Resilience4j
简介
Resilience4j是一款轻量级,易于使用的容错库,其灵感来自于Netflix Hystrix,但是专为Java 8和函数式编程而设计。轻量级,因为库只使用了Vavr,它没有任何其他外部依赖下。相比之下,Netflix Hystrix对Archaius具有编译依赖性,Archaius具有更多的外部库依赖性,例如Guava和Apache Commons Configuration。
要使用Resilience4j,不需要引入所有依赖,只需要选择你需要的。
Resilience4j提供了以下的核心模块和拓展模块:
核心模块:
- resilience4j-circuitbreaker: Circuit breaking
- resilience4j-ratelimiter: Rate limiting
- resilience4j-bulkhead: Bulkheading
- resilience4j-retry: Automatic retrying (sync and async)
- resilience4j-cache: Result caching
- resilience4j-timelimiter: Timeout handling
Circuitbreaker
简介
CircuitBreaker通过具有三种正常状态的有限状态机实现:CLOSED,OPEN和HALF_OPEN以及两个特殊状态DISABLED和FORCED_OPEN。当熔断器关闭时,所有的请求都会通过熔断器。如果失败率超过设定的阈值,熔断器就会从关闭状态转换到打开状态,这时所有的请求都会被拒绝。当经过一段时间后,熔断器会从打开状态转换到半开状态,这时仅有一定数量的请求会被放入,并重新计算失败率,如果失败率超过阈值,则变为打开状态,如果失败率低于阈值,则变为关闭状态。
Resilience4j记录请求状态的数据结构和Hystrix不同,Hystrix是使用滑动窗口来进行存储的,而Resilience4j采用的是Ring Bit Buffer(环形缓冲区)。Ring Bit Buffer在内部使用BitSet这样的数据结构来进行存储,BitSet的结构如下图所示:
每一次请求的成功或失败状态只占用一个bit位,与boolean数组相比更节省内存。BitSet使用long[]数组来存储这些数据,意味着16个值(64bit)的数组可以存储1024个调用状态。
计算失败率需要填满环形缓冲区。例如,如果环形缓冲区的大小为10,则必须至少请求满10次,才会进行故障率的计算,如果仅仅请求了9次,即使9个请求都失败,熔断器也不会打开。但是CLOSE状态下的缓冲区大小设置为10并不意味着只会进入10个 请求,在熔断器打开之前的所有请求都会被放入。
当故障率高于设定的阈值时,熔断器状态会从由CLOSE变为OPEN。这时所有的请求都会抛出CallNotPermittedException异常。当经过一段时间后,熔断器的状态会从OPEN变为HALF_OPEN,HALF_OPEN状态下同样会有一个Ring Bit Buffer,用来计算HALF_OPEN状态下的故障率,如果高于配置的阈值,会转换为OPEN,低于阈值则装换为CLOSE。与CLOSE状态下的缓冲区不同的地方在于,HALF_OPEN状态下的缓冲区大小会限制请求数,只有缓冲区大小的请求数会被放入。
除此以外,熔断器还会有两种特殊状态:DISABLED(始终允许访问)和FORCED_OPEN(始终拒绝访问)。这两个状态不会生成熔断器事件(除状态装换外),并且不会记录事件的成功或者失败。退出这两个状态的唯一方法是触发状态转换或者重置熔断器。
熔断器关于线程安全的保证措施有以下几个部分:
- 熔断器的状态使用AtomicReference保存的
- 更新熔断器状态是通过无状态的函数或者原子操作进行的
- 更新事件的状态用synchronized关键字保护
意味着同一时间只有一个线程能够修改熔断器状态或者记录事件的状态。
可配置参数
配置参数 | 默认值 | 描述 |
---|---|---|
failureRateThreshold | 50 | 熔断器关闭状态和半开状态使用的同一个失败率阈值 |
ringBufferSizeInHalfOpenState | 10 | 熔断器半开状态的缓冲区大小,会限制线程的并发量,例如缓冲区为10则每次只会允许10个请求调用后端服务 |
ringBufferSizeInClosedState | 100 | 熔断器关闭状态的缓冲区大小,不会限制线程的并发量,在熔断器发生状态转换前所有请求都会调用后端服务 |
waitDurationInOpenState | 60(s) | 熔断器从打开状态转变为半开状态等待的时间 |
automaticTransitionFromOpenToHalfOpenEnabled | false | 如果置为true,当等待时间结束会自动由打开变为半开,若置为false,则需要一个请求进入来触发熔断器状态转换 |
recordExceptions | empty | 需要记录为失败的异常列表 |
ignoreExceptions | empty | 需要忽略的异常列表 |
recordFailure | throwable -> true | 自定义的谓词逻辑用于判断异常是否需要记录或者需要忽略,默认所有异常都进行记录 |
测试前准备
pom.xml
测试使用的IDE为idea,使用的springboot进行学习测试,首先引入maven依赖:
io.github.resilience4j
resilience4j-spring-boot
0.9.0
resilience4j-spring-boot集成了circuitbeaker、retry、bulkhead、ratelimiter几个模块,因为后续还要学习其他模块,就直接引入resilience4j-spring-boot依赖。
application.yml配置
resilience4j:
circuitbreaker:
configs:
default:
ringBufferSizeInClosedState: 5 # 熔断器关闭时的缓冲区大小
ringBufferSizeInHalfOpenState: 2 # 熔断器半开时的缓冲区大小
waitDurationInOpenState: 10000 # 熔断器从打开到半开需要的时间
failureRateThreshold: 60 # 熔断器打开的失败阈值
eventConsumerBufferSize: 10 # 事件缓冲区大小
registerHealthIndicator: true # 健康监测
automaticTransitionFromOpenToHalfOpenEnabled: false # 是否自动从打开到半开,不需要触发
recordFailurePredicate: com.example.resilience4j.exceptions.RecordFailurePredicate # 谓词设置异常是否为失败
recordExceptions: # 记录的异常
- com.example.resilience4j.exceptions.BusinessBException
- com.example.resilience4j.exceptions.BusinessAException
ignoreExceptions: # 忽略的异常
- com.example.resilience4j.exceptions.BusinessAException
instances:
backendA:
baseConfig: default
waitDurationInOpenState: 5000
failureRateThreshold: 20
backendB:
baseConfig: default
可以配置多个熔断器实例,使用不同配置或者覆盖配置。
需要保护的后端服务
以一个查找用户列表的后端服务为例,利用熔断器保护该服务。
interface RemoteService {
List process() throws TimeoutException, InterruptedException;
}
连接器调用该服务
这是调用远端服务的连接器,我们通过调用连接器中的方法来调用后端服务。
public RemoteServiceConnector{
public List process() throws TimeoutException, InterruptedException {
List users;
users = remoteServic.process();
return users;
}
}
用于监控熔断器状态及事件的工具类
要想学习各个配置项的作用,需要获取特定时候的熔断器状态,写一个工具类:
@Log4j2
public class CircuitBreakerUtil {
/**
* @Description: 获取熔断器的状态
*/
public static void getCircuitBreakerStatus(String time, CircuitBreaker circuitBreaker){
CircuitBreaker.Metrics metrics = circuitBreaker.getMetrics();
// Returns the failure rate in percentage.
