Tricolor: Multiple classloaders, multiple Hazelcasts, multiple loggers and multiple colours — all in the one JVM



The normal deployment is for a JVM to contain a single Hazelcast instance, a client or a server. This means that the instance can utilise all the resources available to that JVM.

In automated tests, it can frequently be useful to run multiple Hazelcast instances in the one JVM so that it is easy to control the sequence that the instances interact with each other in a predictable way.

Alternatively, it could be necessary to connect to two different clusters at once, to efficiently copy data from one to another.

The Classloader Constraint

If you run two Hazelcast instances in a JVM, by default they will both be instances of the same class, com.hazelcast.core.HazelcastInstance.

What that means in JVM terms is that there are two HazelcastInstance objects but their getClass() method refers to the same class object.

The purpose of this example is to show how to bypass this, to create several HazelcastInstance objects, each with a unique class object.

For the purpose of this example, the goal is to have each using a different coloured logger. Each logger uses a different colour, so that when multiple instances are logging in the one JVM, it is easy to tell the logging messages produced by each instance.

The logger used is Log4j version 2, the successor to the original Log4j which is now retired.

The Example

The example is structured as follows

├── src/main/java/
├── ├── com.hazelcast.samples.tricolor/
├── ├── ├──
├── ├── ├──
├── ├── ├──
├── ├── ├──
├── src/main/resources/
├── ├── blue.xml
├── ├── green.xml
├── ├── hazelcast-client.xml
├── ├── hazelcast.xml
├── ├── red.xml
├── pom.xml

Simply, this process creates three threads — two to invoke Hazelcast servers, and one to invoke a Hazelcast client. It starts them running, lets them finish, then shuts down.

Internally, these three created threads use a CountDownLatch and a ClientListener to control their sequencing. That isn’t visible to Main, it just starts them off and waits.

What this class does differently from normal is in the myLoadClass method. This makes sure that the created threads have different classloaders, and this allows them to have different loggers, which in turn is what allows output to be logged in different colours.


This method loads a class with a new, separate classloader. As we use this to create a thread, the items this thread then loads — Hazelcast and the logging framework — are in that separate classloader.

The mechanism is to sibling the current classloader. A reference is made to the resource locations (URLs) that the current classloader searches, then we create a new classloader from the parent of the current classloader with those resource locations.

In efect, we duplicate the current classloader. The newly created classloader has the same ancestry and looks in the same places for resources (such as Hazelcast), but when it finds them it creates a separate class instance.

The callable object here is fairly straight-forward, but it implements three interfaces.

It implements Callable, so that it can be invoked by a thread executor.

It implements Serializable, so that it can be sent from one Hazelcast instance (the client) to others (the servers).

And it implements HazelcastInstanceAware so that the Hazelcast instance it runs on is injected.

When it runs, it logs out which server it is running on. Because we will run it on all servers, you should see this logging message appear in blue and in green.

When it completes, it returns the id of the server instance it ran on, and the client logs this. So you will see this in red.

The client thread is started with the argument red.xml, so it will pick up its logging configuration from a file of the same name, and produce log output in red.

The thread waits on countdown latch until all the servers are adequately started, then creates a Hazelcast client instance based on hazelcast-client.xml.

Use of a countdown latch is a safer mechanism to control execution of multiple Hazelcast instances in the one thread. They could all be started at once, and the client would try to connect to the servers a few times before giving up — so this might work if the servers start quickly, and might fail if your machine is slow.

This client then submits a callable, logs the output, and shuts down.

All that really deviates from typical usage is the class loader, a different one from the class loader used by the two server threads.

Two instances are created for the MyServer class. One is passed in the argument green.xml, uses that log file, and produces log output in green. The other is given the argument blue.xml for blue log output.

Each of these threads decrements a countdown latch when the Hazelcast server instance inside it is fully started. This is used so the client instance in another thread does not try to start until all servers are started. This gives a degree of predictability to the execution, although the two servers are started in parallel so there is still some randomness.

Once started, the servers use a ClientListener to detect when the client has disconnected. On the assumption the client disconnection means the client has finished, the server shuts down.

Run it

The code uses Spring Boot for packaging, so you need to run the build at least as far as the package phase of Maven, with a command such as:

    mvn install

Then run the executable jar file produced:

    java -jar target/tricolor.jar

Assuming you have multi-coloured output for your terminal window, you should see something like this:

Sample Output

You should get a message from each of the server threads and client thread that each has started. Probably you will get a message that the client thread is waiting for the server threads to be ready, but you may not get this depending on your machine speed.

The client will submit a collable which will log on each server it runs on, then finally all three threads will end and the process as a whole will close.

Leave a Reply