Tenacity

Tenacity is Dropwizard+Hystrix.

Dropwizard is a framework for building REST services. Hystrix is a resiliency library from Netflix and Ben Christensen.

Hystrix's goals are to:

Stop cascading failures.
Fail-fast and rapidly recover.
Reduce mean-time-to-discovery (with dashboards)
Reduce mean-time-to-recovery (with dynamic configuration)

Tenacity makes Hystrix dropwizard-friendly and for dropwizard-developers to quickly leverage the benefits of Hystrix.

Uses dropwizard-bundles for bootstrapping: property strategies, metrics, dynamic configuration, and some resource endpoints (e.g. for dashboards).
Dropwizard-configuration style (YAML) for dependencies.
Abstractions to clearly configure a dependency operation (TenacityCommand<ReturnType>).
Ability to unit-test Hystrix: Resets static state held by Hystrix (metrics, counters, etc.). Increases rate at which a concurrent thread updates metrics.
Publishes measurements via Metrics.

Tenacity is meant to be used with Breakerbox which adds real-time visualization of metrics and dynamic configuration.

Modules

tenacity-core: The building blocks to quickly use Hystrix within the context of Dropwizard.
tenacity-client: Client for consuming the resources that tenacity-core adds.
tenacity-testing: TenacityTestRule allows for easier unit testing. Resets internal state of Hystrix.
tenacity-jdbi: Pulls in dropwizard-jdbi and provides a DBIExceptionLogger and SQLExceptionLogger to be used with the ExceptionLoggingCommandHook.

How To Use

Here is a sample TenacityCommand that always succeeds:

public class AlwaysSucceed extends TenacityCommand<String> {
    public AlwaysSucceed() {
        super(DependencyKey.ALWAYS_SUCCEED);
    }

    @Override
    protected String run() throws Exception {
        return "value";
    }
}

A quick primer on the way to use a TenacityCommand if you are not familiar with Hystrix's execution model:

Synchronous Execution

AlwaysSucceed command = new AlwaysSucceed();
String result = command.execute();

This executes the command synchronously but through the protection of a Future.get(configurableTimeout).

Asynchronous Execution

Future<String> futureResult = command.queue();

Reactive Execution

Observable<String> observable = new AlwaysSucceed().observe();

Fallbacks

When execution fails, it is possible to gracefully degrade with the use of fallbacks.

Execution Flow

TenacityCommand Constructor Arguments

Earlier we saw:

public class AlwaysSucceed extends TenacityCommand<String> {
    public AlwaysSucceed() {
        super(DependencyKey.ALWAYS_SUCCEED);
    }
    ...
}

The arguments are:

commandKey: This creates a circuit-breaker, threadpool, and also the identifier that will be used in dashboards. This should be your implementation of the TenacityPropertyKey interface.

It is possible to create multiple circuit-breakers that leverage a single threadpool, but for simplicity we are not allowing that type of configuration.

How to add Tenacity to your Dropwizard Service

To leverage within dropwizard first add the following to your pom.xml:

<dependency>
    <groupId>com.yammer.tenacity</groupId>
    <artifactId>tenacity-core</artifactId>
    <version>1.1.2</version>
</dependency>

Or you can leverage the tenacity-bom:

<dependencyManagement>
    <dependencies>
        <dependency>
            <groupId>com.yammer.tenacity</groupId>
            <artifactId>tenacity-bom</artifactId>
            <version>1.1.2</version>
            <type>pom</type>
            <scope>import</scope>
        </dependency>
    </dependencies>
</dependencyManagement>

<dependency>
    <groupId>com.yammer.tenacity</groupId>
    <artifactId>tenacity-core</artifactId>
</dependency>

Enumerate your dependencies. These will eventually be used as global identifiers in dashboards. We have found that it works best for us when you include the service and the external dependency at a minimum. Here is an example of completie's dependencies. Note we also shave down some characters to save on space, again for UI purposes. In addition, you'll need to have an implementation of a TenacityPropertyKeyFactory which you can see an example of below.