float failureRate = metrics.getFailureRate();
// Returns the current number of buffered calls.
int bufferedCalls = metrics.getNumberOfBufferedCalls();
// Returns the current number of failed calls.
int failedCalls = metrics.getNumberOfFailedCalls();
// Returns the current number of successed calls.
int successCalls = metrics.getNumberOfSuccessfulCalls();
// Returns the max number of buffered calls.
int maxBufferCalls = metrics.getMaxNumberOfBufferedCalls();
// Returns the current number of not permitted calls.
long notPermittedCalls = metrics.getNumberOfNotPermittedCalls();
log.info(time + "state=" +circuitBreaker.getState() + " , metrics[ failureRate=" + failureRate +
", bufferedCalls=" + bufferedCalls +
", failedCalls=" + failedCalls +
", successCalls=" + successCalls +
", maxBufferCalls=" + maxBufferCalls +
", notPermittedCalls=" + notPermittedCalls +
" ]"
);
}
/**
* @Description: 监听熔断器事件
*/
public static void addCircuitBreakerListener(CircuitBreaker circuitBreaker){
circuitBreaker.getEventPublisher()
.onSuccess(event -> log.info("服务调用成功:" + event.toString()))
.onError(event -> log.info("服务调用失败:" + event.toString()))
.onIgnoredError(event -> log.info("服务调用失败,但异常被忽略:" + event.toString()))
.onReset(event -> log.info("熔断器重置:" + event.toString()))
.onStateTransition(event -> log.info("熔断器状态改变:" + event.toString()))
.onCallNotPermitted(event -> log.info(" 熔断器已经打开:" + event.toString()))
;
}
调用方法
CircuitBreaker目前支持两种方式调用,一种是程序式调用,一种是AOP使用注解的方式调用。
程序式的调用方法
在CircuitService中先注入注册器,然后用注册器通过熔断器名称获取熔断器。如果不需要使用降级函数,可以直接调用熔断器的executeSupplier方法或executeCheckedSupplier方法:
public class CircuitBreakerServiceImpl{
@Autowired
private CircuitBreakerRegistry circuitBreakerRegistry;
public List circuitBreakerNotAOP() throws Throwable {
CircuitBreaker circuitBreaker = circuitBreakerRegistry.circuitBreaker("backendA");
CircuitBreakerUtil.getCircuitBreakerStatus("执行开始前:", circuitBreaker);
circuitBreaker.executeCheckedSupplier(remotServiceConnector::process);
}
}
如果需要使用降级函数,则要使用decorate包装服务的方法,再使用Try.of().recover()进行降级处理,同时也可以根据不同的异常使用不同的降级方法:
public class CircuitBreakerServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private CircuitBreakerRegistry circuitBreakerRegistry;
public List circuitBreakerNotAOP(){
// 通过注册器获取熔断器的实例
CircuitBreaker circuitBreaker = circuitBreakerRegistry.circuitBreaker("backendA");
CircuitBreakerUtil.getCircuitBreakerStatus("执行开始前:", circuitBreaker);
// 使用熔断器包装连接器的方法
CheckedFunction0> checkedSupplier = CircuitBreaker.
decorateCheckedSupplier(circuitBreaker, remoteServiceConnector::process);
// 使用Try.of().recover()调用并进行降级处理
Try> result = Try.of(checkedSupplier).
recover(CallNotPermittedException.class, throwable -> {
log.info("熔断器已经打开,拒绝访问被保护方法~");
CircuitBreakerUtil
.getCircuitBreakerStatus("熔断器打开中:", circuitBreaker);
List users = new ArrayList();
return users;
})
.recover(throwable -> {
log.info(throwable.getLocalizedMessage() + ",方法被降级了~~");
CircuitBreakerUtil
.getCircuitBreakerStatus("降级方法中:",circuitBreaker);
List users = new ArrayList();
return users;
});
CircuitBreakerUtil.getCircuitBreakerStatus("执行结束后:", circuitBreaker);
return result.get();
}
}
AOP式的调用方法
首先在连接器方法上使用@CircuitBreaker(name="",fallbackMethod="")注解,其中name是要使用的熔断器的名称,fallbackMethod是要使用的降级方法,降级方法必须和原方法放在同一个类中,且降级方法的返回值需要和原方法相同,输入参数需要添加额外的exception参数,类似这样:
public RemoteServiceConnector{
@CircuitBreaker(name = "backendA", fallbackMethod = "fallBack")
public List process() throws TimeoutException, InterruptedException {
List users;
users = remoteServic.process();
return users;
}
private List fallBack(Throwable throwable){
log.info(throwable.getLocalizedMessage() + ",方法被降级了~~");
CircuitBreakerUtil.getCircuitBreakerStatus("降级方法中:", circuitBreakerRegistry.circuitBreaker("backendA"));
List users = new ArrayList();
return users;
}
private List fallBack(CallNotPermittedException e){
log.info("熔断器已经打开,拒绝访问被保护方法~");
CircuitBreakerUtil.getCircuitBreakerStatus("熔断器打开中:", circuitBreakerRegistry.circuitBreaker("backendA"));
List users = new ArrayList();
return users;
}
}
可使用多个降级方法,保持方法名相同,同时满足的条件的降级方法会触发最接近的一个(这里的接近是指类型的接近,先会触发离它最近的子类异常),例如如果process()方法抛出CallNotPermittedException,将会触发fallBack(CallNotPermittedException e)方法而不会触发fallBack(Throwable throwable)方法。
之后直接调用方法就可以了:
public class CircuitBreakerServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private CircuitBreakerRegistry circuitBreakerRegistry;
public List circuitBreakerAOP() throws TimeoutException, InterruptedException {
CircuitBreakerUtil
.getCircuitBreakerStatus("执行开始前:",circuitBreakerRegistry.circuitBreaker("backendA"));
List result = remoteServiceConnector.process();
CircuitBreakerUtil
.getCircuitBreakerStatus("执行结束后:", circuitBreakerRegistry.circuitBreaker("backendA"));
return result;
}
}
使用测试
接下来进入测试,首先我们定义了两个异常,异常A同时在黑白名单中,异常B只在黑名单中:
recordExceptions: # 记录的异常
- com.example.resilience4j.exceptions.BusinessBException
- com.example.resilience4j.exceptions.BusinessAException
ignoreExceptions: # 忽略的异常
- com.example.resilience4j.exceptions.BusinessAException
然后对被保护的后端接口进行如下的实现:
public class RemoteServiceImpl implements RemoteService {
private static AtomicInteger count = new AtomicInteger(0);
public List process() {
int num = count.getAndIncrement();
log.info("count的值 = " + num);
if (num % 4 == 1){
throw new BusinessAException("异常A,不需要被记录");
}
if (num % 4 == 2 || num % 4 == 3){
throw new BusinessBException("异常B,需要被记录");
}
log.info("服务正常运行,获取用户列表");
// 模拟数据库的正常查询
return repository.findAll();
}
}
使用CircuitBreakerServiceImpl中的AOP或者程序式调用方法进行单元测试,循环调用10次:
public class CircuitBreakerServiceImplTest{
@Autowired
private CircuitBreakerServiceImpl circuitService;
@Test
public void circuitBreakerTest() {
for (int i=0; i<10; i++){
// circuitService.circuitBreakerAOP();
circuitService.circuitBreakerNotAOP();
}
}
}
看下运行结果:
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=0, failedCalls=0, successCalls=0, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 0
服务正常运行,获取用户列表
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1,
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 1
异常A,不需要被记录,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 2
异常B,需要被记录,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=1, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=1, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=1, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 3
异常B,需要被记录,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=2, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=2, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=2, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 4
服务正常运行,获取用户列表
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=4, failedCalls=2, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=4, failedCalls=2, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 5
异常A,不需要被记录,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=4, failedCalls=2, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=4, failedCalls=2, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=4, failedCalls=2, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 6
异常B,需要被记录,方法被降级了~~
降级方法中:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
熔断器已经打开,拒绝访问被保护方法~
熔断器打开中:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=1 ]
执行结束后:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=1 ]
注意到异常A发生的前后bufferedCalls、failedCalls、successCalls三个参数的值都没有没有发生变化,说明白名单的优先级高于黑名单,源码中也有提到Ignoring an exception has priority over recording an exception:
/**
* @see #ignoreExceptions(Class[]) ). Ignoring an exception has priority over recording an exception.
