PubOnly 32b          10,000,000         12,534,612 msg/s    382.53 mb/s
PubOnly 256b         10,000,000          7,996,057 msg/s      1.91 gb/s
PubOnly 512b         10,000,000          5,942,165 msg/s      2.83 gb/s
PubOnly 1k            1,000,000          4,043,937 msg/s      3.86 gb/s
PubOnly 4k              500,000          1,114,947 msg/s      4.25 gb/s
PubOnly 8k              100,000            460,630 msg/s      3.51 gb/s
PubSub 0b            10,000,000          3,155,673 msg/s       0.00 b/s
PubSub 8b            10,000,000          3,218,427 msg/s     24.55 mb/s
PubSub 32b           10,000,000          2,681,550 msg/s     81.83 mb/s
PubSub 256b          10,000,000          2,020,481 msg/s    493.28 mb/s
PubSub 512b           5,000,000          2,000,918 msg/s    977.01 mb/s
PubSub 1k             1,000,000          1,170,448 msg/s      1.12 gb/s
PubSub 4k               100,000            382,964 msg/s      1.46 gb/s
PubSub 8k               100,000            196,474 msg/s      1.50 gb/s
PubDispatch 0b       10,000,000          4,645,438 msg/s       0.00 b/s
PubDispatch 8b       10,000,000          4,500,006 msg/s     34.33 mb/s
PubDispatch 32b      10,000,000          4,458,481 msg/s    136.06 mb/s
PubDispatch 256b     10,000,000          2,586,563 msg/s    631.49 mb/s
PubDispatch 512b      5,000,000          2,187,592 msg/s      1.04 gb/s
PubDispatch 1k        1,000,000          1,369,985 msg/s      1.31 gb/s
PubDispatch 4k          100,000            403,314 msg/s      1.54 gb/s
PubDispatch 8k          100,000            203,320 msg/s      1.55 gb/s
ReqReply 0b              20,000              9,548 msg/s       0.00 b/s
ReqReply 8b              20,000              9,491 msg/s     74.15 kb/s
ReqReply 32b             10,000              9,778 msg/s    305.59 kb/s
ReqReply 256b            10,000              8,394 msg/s      2.05 mb/s
ReqReply 512b            10,000              8,259 msg/s      4.03 mb/s
ReqReply 1k              10,000              8,193 msg/s      8.00 mb/s
ReqReply 4k              10,000              7,915 msg/s     30.92 mb/s
ReqReply 8k              10,000              7,454 msg/s     58.24 mb/s
Latency 0b    5,000     35 /  49.20 / 134    +/- 0.77  (microseconds)
Latency 8b    5,000     35 /  49.54 / 361    +/- 0.80  (microseconds)
Latency 32b   5,000     35 /  49.27 / 135    +/- 0.79  (microseconds)
Latency 256b  5,000     41 /  56.41 / 142    +/- 0.90  (microseconds)
Latency 512b  5,000     40 /  56.41 / 174    +/- 0.91  (microseconds)
Latency 1k    5,000     35 /  49.76 / 160    +/- 0.80  (microseconds)
Latency 4k    5,000     36 /  50.64 / 193    +/- 0.83  (microseconds)
Latency 8k    5,000     38 /  55.45 / 206    +/- 0.88  (microseconds)

It is worth noting that in both cases memory was not a factor, the processor and OS were more of a consideration. To test this, take a look at the NatsBench results again. Those are run without any constraint on the Java heap and end up doubling the used memory. However, if we run the same test again with a constraint of 1Gb using -Xmx1g, the performance is comparable, differentiated primarily by "noise" that we can see between test runs with the same settings.

Starting benchmark(s) [msgs=5000000, msgsize=256, pubs=2, subs=2]
Current memory usage is 976.38 mb / 981.50 mb / 981.50 mb free/total/max
Use ctrl-C to cancel.

Pub Only stats: 10,123,382 msgs/sec ~ 2.41 gb/sec
 [ 1] 5,068,256 msgs/sec ~ 1.21 gb/sec (2500000 msgs)
 [ 2] 5,061,691 msgs/sec ~ 1.21 gb/sec (2500000 msgs)
  min 5,061,691 | avg 5,064,973 | max 5,068,256 | stddev 3,282.50 msgs

Pub/Sub stats: 3,563,770 msgs/sec ~ 870.06 mb/sec
 Pub stats: 1,188,261 msgs/sec ~ 290.10 mb/sec
  [ 1] 594,701 msgs/sec ~ 145.19 mb/sec (2500000 msgs)
  [ 2] 594,130 msgs/sec ~ 145.05 mb/sec (2500000 msgs)
   min 594,130 | avg 594,415 | max 594,701 | stddev 285.50 msgs
 Sub stats: 2,375,839 msgs/sec ~ 580.04 mb/sec
  [ 1] 1,187,919 msgs/sec ~ 290.02 mb/sec (5000000 msgs)
  [ 2] 1,187,920 msgs/sec ~ 290.02 mb/sec (5000000 msgs)
   min 1,187,919 | avg 1,187,919 | max 1,187,920 | stddev .71 msgs

Final memory usage is 317.62 mb / 960.50 mb / 960.50 mb free/total/max

Building From Source

The build depends on Gradle, and contains gradlew to simplify the process. After cloning, you can build the repository and run the tests with a single command:

> git clone
> cd java-nats
> ./gradlew build

This will place the class files in a new build folder. To just build the jar:

> ./gradlew jar

The jar will be placed in build/libs.

You can also build the java doc, and the samples jar using:

> ./gradlew javadoc
> ./gradlew exampleJar

The java doc is located in build/docs and the example jar is in build/libs. Finally, to run the tests with the coverage report:

> ./gradlew test jacocoTestReport

which will create a folder called build/reports/jacoco containing the file index.html you can open and use to browse the coverage. Keep in mind we have focused on library test coverage, not coverage for the examples.

Many of the tests run gnatsd on a custom port. If gnatsd is in your path they should just work, but in cases where it is not, or an IDE running tests has issues with the path you can specify the gnatsd location with the environment variable gnatsd_path.


Unless otherwise noted, the NATS source files are distributed under the Apache Version 2.0 license found in the LICENSE file.