brave

Java distributed tracing implementation compatible with Zipkin.

Zipkin is based on Dapper.

dapper (dutch) = brave (english)... so that's where the name comes from.

introduction

I advise you to read the Dapper paper, but in short:

What we want to achieve is understand system behavior and performance of complex distributed systems. We want to do this with minimal impact on existing code by introducing some small common libraries that are reusable and don't interfere with the existing business logic or architecture. Besides not impacting business logic or architecute we off course also want it to have a neglectable impact on performance.

You should use brave instead of Zipkin if:

• You can't use Finagle.
• You don't want to add Scala as a dependency to your Java project.
• You want out of the box integration support for RESTEasy, Jersey, Apache HttpClient.

Brave uses the Zipkin thrift generated classes as part of its api so it is easy to use existing Zipkin components with Brave (zipkin-collector, zipkin-query, zipkin-ui, cassandra store,...).

I'm very thankful to Twitter for open sourcing Zipkin! Is is by seeing their Zipkin video and presentation that I got to know Zipkin/Dapper and that I saw the potential and the simplicity of the solution.

about spans, traces and architecture

• span: A single client/server request/response. Can have an optional parent span id and is part of a trace.
• trace: A tree of spans.

So as you can see a single span is submitted twice:

• from the client side, the initiator, with cs (client send) and cr (client received) annotations
• from the server side with sr (server received) and ss (server send) annotations.

The above image shows how I intend to use Brave in production and also how it integrates with the Zipkin back-end components.

The applications and services have Brave integration at client and server side, for example through the RestEasy support module. They submit spans to the ZipkinSpanCollector which submits them to Flume.

I introduced Flume instead of Scribe as Flume is still actively maintained, easier to deploy, has good documentation and extensions are written in Java.

The ZipkinSpanCollectorSink, part of flume-zipkin-collector-sink module submits spans to the Zipkin collector service. The ZipkinMetricsSink, part of flume-zipkin-metrics-sink module submits custom annotations with duration (to measure certain sections of the code) to the Metrics library. Metrics builds histograms for the metrics and sends the data to a back-end system for storage and visualisation. Metrics supports multiple back-ends but the sink implementation today supports graphite.

If you use the ZooKeeperSamplingTraceFilter from brave-tracefilters module you can enable/disable tracing or adjust sample rate using ZooKeeper as also indicated on the drawing.

Example implementation

The brave-resteasy-example is a good starting point to get you up to speed on how you can implement brave in your own apps.

Maven artifacts

Version 2.3 is available in Maven central. So you can simply add the dependencies you want to your pom.xml. You will need at least:

<dependency>
    <groupId>com.github.kristofa</groupId>
    <artifactId>brave-impl</artifactId>
    <version>2.3</version>
</dependency>

For other dependencies see README.md files for sub modules.

Contribution

If you would like to contribute to brave here are some topics that might give you inspiration 😉

• Database tracing: Hibernate / Spring JDBC template support
• Similar as introduction of brave-client in version 2.2.1, introduce brave-server which adds an abstraction about ServerTracer which should remove code duplication and improve consistency between implementations.
• ca / sa annotation support: Submitting Client Address and Server Address annotations.
• Aggregation of data. Realtime (eg using Storm).
• Extend flume-zipkin-metrics-sink with support for other back-ends next to graphite.
• Improve Jersey support by updating Jersey to latest version and implements ContainerRequestFilter and ContainerResponseFilter iso current servlet filter.
• Support for other frameworks.

Feel free to submit pull requests.

Version history

Unfortunately I did not use Semantic Versioning from the start. The 2.x.x versions of brave are currently being extensively tested and will soon end up in production. However from now on I'll try to stick to Semantic Versioning when it comes to bug fixing / new functionality / backwards compatibility.

2.4-SNAPSHOT

• New feature: Add server side tracing support for Spring mvc based services. #42
• New feature: Add MySQL support (trace queries). #43
• Improvement: Don't enforce log4j to users of brave. Make dependency provided. #41
• New feature: Add Jersey 2 support #38
• New feature: Add new SpanNameFilter that uses pattern matching. #46

2.3

• Add basic building blocks for asynch client support. See #30.
• Bugfix for aligning service names in rest-easy implementation. See #32.
• Lower log messages from ERROR to WARN. See #36.
• Support for filtering span names with variable content. See #33.