*
* Example:
* recordExceptions(Throwable.class) and ignoreExceptions(RuntimeException.class)
* would capture all Errors and checked Exceptions, and ignore unchecked
*
*/
同时也可以看出白名单所谓的忽略,是指不计入缓冲区中(即不算成功也不算失败),有降级方法会调用降级方法,没有降级方法会抛出异常,和其他异常无异。
执行开始前:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
熔断器已经打开,拒绝访问被保护方法~
熔断器打开中:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=1 ]
执行结束后:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=1 ]
当环形缓冲区大小被填满时会计算失败率,这时请求会被拒绝获取不到count的值,且notPermittedCalls会增加。
接下来我们实验一下多线程下熔断器关闭和熔断器半开两种情况下缓冲环的区别,我们先开15个线程进行调用测试熔断器关闭时的缓冲环,熔断之后等10s再开15个线程进行调用测试熔断器半开时的缓冲环:
public class CircuitBreakerServiceImplTest{
@Autowired
private CircuitBreakerServiceImpl circuitService;
@Test
public void circuitBreakerThreadTest() throws InterruptedException {
ExecutorService pool = Executors.newCachedThreadPool();
for (int i=0; i<15; i++){
pool.submit(
// circuitService::circuitBreakerAOP
circuitService::circuitBreakerNotAOP);
}
pool.shutdown();
while (!pool.isTerminated());
Thread.sleep(10000);
log.info("熔断器状态已转为半开");
pool = Executors.newCachedThreadPool();
for (int i=0; i<15; i++){
pool.submit(
// circuitService::circuitBreakerAOP
circuitService::circuitBreakerNotAOP);
}
pool.shutdown();
while (!pool.isTerminated());
for (int i=0; i<10; i++){
}
}
}
前15个线程都通过了熔断器,由于正常返回需要查数据库,所以会慢很多,失败率很快就达到了100%,而且观察到如下的记录:
异常B,需要被记录,方法被降级了~~
降级方法中:state=OPEN , metrics[ failureRate=100.0, bufferedCalls=5, failedCalls=5, successCalls=0, maxBufferCalls=5, notPermittedCalls=0 ]
可以看出,虽然熔断器已经打开了,可是异常B还是进入了降级方法,抛出的异常不是notPermittedCalls数量为0,说明在熔断器转换成打开之前所有请求都通过了熔断器,缓冲环不会控制线程的并发。
执行结束后:state=OPEN , metrics[ failureRate=80.0, bufferedCalls=5, failedCalls=4, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=OPEN , metrics[ failureRate=60.0, bufferedCalls=5, failedCalls=3, successCalls=2, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=OPEN , metrics[ failureRate=40.0, bufferedCalls=5, failedCalls=2, successCalls=3, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=OPEN , metrics[ failureRate=20.0, bufferedCalls=5, failedCalls=1, successCalls=4, maxBufferCalls=5, notPermittedCalls=0 ]
同时以上几条正常执行的服务完成后,熔断器的失败率在下降,说明熔断器打开状态下还是会计算失败率,由于环形缓冲区大小为5,初步推断成功的状态会依次覆盖最开始的几个状态,所以得到了上述结果。
接下来分析后15个线程的结果
熔断器状态已转为半开
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 16
服务正常运行,获取用户列表
执行开始前:state=OPEN , metrics[ failureRate=0.0, bufferedCalls=5, failedCalls=0, successCalls=5, maxBufferCalls=5, notPermittedCalls=0 ]
熔断器状态改变:2019-07-29T17:19:19.959+08:00[Asia/Shanghai]: CircuitBreaker 'backendA' changed state from OPEN to HALF_OPEN
count的值 = 18
count的值 = 17
服务正常运行,获取用户列表
count的值 = 19
count的值 = 15
熔断器状态从打开到半开我设置的是5s,前15个线程调用之后我等待了10s,熔断器应该已经变为半开了,但是执行开始前熔断器的状态却是OPEN,这是因为默认的配置项automaticTransitionFromOpenToHalfOpenEnabled=false,时间到了也不会自动转换,需要有新的请求来触发熔断器的状态转换。同时我们发现,好像状态改变后还是进了超过4个请求,似乎半开状态的环并不能限制线程数?这是由于这些进程是在熔断器打开时一起进来的。为了更好的观察环半开时候环大小是否限制线程数,我们修改一下配置:
resilience4j:
circuitbreaker:
configs:
myDefault:
automaticTransitionFromOpenToHalfOpenEnabled: true # 是否自动从打开到半开
我们再试一次:
熔断器状态已转为半开
执行开始前:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=0, failedCalls=0, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
执行开始前:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=0, failedCalls=0, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
执行开始前:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=0, failedCalls=0, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
count的值 = 15
count的值 = 16
服务正常运行,获取用户列表
异常B,需要被记录,方法被降级了~~
降级方法中:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=1, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
执行结束后:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=1, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
count的值 = 17
异常A,不需要被记录,方法被降级了~~
降级方法中:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=2, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
执行开始前:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=2, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
count的值 = 18
执行开始前:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=2, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
异常B,需要被记录,方法被降级了~~
降级方法中:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=3, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
执行结束后:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=3, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
熔断器已经打开:2019-07-29T17:36:14.189+08:00[Asia/Shanghai]: CircuitBreaker 'backendA' recorded a call which was not permitted.