public enum CompletieDependencyKeys implements TenacityPropertyKey {
    CMPLT_PRNK_USER, CMPLT_PRNK_GROUP, CMPLT_PRNK_SCND_ORDER, CMPLT_PRNK_NETWORK,
    CMPLT_TOKIE_AUTH,
    CMPLT_TYRANT_AUTH,
    CMPLT_WHVLL_PRESENCE
}

public class CompletieDependencyKeyFactory implements TenacityPropertyKeyFactory {
    @Override
    public TenacityPropertyKey from(String value) {
        return CompletieDependencyKeys.valueOf(value.toUpperCase());
    }
}

Create a TenacityBundleConfigurationFactory implementation - you can use the BaseTenacityBundleConfigurationFactory as your starting point. This will be used to register your custom tenacity dependencies and custom configurations.

public class CompletieTenacityBundleConfigurationFactory extends BaseTenacityBundleConfigurationFactory<CompletieConfiguration> {

  @Override
  public Map<TenacityPropertyKey, TenacityConfiguration> getTenacityConfigurations(CompletieConfiguration configuration) {
        final ImmutableMap.Builder<TenacityPropertyKey, TenacityConfiguration> builder = ImmutableMap.builder();

        builder.put(CompletieDependencyKeys.CMPLT_PRNK_USER, configuration.getRanking().getHystrixUserConfig());
        builder.put(CompletieDependencyKeys.CMPLT_PRNK_GROUP, configuration.getRanking().getHystrixGroupConfig());
        builder.put(CompletieDependencyKeys.CMPLT_PRNK_SCND_ORDER, configuration.getRanking().getHystrixSecondOrderConfig());
        builder.put(CompletieDependencyKeys.CMPLT_PRNK_NETWORK, configuration.getRanking().getHystrixNetworkConfig());
        builder.put(CompletieDependencyKeys.CMPLT_TOKIE_AUTH, configuration.getAuthentication().getHystrixConfig());
        builder.put(CompletieDependencyKeys.CMPLT_WHVLL_PRESENCE, configuration.getPresence().getHystrixConfig())

        return builder.build();
  }
}

Then make sure you add the bundle in your Application.

Map<TenacityPropertyKey, TenacityConfiguration> type.

@Override
public void initialize(Bootstrap<MyConfiguration> bootstrap) {
    ...
    bootstrap.addBundle(TenacityBundleBuilder
                                        .<MyConfiguration> newBuilder()
                                        .configurationFactory(new CompletieTenacityBundleConfigurationFactory())
                                        .build());
    ...
}

Use TenacityCommand to select which custom tenacity configuration you want to use.

public class CompletieDependencyOnTokie extends TenacityCommand<String> {
    public CompletieDependencyOnTokie() {
        super(CompletieDependencyKeys.CMPLT_TOKIE_AUTH);
    }
    ...
}

When testing use the tenacity-testing module. This registers appropriate custom publishers/strategies, clears global Archaius configuration state (Hystrix uses internally to manage configuration), and tweaks threads that calculate metrics which influence circuit breakers to update a more frequent interval. Simply use the TenacityTestRule.
```
@Rule
public final TenacityTestRule tenacityTestRule = new TenacityTestRule();
```
```
<dependency>
    <groupId>com.yammer.tenacity</groupId>
    <artifactId>tenacity-testing</artifactId>
    <version>1.1.2</version>
    <scope>test</scope>
</dependency>
```
Last is to actually configure your dependencies once they are wrapped with TenacityCommand.

Configuration

Once you have identified your dependencies you need to configure them appropriately. Here is the basic structure of a single TenacityConfiguration that may be leverage multiple times through your service configuration:

Defaults

executionIsolationThreadTimeoutInMillis: 1000
executionIsolationStrategy: THREAD
threadpool:
    threadPoolCoreSize: 10
    keepAliveTimeMinutes: 1
    maxQueueSize: -1
    queueSizeRejectionThreshold: 5
    metricsRollingStatisticalWindowInMilliseconds: 10000
    metricsRollingStatisticalWindowBuckets: 10
circuitBreaker:
    requestVolumeThreshold: 20
    errorThresholdPercentage: 50
    sleepWindowInMillis: 5000
    metricsRollingStatisticalWindowInMilliseconds: 10000
    metricsRollingStatisticalWindowBuckets: 10
semaphore:
    maxConcurrentRequests: 10
    fallbackMaxConcurrentRequests: 10

The following two are the most important and you can probably get by just fine by defining just these two and leveraging the defaults.

executionIsolationStrategy: Which to use THREAD or SEMAPHORE. Defaults to THREAD. Execution Isolation Strategy.
executionIsolationThreadTimeoutInMillis: How long the entire dependency command should take.
threadPoolCoreSize: Self explanatory.

