October 20th, Q&A session: Get you issues solved and questions answered!

GitHub logo
Edit

Tracing

Warning
This feature is experimental.

A number of APIs in Ignite are instrumented for tracing with OpenCensus. You can collect distributed traces of various tasks executed in your cluster and use this information to diagnose latency problems.

We suggest you get familiar with OpenCensus tracing documentation before reading this chapter: https://opencensus.io/tracing/.

The following Ignite APIs are instrumented for tracing:

  • Discovery

  • Communication

  • Exchange

  • Transactions

To view traces, you must export them into an external system. You can use one of the OpenCensus exporters or write your own, but in any case, you will have to write code that registers an exporter in Ignite. Refer to Exporting Traces for details.

Configuring Tracing

Enable OpenCensus tracing in the node configuration. All nodes in the cluster must use the same tracing configuration.

<bean class="org.apache.ignite.configuration.IgniteConfiguration">

    <property name="tracingSpi">
        <bean class="org.apache.ignite.spi.tracing.opencensus.OpenCensusTracingSpi"/>
    </property>

</bean>
IgniteConfiguration cfg = new IgniteConfiguration();

cfg.setTracingSpi(new org.apache.ignite.spi.tracing.opencensus.OpenCensusTracingSpi());

Ignite ignite = Ignition.start(cfg);
This API is not presently available for C++. You can use XML configuration.

Enabling Trace Sampling

When you start your cluster with the above configuration, Ignite does not collect traces. You have to enable trace sampling for a specific API at runtime. You can turn trace sampling on and off at will, for example, only for the period when you are troubleshooting a problem.

You can do this in two ways:

  • via the control script from the command line

  • programmatically

Traces are collected at a given probabilistic sampling rate. The rate is specified as a value between 0.0 and 1.0 inclusive: 0 means no sampling, 1 means always sampling.

When the sampling rate is set to a value greater than 0, Ignite collects traces. To disable trace collection, set the sampling rate to 0.

The following sections describe the two ways of enabling trace sampling.

Using Control Script

Go to the {IGNITE_HOME}/bin directory of your Ignite installation. Enable experimental commands in the control script:

export IGNITE_ENABLE_EXPERIMENTAL_COMMAND=true

Enable tracing for a specific API:

./control.sh --tracing-configuration set --scope TX --sampling-rate 1

Refer to the Control Script sections for the list of all parameters.

Programmatically

Once you start the node, you can enable trace sampling as follows:

Ignite ignite = Ignition.start();

ignite.tracingConfiguration().set(
        new TracingConfigurationCoordinates.Builder(Scope.TX).build(),
        new TracingConfigurationParameters.Builder().withSamplingRate(1).build());

The --scope parameter specifies the API you want to trace. The following APIs are instrumented for tracing:

  • DISCOVERY — discovery events

  • EXCHANGE — exchange events

  • COMMUNICATION — communication events

  • TX — transactions

The --sampling-rate is the probabilistic sampling rate, a number between 0 and 1:

  • 0 means no sampling,

  • 1 means always sampling.

Exporting Traces

To view traces, you need to export them to an external backend using one of the available exporters. OpenCensus supports a number of exporters out-of-the-box, and you can write a custom one. Refer to the OpenCensus Exporters for details.

In this section, we will show how to export traces to Zipkin.

  1. Follow this guide to launch Zipkin on your machine.

  2. Register ZipkinTraceExporter in the application where you start Ignite:

    //register Zipkin exporter
    ZipkinTraceExporter.createAndRegister(
            ZipkinExporterConfiguration.builder().setV2Url("http://localhost:9411/api/v2/spans")
                    .setServiceName("ignite-cluster").build());
    
    IgniteConfiguration cfg = new IgniteConfiguration().setClientMode(true)
            .setTracingSpi(new org.apache.ignite.spi.tracing.opencensus.OpenCensusTracingSpi());
    
    Ignite ignite = Ignition.start(cfg);
    
    //enable trace sampling for transactions with 100% sampling rate
    ignite.tracingConfiguration().set(
            new TracingConfigurationCoordinates.Builder(Scope.TX).build(),
            new TracingConfigurationParameters.Builder().withSamplingRate(1).build());
    
    //create a transactional cache
    IgniteCache<Integer, String> cache = ignite
            .getOrCreateCache(new CacheConfiguration<Integer, String>("myCache")
                    .setAtomicityMode(CacheAtomicityMode.TRANSACTIONAL));
    
    IgniteTransactions transactions = ignite.transactions();
    
    // start a transaction
    try (Transaction tx = transactions.txStart()) {
        //do some operations
        cache.put(1, "test value");
    
        System.out.println(cache.get(1));
    
        cache.put(1, "second value");
    
        tx.commit();
    }
    
    try {
        //This code here is to wait until the trace is exported to Zipkin.
        //If your application doesn't stop here, you don't need this piece of code.
        Thread.sleep(5_000);
    } catch (InterruptedException e) {
        e.printStackTrace();
    }
  3. Open http://localhost:9411/zipkin in your browser and click the search icon.

    This is what a trace of the transaction looks like:

    trace in zipkin

Analyzing Trace Data

A trace is recorded information about the execution of a specific event. Each trace consists of a tree of spans. A span is an individual unit of work performed by the system in order to process the event.

Because of the distributed nature of Ignite, an operation usually involves multiple nodes. Therefore, a trace can include spans from multiple nodes. Each span always contains the information about the node where the corresponding operation was executed.

In the image of the transaction trace presented above, you can see that the trace contains the spans associated with the following operations:

  • acquire locks (transactions.colocated.lock.map),

  • get (transactions.near.enlist.read),

  • put (transactions.near.enlist.write),

  • commit (transactions.commit), and

  • close (transactions.close).

The commit operation, in turn, consists of two operations: prepare and finish.

You can click on each span to view the annotations and tags attached to it.

Span