执行开始前:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=2, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
执行结束后:state=HALF_OPEN , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=2, successCalls=0, maxBufferCalls=4, notPermittedCalls=0 ]
熔断器已经打开,拒绝访问被保护方法~
结果只有4个请求进去了,可以看出虽然熔断器状态还是半开,但是已经熔断了,说明在半开状态下,超过环大小的请求会被直接拒绝。
综上,circuitbreaker的机制已经被证实,且十分清晰,以下为几个需要注意的点:
- 失败率的计算必须等环装满才会计算
- 白名单优先级高于黑名单且白名单上的异常会被忽略,不会占用缓冲环位置,即不会计入失败率计算
- 熔断器打开时同样会计算失败率,当状态转换为半开时重置为-1
- 只要出现异常都可以调用降级方法,不论是在白名单还是黑名单
- 熔断器的缓冲环有两个,一个关闭时的缓冲环,一个打开时的缓冲环
- 熔断器关闭时,直至熔断器状态转换前所有请求都会通过,不会受到限制
- 熔断器半开时,限制请求数为缓冲环的大小,其他请求会等待
- 熔断器从打开到半开的转换默认还需要请求进行触发,也可通过automaticTransitionFromOpenToHalfOpenEnabled=true设置为自动触发
TimeLimiter
简介
与Hystrix不同,Resilience4j将超时控制器从熔断器中独立出来,成为了一个单独的组件,主要的作用就是对方法调用进行超时控制。实现的原理和Hystrix相似,都是通过调用Future的get方法来进行超时控制。
可配置参数
配置参数 | 默认值 | 描述 |
---|---|---|
timeoutDuration | 1(s) | 超时时间限定 |
cancelRunningFuture | true | 当超时时是否关闭取消线程 |
测试前准备
pom.xml
io.github.resilience4j
resilience4j-timelimiter
0.16.0
TimeLimiter没有整合进resilience4j-spring-boot中,需要单独添加依赖
application.yml配置
timelimiter:
timeoutDuration: 3000 # 超时时长
cancelRunningFuture: true # 发生异常是否关闭线程
TimeLimiter没有配置自动注入,需要自己进行注入,写下面两个文件进行配置自动注入:
TimeLimiterProperties
用于将application.yml中的配置转换为TimeLimiterProperties对象:
@Data
@Component
@ConfigurationProperties(prefix = "resilience4j.timelimiter")
public class TimeLimiterProperties {
private Duration timeoutDuration;
private boolean cancelRunningFuture;
}
TimeLimiterConfiguration
将TimeLimiterProperties对象写入到TimeLimiter的配置中:
@Configuration
public class TimeLimiterConfiguration {
@Autowired
private TimeLimiterProperties timeLimiterProperties;
@Bean
public TimeLimiter timeLimiter(){
return TimeLimiter.of(timeLimiterConfig());
}
private TimeLimiterConfig timeLimiterConfig(){
return TimeLimiterConfig.custom()
.timeoutDuration(timeLimiterProperties.getTimeoutDuration())
.cancelRunningFuture(timeLimiterProperties.isCancelRunningFuture()).build();
}
}
调用方法
还是以之前查询用户列表的后端服务为例。TimeLimiter目前仅支持程序式调用,还不能使用AOP的方式调用。
因为TimeLimiter通常与CircuitBreaker联合使用,很少单独使用,所以直接介绍联合使用的步骤。
TimeLimiter没有注册器,所以通过@Autowired注解自动注入依赖直接使用,因为TimeLimter是基于Future的get方法的,所以需要创建线程池,然后通过线程池的submit方法获取Future对象:
public class CircuitBreakerServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private CircuitBreakerRegistry circuitBreakerRegistry;
@Autowired
private TimeLimiter timeLimiter;
public List circuitBreakerTimeLimiter(){
// 通过注册器获取熔断器的实例
CircuitBreaker circuitBreaker = circuitBreakerRegistry.circuitBreaker("backendA");
CircuitBreakerUtil.getCircuitBreakerStatus("执行开始前:", circuitBreaker);
// 创建单线程的线程池
ExecutorService pool = Executors.newSingleThreadExecutor();
//将被保护方法包装为能够返回Future的supplier函数
Supplier>> futureSupplier = () -> pool.submit(remoteServiceConnector::process);
// 先用限时器包装,再用熔断器包装
Callable> restrictedCall = TimeLimiter.decorateFutureSupplier(timeLimiter, futureSupplier);
Callable> chainedCallable = CircuitBreaker.decorateCallable(circuitBreaker, restrictedCall);
// 使用Try.of().recover()调用并进行降级处理
Try> result = Try.of(chainedCallable::call)
.recover(CallNotPermittedException.class, throwable ->{
log.info("熔断器已经打开,拒绝访问被保护方法~");
CircuitBreakerUtil.getCircuitBreakerStatus("熔断器打开中", circuitBreaker);
List users = new ArrayList();
return users;
})
.recover(throwable -> {
log.info(throwable.getLocalizedMessage() + ",方法被降级了~~");
CircuitBreakerUtil.getCircuitBreakerStatus("降级方法中:",circuitBreaker);
List users = new ArrayList();
return users;
});
CircuitBreakerUtil.getCircuitBreakerStatus("执行结束后:", circuitBreaker);
return result.get();
}
}
使用测试
异常A和B在application.yml文件中没有修改:
recordExceptions: # 记录的异常
- com.example.resilience4j.exceptions.BusinessBException
- com.example.resilience4j.exceptions.BusinessAException
ignoreExceptions: # 忽略的异常
- com.example.resilience4j.exceptions.BusinessAException
使用另一个远程服务接口的实现,将num%4==3的情况让线程休眠5s,大于我们TimeLimiter的限制时间:
public class RemoteServiceImpl implements RemoteService {
private static AtomicInteger count = new AtomicInteger(0);
public List process() {
int num = count.getAndIncrement();
log.info("count的值 = " + num);
if (num % 4 == 1){
throw new BusinessAException("异常A,不需要被记录");
}
if (num % 4 == 2){
throw new BusinessBException("异常B,需要被记录");
}
if (num % 4 == 3){
Thread.sleep(5000);
}
log.info("服务正常运行,获取用户列表");
// 模拟数据库的正常查询
return repository.findAll();
}
}
把调用方法进行单元测试,循环10遍:
public class CircuitBreakerServiceImplTest{
@Autowired
private CircuitBreakerServiceImpl circuitService;
@Test
public void circuitBreakerTimeLimiterTest() {
for (int i=0; i<10; i++){
circuitService.circuitBreakerTimeLimiter();
}
}
}
看下运行结果:
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=0, failedCalls=0, successCalls=0, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 0
服务正常运行,获取用户列表
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 1
com.example.resilience4j.exceptions.BusinessAException: 异常A,不需要被记录,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 2
com.example.resilience4j.exceptions.BusinessBException: 异常B,需要被记录,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行开始前:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 3