Here are the rest of the descriptions:

keepAliveTimeMinutes: Thread keepAlive time in the thread pool.
maxQueueSize: -1 uses a SynchronousQueue. Anything >0 leverages a BlockingQueue and enables the queueSizeRejectionThreshold variable.
queueSizeRejectionThreshold: Disabled when using -1 for maxQueueSize otherwise self explanatory.
requestVolumeThreshold: The minimum number of requests that need to be received within the metricsRollingStatisticalWindowInMilliseconds in order to open a circuit breaker.
errorThresholdPercentage: The percentage of errors needed to trip a circuit breaker. In order for this to take effect the requestVolumeThreshold must first be satisfied.
sleepWindowInMillis: How long to keep the circuit breaker open, before trying again.

Here are the semaphore related items:

maxConcurrentRequests: The number of concurrent requests for a given key at any given time.
fallbackMaxConcurrentRequests: The number of concurrent requests for the fallback at any given time.

These are recommended to be left alone unless you know what you're doing:

metricsRollingStatisticalWindowInMilliseconds: How long to keep around metrics for calculating rates.
metricsRollingStatisticalWindowBuckets: How many different metric windows to keep in memory.

Once you are done configuring your Tenacity dependencies. Don't forget to tweak the necessary connect/read timeouts on HTTP clients. We have some suggestions for how you go about this in the Equations section.

Breakerbox

One of the great things about Tenacity is the ability to aid in the reduction of mean-time-to-discovery and mean-time-to-recovery for issues. This is available through a separate service Breakerbox.

Breakerbox is a central dashboard and an on-the-fly configuration tool for Tenacity. In addition to the per-tenacity-command configurations shown above this configuration piece let's you define where and how often to check for newer configurations.

breakerbox:
  urls: http://breakerbox.yourcompany.com:8080/archaius/{service}
  initialDelay: 0s
  delay: 60s
  waitForInitialLoad: 0s

urls is a list of comma-deliminated list of urls for where to pull tenacity configurations. This will pull override configurations for all dependency keys for requested service.
initialDelay how long before the first poll for newer configuration executes.
delay the ongoing schedule to poll for newer configurations.
waitForInitialLoad is the amount of time to block Dropwizard from starting while waiting for Breakerbox configurations.

In order to add integration with Breakerbox you need to implement the following method in your TenacityBundleConfigurationFactory implementation:

@Override
public BreakerboxConfiguration getBreakerboxConfiguration(CompletieConfiguration configuration) {
   return configuration.getBreakerbox();
}

Configuration Hierarchy Order

Configurations can happen in a lot of different spots so it's good to just spell it out clearly. The order in this list matters, the earlier items override those that come later.

Breakerbox
Local service configuration YAML
Defaults

Configuration Equations

How to configure your dependent services can be confusing. A good place to start if don't have a predefined SLA is to just look at actual measurements. At Yammer, we set our max operational time for our actions somewhere between p99 and p999 for response times. We do this because we have found it to actually be faster to fail those requests, retry, and optimistically get a p50 response time.

Tenacity

executionIsolationThreadTimeoutInMillis = p99 + median + extra
p99 < executionIsolationThreadTimeoutInMillis < p999
threadpool
- size = (p99 in seconds) * (m1 rate req/sec) + extra
- 10 is usually fine for most operations. Anything with a large pool should be understood why that is necessary (e.g. long response times)
- Extra: this number only meets the current traffic needs. Make sure to add some extra for bursts as well as growth.

HTTP client

connectTimeout = 33% of executionIsolationThreadTimeoutInMillis
timeout (readTimeout) = 110% of executionIsolationThreadTimeoutInMillis
- We put this timeout higher so that it doesn't raise a TimeoutException from the HTTP client, but instead from Hystrix.

Note: These are just suggestions, feel free to look at Hystrix's configuration documentation, or implement your own.