2.2.1

• Introduce brave-client module that prevents code duplication and adds consistency to different brave integration implementations. brave-client is used by all client implementations (RESTEasy, Jersey, Apache HttpClient). See #27 and #29 by srapp.
• Improve compatibility with Zipkin. Hex encode trace id's. See #26 by klette and fix #28 by srapp.
• Bug fix in Jersey implementation by srapp . See #22
• Fix dependency issue. See #20 by K-jo.

2.2.0

• Bugfix in spring resteasy integration. Mentioned in issue #3. Logic to set span name resulted in different values for client vs server side. As a result span names were not consistent in zipkin-web.
• Bugfix for flume-zipkin-collector-sink, see #16. Bug in Flume Zipkin SpanCollectorSink resulted in lost spans. Thanks to leonly0224 for catching this.
• Bugfix for flume-zipkin-metrics-sink. Same issue as #16.
• Add brave-apache-http-interceptors module. Which adds Apache HttpClient support.
• Bugfix / improvement, see issues #15, #18, #19. Makes it possible to set Endpoint service name from ClientTracer. This means you can name a service yourself (eg when accessing cache, database,...). But it also means you can set same service name for client and server parts of span which result in better inspection in zipkin-web (see #18 ). Thanks to eirslett for implementing this. Both RestEasy and apache http client integration have been adapted to make use of this.
• Test flume-zipkin-collector-sink and flume-zipkin-metrics-sink with flume 1.5.0 + update libraries to 1.5.0
• Update flume-zipkin-collector-sink to add connection and socket timeout config entry.

2.1.1

• Bugfix: Make submitting annotations thread-safe. Add synchronisation to avoid ArrayIndexOutOfBoundsException or lost annotations.
• RestEasy BravePostProcessInterceptor clears span after 'server send' annotation is submitted. This prevents ThreadLocal classloader leak when used in Tomcat.
• Ryan Tenney : Make logger name configurable for LoggingSpanCollectorImpl.
• Ryan Tenney : Update slf4j to version 1.7.5 and use parameters in log statements iso string concatenation.

2.1.0

• Henrik Nordvik: Add Jersey support. See brave-jersey module.
• Add flume-zipkin-metrics module. See README.md of that module for more information.
• pettyjamesm: Added alternative Brave public api for use cases where the ThreadLocal approach is not working and you want to have more control over the context scope.
• Update brave-resteasy-spring module by adding BraveClientExecutionInterceptor. This interceptor can be used with the RestEasy client framework and makes the usage of Brave with RestEasy even easier as it now supports both server and client side.

2.0.2

• Done: ZipkinSpanCollector: Give the option to log error but not throw exception when connection with collector can't be set up during initialisation.
• Done: Make ZipkinSpanCollector more configurable. Allow configuration of queue size, batch size, nr of processing threads, socket time out.
• Done: Important bugfix in ZipkinSpanCollector. The processing thread(s) catch and log all exceptions to prevent they end prematurely.
• Done: Rename 'batchSize' config property of ZipkinSpanCollectorSink to 'batchsize' to bring it inline with other config parameters (all lower case).

2.0.1

• Done: Adapt ZipkinSpanCollector so it is able to reconnect to Scribe / Zipkin Collector in case of temporary network connectivity issues.
• Done: Adapt Flume ZipkinSpanCollectorSink so it is able to reconnect to Zipkin Collector in case of temporary network connectivity issues.
• Done: BraveCallable constructor is public.

2.0

Brave 1.0 was an alternative implementation of Dapper with integration with the Zipkin backend components. Brave 2.0 has the Zipkin core thrift classes as part of its api. This has as consequence that it will be easier to share components or extensions between Zipkin/Brave. It is compatible with Zipkin 1.1

Implemented changes:

• Done: Use Zipkin-core thrift generated classes as part of api.
• Done: Binary annotation support.
• Done: Cut dependencies with Twitter specific libraries. Only rely on Thrift.
• Done: Performance and through-put optimizations for zipkin-span-collector. It uses queue to store spans and separate thread for submitting spans to span collector / scribe. It tries to buffer spans to send them in batches to avoid communication overhead.
• Done: Make sampling / trace filtering work properly. See here for details.
• Done: Add flume support for transporting spans. See flume-zipkin-collector-sink.
• Done: Add TraceFilter implementation that uses Zookeeper for globally adjusting sample rate or enable/disable tracing all together. See brave-tracefilters
• Done: Extend threading support by introducing BraveExecutorService, BraveCallable and BraveRunnable.

1.0

• Functional but not usable for production usage, only for development and testing.