null,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
发现熔断器任何异常和超时都没有失败。。完全不会触发熔断,这是为什么呢?我们把异常toString()看一下:
java.util.concurrent.ExecutionException: com.example.resilience4j.exceptions.BusinessBException: 异常B,需要被记录,方法被降级了~~
java.util.concurrent.TimeoutException,方法被降级了~~
这下原因就很明显了,线程池会将线程中的任何异常包装为ExecutionException,而熔断器没有把异常解包,由于我们设置了黑名单,而熔断器又没有找到黑名单上的异常,所以失效了。这是一个已知的bug,会在下个版本(0.16.0之后)中修正,目前来说如果需要同时使用TimeLimiter和CircuitBreaker的话,黑白名单的设置是不起作用的,需要自定义自己的谓词逻辑,并在test()方法中将异常解包进行判断,比如像下面这样:
public class RecordFailurePredicate implements Predicate {
@Override
public boolean test(Throwable throwable) {
if (throwable.getCause() instanceof BusinessAException) return false;
else return true;
}
}
然后在application.yml文件中指定这个类作为判断类:
circuitbreaker:
configs:
default:
recordFailurePredicate: com.example.resilience4j.predicate.RecordFailurePredicate
就能自定义自己的黑白名单了,我们再运行一次试试:
java.util.concurrent.TimeoutException,方法被降级了~~
降级方法中:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=2, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
执行结束后:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=3, failedCalls=2, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
可以看出,TimeLimiter已经生效了,同时CircuitBreaker也正常工作。
Note:
最新版0.17.0,该bug已经修复,黑白名单可以正常使用。
Retry
简介
同熔断器一样,重试组件也提供了注册器,可以通过注册器获取实例来进行重试,同样可以跟熔断器配合使用。
可配置参数
配置参数 | 默认值 | 描述 |
---|---|---|
maxAttempts | 3 | 最大重试次数 |
waitDuration | 500[ms] | 固定重试间隔 |
intervalFunction | numberOfAttempts -> waitDuration | 用来改变重试时间间隔,可以选择指数退避或者随机时间间隔 |
retryOnResultPredicate | result -> false | 自定义结果重试规则,需要重试的返回true |
retryOnExceptionPredicate | throwable -> true | 自定义异常重试规则,需要重试的返回true |
retryExceptions | empty | 需要重试的异常列表 |
ignoreExceptions | empty | 需要忽略的异常列表 |
测试前准备
pom.xml
不需要引入新的依赖,已经集成在resilience4j-spring-boot中了
application.yml配置
resilience4j:
retry:
configs:
default:
maxRetryAttempts: 3
waitDuration: 10s
enableExponentialBackoff: true # 是否允许使用指数退避算法进行重试间隔时间的计算
expontialBackoffMultiplier: 2 # 指数退避算法的乘数
enableRandomizedWait: false # 是否允许使用随机的重试间隔
randomizedWaitFactor: 0.5 # 随机因子
resultPredicate: com.example.resilience4j.predicate.RetryOnResultPredicate
retryExceptionPredicate: com.example.resilience4j.predicate.RetryOnExceptionPredicate
retryExceptions:
- com.example.resilience4j.exceptions.BusinessBException
- com.example.resilience4j.exceptions.BusinessAException
- io.github.resilience4j.circuitbreaker.CallNotPermittedException
ignoreExceptions:
- io.github.resilience4j.circuitbreaker.CallNotPermittedException
instances:
backendA:
baseConfig: default
waitDuration: 5s
backendB:
baseConfig: default
maxRetryAttempts: 2
application.yml可以配置的参数多出了几个enableExponentialBackoff、expontialBackoffMultiplier、enableRandomizedWait、randomizedWaitFactor,分别代表是否允许指数退避间隔时间,指数退避的乘数、是否允许随机间隔时间、随机因子,注意指数退避和随机间隔不能同时启用。
用于监控重试组件状态及事件的工具类
同样为了监控重试组件,写一个工具类:
@Log4j2
public class RetryUtil {
/**
* @Description: 获取重试的状态
*/
public static void getRetryStatus(String time, Retry retry){
Retry.Metrics metrics = retry.getMetrics();
long failedRetryNum = metrics.getNumberOfFailedCallsWithRetryAttempt();
long failedNotRetryNum = metrics.getNumberOfFailedCallsWithoutRetryAttempt();
long successfulRetryNum = metrics.getNumberOfSuccessfulCallsWithRetryAttempt();
long successfulNotyRetryNum = metrics.getNumberOfSuccessfulCallsWithoutRetryAttempt();
log.info(time + "state=" + " metrics[ failedRetryNum=" + failedRetryNum +
", failedNotRetryNum=" + failedNotRetryNum +
", successfulRetryNum=" + successfulRetryNum +
", successfulNotyRetryNum=" + successfulNotyRetryNum +
" ]"
);
}
/**
* @Description: 监听重试事件
*/
public static void addRetryListener(Retry retry){
retry.getEventPublisher()
.onSuccess(event -> log.info("服务调用成功:" + event.toString()))
.onError(event -> log.info("服务调用失败:" + event.toString()))
.onIgnoredError(event -> log.info("服务调用失败,但异常被忽略:" + event.toString()))
.onRetry(event -> log.info("重试:第" + event.getNumberOfRetryAttempts() + "次"))
;
}
}
调用方法
还是以之前查询用户列表的服务为例。Retry支持AOP和程序式两种方式的调用.
程序式的调用方法
和CircuitBreaker的调用方式差不多,和熔断器配合使用有两种调用方式,一种是先用重试组件装饰,再用熔断器装饰,这时熔断器的失败需要等重试结束才计算,另一种是先用熔断器装饰,再用重试组件装饰,这时每次调用服务都会记录进熔断器的缓冲环中,需要注意的是,第二种方式需要把CallNotPermittedException放进重试组件的白名单中,因为熔断器打开时重试是没有意义的:
public class CircuitBreakerServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private CircuitBreakerRegistry circuitBreakerRegistry;
@Autowired
private RetryRegistry retryRegistry;
public List circuitBreakerRetryNotAOP(){
// 通过注册器获取熔断器的实例
CircuitBreaker circuitBreaker = circuitBreakerRegistry.circuitBreaker("backendA");
// 通过注册器获取重试组件实例
Retry retry = retryRegistry.retry("backendA");
CircuitBreakerUtil.getCircuitBreakerStatus("执行开始前:", circuitBreaker);
// 先用重试组件包装,再用熔断器包装
CheckedFunction0> checkedSupplier = Retry.decorateCheckedSupplier(retry, remoteServiceConnector::process);
CheckedFunction0> chainedSupplier = CircuitBreaker .decorateCheckedSupplier(circuitBreaker, checkedSupplier);
// 使用Try.of().recover()调用并进行降级处理
Try> result = Try.of(chainedSupplier).