Resources

Tenacity adds resources under /tenacity:

GET /tenacity/propertykeys: List of strings which are all the registered propertykeys with Tenacity.
GET /tenacity/configuration/{key}: JSON representation of a TenacityConfiguration for the supplied {key}.
GET /tenacity/circuitbreakers: Simple JSON representation of all circuitbreakers and their circuitbreaker status. GET /tenacity/circuitbreakers/{key}: Single circuitbreaker status PUT /tenacity/circuitbreakers/{key}: Expected "FORCED_CLOSED, FORCED_OPEN, or FORCED_RESET" as the body.
GET /tenacity/metrics.stream: text/event-stream of Hystrix metrics.

By default these are put onto the main application port. If you want to place these instead on the admin port, you can configure this when building the Tenacity bundle.

TenacityBundleBuilder
                .<MyConfiguration> newBuilder()
                ...
                .usingAdminPort()
                .build();

TenacityExceptionMapper

An exception mapper exists to serve as an aid for unhandled HystrixRuntimeExceptions. It is used to convert all the types of unhandled exceptions to be converted to a simple HTTP status code. A common pattern here is to convert the unhandled HystrixRuntimeExceptions to 429 Too Many Requests:

TenacityBundleBuilder
                .<MyConfiguration> newBuilder()
                .configurationFactory(configurationFactory)
                .mapAllHystrixRuntimeExceptionsTo(429)
                .build();

ExceptionLoggingCommandHook

If you don't handle logging exceptions explicitly within each TenacityCommand, you can easily miss problems or at-least find them very hard to debug. Instead you can add the ExceptionLoggingCommandHook to the TenacityBundle and register ExceptionLoggers to handle the logging of different kinds of Exceptions. The ExecutionLoggingCommandHook acts as a HystrixCommandExecutionHook and intercepts all Exceptions that occur during the run() method of your TenacityCommands. By sequencing ExceptionLoggers from most specific to most general, the ExceptionLoggingCommandHook will be able to find the best ExceptionLogger for the type of Exception.

TenacityBundleBuilder
                .<MyConfiguration> newBuilder()
                .configurationFactory(configurationFactory)
                .mapAllHystrixRuntimeExceptionsTo(429)
                .commandExecutionHook(new ExceptionLoggingCommandHook(
                    new DBIExceptionLogger(registry),
                    new SQLExceptionLogger(registry),
                    new DefaultExceptionLogger()
                ))
                .build();

TenacityJerseyClientBuilder

TenacityJerseyClient and TenacityJerseyClientBuilder to reduce configuration complexity when using Tenacity and JerseyClient. At the moment these two have competing timeout configurations that can end up looking like application exceptions when they are simply TimeoutExceptions being thrown by JerseyClient. TenacityJerseyClient aims to fix this by adjusting the socket read timeout on a per-request basis on the currently set execution timeout value for resources built from TenacityJerseyClient and its associated TenacityPropertyKey. Note: Metrics associated with the TenacityPropertyKey are NOT updated whenever the underlying Client is used. The TenacityPropertyKey metrics are only ever updated when using TenacityCommand or TenacityObservableCommand at the moment.

Client client = new JerseyClientBuilder(environment).build("some-external-dependency");
Client tenacityClient = TenacityJerseyClientBuilder
    .builder(YOUR_TENACITY_PROPERTY_KEY)
    .usingTimeoutPadding(Duration.milliseconds(50))
        //Padding to add in addition to the Tenacity set time. Default is 50ms.
        //Result: TenacityTimeout + TimeoutPadding = SocketReadTimeout
    .build(client);

//Then use tenacityClient the same way as you'd use client. TenacityClient overrides resource/asyncResource and those in turn are Tenacity*Resources.
//They adjust timeouts on every use or on a per-request basis.

TenacityCircuitBreakerHealthCheck

There is now the ability to add a HealthCheck which returns unhealthy when any circuit is open. This could be because a circuit is forced open or because of environment circumstances. The default is for this not to be turned on. You can enable this HealthCheck by configuring it on the bundle that is added to your application.

TenacityBundleBuilder
                .newBuilder()
                ...
                .withCircuitBreakerHealthCheck()
                .build();

TenacityCommand.Builder

Java8 brings functional interfaces and TenacityCommand/TenacityObservableCommand offers support.

TenacityCommand
      .builder(DependencyKey.GENERAL)
      .run(() -> 1)
      .fallback(() -> 2)
      .execute()

TenacityObservableCommand
      .builder(DependencyKey.GENERAL)
      .run(() -> Observable.just(1))
      .fallback(() -> Observable.just(2))
      .observe()

chrisgray/tenacity