recover(CallNotPermittedException.class, throwable -> {
log.info("已经被熔断,停止重试");
return new ArrayList<>();
})
.recover(throwable -> {
log.info("重试失败: " + throwable.getLocalizedMessage());
return new ArrayList<>();
});
RetryUtil.getRetryStatus("执行结束: ", retry);
CircuitBreakerUtil.getCircuitBreakerStatus("执行结束:", circuitBreaker);
return result.get();
}
}
AOP式的调用方法
首先在连接器方法上使用@Retry(name="",fallbackMethod="")注解,其中name是要使用的重试器实例的名称,fallbackMethod是要使用的降级方法:
public RemoteServiceConnector{
@CircuitBreaker(name = "backendA", fallbackMethod = "fallBack")
@Retry(name = "backendA", fallbackMethod = "fallBack")
public List process() throws TimeoutException, InterruptedException {
List users;
users = remoteServic.process();
return users;
}
}
要求和熔断器一致,但是需要注意同时注解重试组件和熔断器的话,是按照第二种方案来的,即每一次请求都会被熔断器记录。
之后直接调用方法:
public class CircuitBreakerServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private CircuitBreakerRegistry circuitBreakerRegistry;
@Autowired
private RetryRegistry retryRegistry;
public List circuitBreakerRetryAOP() throws TimeoutException, InterruptedException {
List result = remoteServiceConnector.process();
RetryUtil.getRetryStatus("执行结束:", retryRegistry.retry("backendA"));
CircuitBreakerUtil
.getCircuitBreakerStatus("执行结束:", circuitBreakerRegistry.circuitBreaker("backendA"));
return result;
}
}
使用测试
异常A和B在application.yml文件中设定为都需要重试,因为使用第一种方案,所以不需要将CallNotPermittedException设定在重试组件的白名单中,同时为了测试重试过程中的异常是否会被熔断器记录,将异常A从熔断器白名单中去除:
recordExceptions: # 记录的异常
- com.example.resilience4j.exceptions.BusinessBException
- com.example.resilience4j.exceptions.BusinessAException
ignoreExceptions: # 忽略的异常
# - com.example.resilience4j.exceptions.BusinessAException
# ...
resultPredicate: com.example.resilience4j.predicate.RetryOnResultPredicate
retryExceptions:
- com.example.resilience4j.exceptions.BusinessBException
- com.example.resilience4j.exceptions.BusinessAException
- io.github.resilience4j.circuitbreaker.CallNotPermittedException
ignoreExceptions:
# - io.github.resilience4j.circuitbreaker.CallNotPermittedException
使用另一个远程服务接口的实现,将num%4==2的情况返回null,测试根据返回结果进行重试的功能:
public class RemoteServiceImpl implements RemoteService {
private static AtomicInteger count = new AtomicInteger(0);
public List process() {
int num = count.getAndIncrement();
log.info("count的值 = " + num);
if (num % 4 == 1){
throw new BusinessAException("异常A,需要重试");
}
if (num % 4 == 2){
return null;
}
if (num % 4 == 3){
throw new BusinessBException("异常B,需要重试");
}
log.info("服务正常运行,获取用户列表");
// 模拟数据库的正常查询
return repository.findAll();
}
}
同时添加一个类自定义哪些返回值需要重试,设定为返回值为空就进行重试,这样num % 4 == 2时就可以测试不抛异常,根据返回结果进行重试了:
public class RetryOnResultPredicate implements Predicate {
@Override
public boolean test(Object o) {
return o == null ? true : false;
}
}
使用CircuitBreakerServiceImpl中的AOP或者程序式调用方法进行单元测试,循环调用10次:
public class CircuitBreakerServiceImplTest{
@Autowired
private CircuitBreakerServiceImpl circuitService;
@Test
public void circuitBreakerRetryTest() {
for (int i=0; i<10; i++){
// circuitService.circuitBreakerRetryAOP();
circuitService.circuitBreakerRetryNotAOP();
}
}
}
看一下运行结果:
count的值 = 0
服务正常运行,获取用户列表
执行结束: state= metrics[ failedRetryNum=0, failedNotRetryNum=0, successfulRetryNum=0, successfulNotyRetryNum=1 ]
执行结束:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 1
重试:第1次
count的值 = 2
重试:第2次
count的值 = 3
服务调用失败:2019-07-09T19:06:59.705+08:00[Asia/Shanghai]: Retry 'backendA' recorded a failed retry attempt. Number of retry attempts: '3', Last exception was: 'com.example.resilience4j.exceptions.BusinessBException: 异常B,需要重试'.
重试失败: 异常B,需要重试
执行结束: state= metrics[ failedRetryNum=1, failedNotRetryNum=0, successfulRetryNum=0, successfulNotyRetryNum=1 ]
执行结束:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=2, failedCalls=1, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
这部分结果可以看出来,重试最大次数设置为3结果其实只重试了2次,服务共执行了3次,重试3次后熔断器只记录了1次。而且返回值为null时也确实进行重试了。
服务正常运行,获取用户列表
执行结束: state= metrics[ failedRetryNum=2, failedNotRetryNum=0, successfulRetryNum=0, successfulNotyRetryNum=3 ]
执行结束:state=OPEN , metrics[ failureRate=40.0, bufferedCalls=5, failedCalls=2, successCalls=3, maxBufferCalls=5, notPermittedCalls=0 ]
已经被熔断,停止重试
执行结束: state= metrics[ failedRetryNum=2, failedNotRetryNum=0, successfulRetryNum=0, successfulNotyRetryNum=3 ]
执行结束:state=OPEN , metrics[ failureRate=40.0, bufferedCalls=5, failedCalls=2, successCalls=3, maxBufferCalls=5, notPermittedCalls=1 ]
当熔断之后不会再进行重试。
接下来我修改一下调用服务的实现:
public class RemoteServiceImpl implements RemoteService {
private static AtomicInteger count = new AtomicInteger(0);
public List process() {
int num = count.getAndIncrement();
log.info("count的值 = " + num);
if (num % 4 == 1){
throw new BusinessAException("异常A,需要重试");
}
if (num % 4 == 3){
return null;
}
if (num % 4 == 2){
throw new BusinessBException("异常B,需要重试");
}
log.info("服务正常运行,获取用户列表");
// 模拟数据库的正常查询
return repository.findAll();
}
}
将num%4==2变成异常B,num%4==3变成返回null,看一下最后一次重试返回值为null属于重试成功还是重试失败。
运行结果如下:
count的值 = 0
服务正常运行,获取用户列表
执行结束: state= metrics[ failedRetryNum=0, failedNotRetryNum=0, successfulRetryNum=0, successfulNotyRetryNum=1 ]
执行结束:state=CLOSED , metrics[ failureRate=-1.0, bufferedCalls=1, failedCalls=0, successCalls=1, maxBufferCalls=5, notPermittedCalls=0 ]
count的值 = 1
重试:第1次
count的值 = 2
重试:第2次
count的值 = 3
服务调用成功:2019-07-09T19:17:35.836+08:00[Asia/Shanghai]: Retry 'backendA' recorded a successful retry attempt. Number of retry attempts: '3', Last exception was: 'com.example.resilience4j.exceptions.BusinessBException: 异常B,需要重试'.
如上可知如果最后一次重试不抛出异常就算作重试成功,不管结果是否需要继续重试。
Bulkhead
简介
Resilence4j的Bulkhead提供两种实现,一种是基于信号量的,另一种是基于有等待队列的固定大小的线程池的,由于基于信号量的Bulkhead能很好地在多线程和I/O模型下工作,所以选择介绍基于信号量的Bulkhead的使用。
可配置参数
配置参数 | 默认值 | 描述 |
---|---|---|
maxConcurrentCalls | 25 | 可允许的最大并发线程数 |
maxWaitDuration | 0 | 尝试进入饱和舱壁时应阻止线程的最大时间 |
测试前准备
pom.xml
不需要引入新的依赖,已经集成在resilience4j-spring-boot中了
application.yml配置
resilience4j:
bulkhead:
configs:
default:
maxConcurrentCalls: 10
maxWaitDuration: 1000
instances:
backendA:
baseConfig: default
maxConcurrentCalls: 3
backendB:
baseConfig: default
maxWaitDuration: 100
和CircuitBreaker差不多,都是可以通过继承覆盖配置设定实例的。
用于监控Bulkhead状态及事件的工具类
同样为了监控Bulkhead组件,写一个工具类:
@Log4j2
public class BulkhdadUtil {
/**
* @Description: 获取bulkhead的状态
*/
public static void getBulkheadStatus(String time, Bulkhead bulkhead){
Bulkhead.Metrics metrics = bulkhead.getMetrics();
// Returns the number of parallel executions this bulkhead can support at this point in time.
int availableConcurrentCalls = metrics.getAvailableConcurrentCalls();
// Returns the configured max amount of concurrent calls
int maxAllowedConcurrentCalls = metrics.getMaxAllowedConcurrentCalls();
log.info(time + ", metrics[ availableConcurrentCalls=" + availableConcurrentCalls +
", maxAllowedConcurrentCalls=" + maxAllowedConcurrentCalls + " ]");
}
/**
* @Description: 监听bulkhead事件
*/
public static void addBulkheadListener(Bulkhead bulkhead){
bulkhead.getEventPublisher()
.onCallFinished(event -> log.info(event.toString()))
.onCallPermitted(event -> log.info(event.toString()))
.onCallRejected(event -> log.info(event.toString()));
}
}
调用方法
还是以之前查询用户列表的服务为例。Bulkhead支持AOP和程序式两种方式的调用。
程序式的调用方法
调用方法都类似,装饰方法之后用Try.of().recover()来执行:
public class BulkheadServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private BulkheadRegistry bulkheadRegistry;
public List bulkheadNotAOP(){
// 通过注册器获得Bulkhead实例
Bulkhead bulkhead = bulkheadRegistry.bulkhead("backendA");
BulkhdadUtil.getBulkheadStatus("开始执行前: ", bulkhead);
// 通过Try.of().recover()调用装饰后的服务
Try> result = Try.of(
Bulkhead.decorateCheckedSupplier(bulkhead, remoteServiceConnector::process))
.recover(BulkheadFullException.class, throwable -> {
log.info("服务失败: " + throwable.getLocalizedMessage());
return new ArrayList();
});
BulkhdadUtil.getBulkheadStatus("执行结束: ", bulkhead);
return result.get();
}
}
AOP式的调用方法
首先在连接器方法上使用@Bulkhead(name="", fallbackMethod="", type="")注解,其中name是要使用的Bulkhead实例的名称,fallbackMethod是要使用的降级方法,type是选择信号量或线程池的Bulkhead:
public RemoteServiceConnector{
@Bulkhead(name = "backendA", fallbackMethod = "fallback", type = Bulkhead.Type.SEMAPHORE)
public List process() throws TimeoutException, InterruptedException {
List users;
users = remoteServic.process();
return users;
}
private List fallback(BulkheadFullException e){
log.info("服务失败: " + e.getLocalizedMessage());
return new ArrayList();
}
}
如果Retry、CircuitBreaker、Bulkhead同时注解在方法上,默认的顺序是Retry>CircuitBreaker>Bulkhead,即先控制并发再熔断最后重试,之后直接调用方法:
public class BulkheadServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private BulkheadRegistry bulkheadRegistry;
public List bulkheadAOP() throws TimeoutException, InterruptedException {
List result = remoteServiceConnector.process();
BulkheadUtil.getBulkheadStatus("执行结束:", bulkheadRegistry.retry("backendA"));
return result;
}
}
使用测试
在application.yml文件中将backenA线程数限制为1,便于观察,最大等待时间为1s,超过1s的会走降级方法:
instances:
backendA:
baseConfig: default
maxConcurrentCalls: 1
使用另一个远程服务接口的实现,不抛出异常,当做正常服务进行:
public class RemoteServiceImpl implements RemoteService {
private static AtomicInteger count = new AtomicInteger(0);
public List process() {
int num = count.getAndIncrement();
log.info("count的值 = " + num);
log.info("服务正常运行,获取用户列表");
// 模拟数据库正常查询
return repository.findAll();
}
}
用线程池调5个线程去请求服务:
public class BulkheadServiceImplTest{
@Autowired
private BulkheadServiceImpl bulkheadService;
@Autowired
private BulkheadRegistry bulkheadRegistry;
@Test
public void bulkheadTest() {
BulkhdadUtil.addBulkheadListener(bulkheadRegistry.bulkhead("backendA"));
ExecutorService pool = Executors.newCachedThreadPool();
for (int i=0; i<5; i++){
pool.submit(() -> {
// bulkheadService.bulkheadAOP();
bulkheadService.bulkheadNotAOP();
});
}
pool.shutdown();
while (!pool.isTerminated());
}
}
}
看一下运行结果:
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
Bulkhead 'backendA' permitted a call.
count的值 = 0
服务正常运行,获取用户列表
开始执行前: , metrics[ availableConcurrentCalls=0, maxAllowedConcurrentCalls=1 ]
Bulkhead 'backendA' rejected a call.
Bulkhead 'backendA' rejected a call.
Bulkhead 'backendA' rejected a call.
Bulkhead 'backendA' rejected a call.
服务失败: Bulkhead 'backendA' is full and does not permit further calls
执行结束: , metrics[ availableConcurrentCalls=0, maxAllowedConcurrentCalls=1 ]
服务失败: Bulkhead 'backendA' is full and does not permit further calls
执行结束: , metrics[ availableConcurrentCalls=0, maxAllowedConcurrentCalls=1 ]
服务失败: Bulkhead 'backendA' is full and does not permit further calls
执行结束: , metrics[ availableConcurrentCalls=0, maxAllowedConcurrentCalls=1 ]
服务失败: Bulkhead 'backendA' is full and does not permit further calls
执行结束: , metrics[ availableConcurrentCalls=0, maxAllowedConcurrentCalls=1 ]
Bulkhead 'backendA' has finished a call.
执行结束: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
由上可以看出,5个请求只有一个进入,其余触发rejected事件,然后自动进入降级方法。接下来我们把等待时间稍微加长一些:
instances:
backendA:
baseConfig: default
maxConcurrentCalls: 1
maxWaitDuration: 5000
再运行一次:
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
开始执行前: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
Bulkhead 'backendA' permitted a call.
count的值 = 0
服务正常运行,获取用户列表
Bulkhead 'backendA' permitted a call.
count的值 = 1
Bulkhead 'backendA' has finished a call.
服务正常运行,获取用户列表
执行结束: , metrics[ availableConcurrentCalls=0, maxAllowedConcurrentCalls=1 ]
Bulkhead 'backendA' has finished a call.
执行结束: , metrics[ availableConcurrentCalls=1, maxAllowedConcurrentCalls=1 ]
Bulkhead 'backendA' permitted a call.
前面的线程没有马上被拒绝,而是等待了一段时间再执行。
RateLimiter
简介
高频控制是可以限制服务调用频率,Resilience4j的RateLimiter可以对频率进行纳秒级别的控制,在每一个周期刷新可以调用的次数,还可以设定线程等待权限的时间。
可配置参数
配置参数 | 默认值 | 描述 |
---|---|---|
timeoutDuration | 5[s] | 线程等待权限的默认等待时间 |
limitRefreshPeriod | 500[ns] | 权限刷新的时间,每个周期结束后,RateLimiter将会把权限计数设置为limitForPeriod的值 |
limiteForPeriod | 50 | 一个限制刷新期间的可用权限数 |
测试前准备
pom.xml
不需要引入新的依赖,已经集成在resilience4j-spring-boot中了
application.yml配置
resilience4j:
ratelimiter:
configs:
default:
limitForPeriod: 5
limitRefreshPeriod: 1s
timeoutDuration: 5s
instances:
backendA:
baseConfig: default
limitForPeriod: 1
backendB:
baseConfig: default
timeoutDuration: 0s
用于监控RateLimiter状态及事件的工具类
同样为了监控RateLimiter组件,写一个工具类:
@Log4j2
public class RateLimiterUtil {
/**
* @Description: 获取rateLimiter的状态
*/
public static void getRateLimiterStatus(String time, RateLimiter rateLimiter){
RateLimiter.Metrics metrics = rateLimiter.getMetrics();
// Returns the number of availablePermissions in this duration.
int availablePermissions = metrics.getAvailablePermissions();
// Returns the number of WaitingThreads
int numberOfWaitingThreads = metrics.getNumberOfWaitingThreads();
log.info(time + ", metrics[ availablePermissions=" + availablePermissions +
", numberOfWaitingThreads=" + numberOfWaitingThreads + " ]");
}
/**
* @Description: 监听rateLimiter事件
*/
public static void addRateLimiterListener(RateLimiter rateLimiter){
rateLimiter.getEventPublisher()
.onSuccess(event -> log.info(event.toString()))
.onFailure(event -> log.info(event.toString()));
}
}
调用方法
还是以之前查询用户列表的服务为例。RateLimiter支持AOP和程序式两种方式的调用。
程序式的调用方法
调用方法都类似,装饰方法之后用Try.of().recover()来执行:
public class RateLimiterServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private RateLimiterRegistry rateLimiterRegistry;
public List ratelimiterNotAOP(){
// 通过注册器获得RateLimiter实例
RateLimiter rateLimiter = rateLimiterRegistry.rateLimiter("backendA");
RateLimiterUtil.getRateLimiterStatus("开始执行前: ", rateLimiter);
// 通过Try.of().recover()调用装饰后的服务
Try> result = Try.of(
Bulkhead.decorateCheckedSupplier(rateLimiter, remoteServiceConnector::process))
.recover(BulkheadFullException.class, throwable -> {
log.info("服务失败: " + throwable.getLocalizedMessage());
return new ArrayList();
});
RateLimiterUtil.getRateLimiterStatus("执行结束: ", rateLimiter);
return result.get();
}
}
AOP式的调用方法
首先在连接器方法上使用@RateLimiter(name="", fallbackMethod="")注解,其中name是要使用的RateLimiter实例的名称,fallbackMethod是要使用的降级方法:
public RemoteServiceConnector{
@RateLimiter(name = "backendA", fallbackMethod = "fallback")
public List process() throws TimeoutException, InterruptedException {
List users;
users = remoteServic.process();
return users;
}
private List fallback(BulkheadFullException e){
log.info("服务失败: " + e.getLocalizedMessage());
return new ArrayList();
}
}
如果Retry、CircuitBreaker、Bulkhead、RateLimiter同时注解在方法上,默认的顺序是Retry>CircuitBreaker>RateLimiter>Bulkhead,即先控制并发再限流然后熔断最后重试
接下来直接调用方法:
public class RateLimiterServiceImpl {
@Autowired
private RemoteServiceConnector remoteServiceConnector;
@Autowired
private RateLimiterRegistry rateLimiterRegistry;
public List rateLimiterAOP() throws TimeoutException, InterruptedException {
List result = remoteServiceConnector.process();
BulkheadUtil.getBulkheadStatus("执行结束:", rateLimiterRegistry.retry("backendA"));
return result;
}
}
使用测试
在application.yml文件中将backenA设定为20s只能处理1个请求,为便于观察,刷新时间设定为20s,等待时间设定为5s:
configs:
default:
limitForPeriod: 5
limitRefreshPeriod: 20s
timeoutDuration: 5s
instances:
backendA:
baseConfig: default
limitForPeriod: 1
使用另一个远程服务接口的实现,不抛出异常,当做正常服务进行,为了让结果明显一些,让方法sleep 5秒:
public class RemoteServiceImpl implements RemoteService {
private static AtomicInteger count = new AtomicInteger(0);
public List process() throws InterruptedException {
int num = count.getAndIncrement();
log.info("count的值 = " + num);
Thread.sleep(5000);
log.info("服务正常运行,获取用户列表");
// 模拟数据库正常查询
return repository.findAll();
}
}
用线程池调5个线程去请求服务:
public class RateLimiterServiceImplTest{
@Autowired
private RateLimiterServiceImpl rateLimiterService;
@Autowired
private RateLimiterRegistry rateLimiterRegistry;
@Test
public void rateLimiterTest() {
RateLimiterUtil.addRateLimiterListener(rateLimiterRegistry.rateLimiter("backendA"));
ExecutorService pool = Executors.newCachedThreadPool();
for (int i=0; i<5; i++){
pool.submit(() -> {
// rateLimiterService.rateLimiterAOP();
rateLimiterService.rateLimiterNotAOP();
});
}
pool.shutdown();
while (!pool.isTerminated());
}
}
}
看一下测试结果:
开始执行前: , metrics[ availablePermissions=1, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=1, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=1, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=1, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=0, numberOfWaitingThreads=0 ]
RateLimiterEvent{type=SUCCESSFUL_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T17:06:15.735+08:00[Asia/Shanghai]}
count的值 = 0
RateLimiterEvent{type=FAILED_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T17:06:20.737+08:00[Asia/Shanghai]}
RateLimiterEvent{type=FAILED_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T17:06:20.739+08:00[Asia/Shanghai]}
RateLimiterEvent{type=FAILED_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T17:06:20.740+08:00[Asia/Shanghai]}
服务失败: RateLimiter 'backendA' does not permit further calls
服务失败: RateLimiter 'backendA' does not permit further calls
执行结束: , metrics[ availablePermissions=0, numberOfWaitingThreads=1 ]
执行结束: , metrics[ availablePermissions=0, numberOfWaitingThreads=1 ]
RateLimiterEvent{type=FAILED_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T17:06:20.745+08:00[Asia/Shanghai]}
服务正常运行,获取用户列表
服务失败: RateLimiter 'backendA' does not permit further calls
执行结束: , metrics[ availablePermissions=0, numberOfWaitingThreads=0 ]
服务失败: RateLimiter 'backendA' does not permit further calls
执行结束: , metrics[ availablePermissions=0, numberOfWaitingThreads=0 ]
执行结束: , metrics[ availablePermissions=1, numberOfWaitingThreads=0 ]
只有一个服务调用成功,其他都执行失败了。现在我们把刷新时间调成1s:
configs:
default:
limitForPeriod: 5
limitRefreshPeriod: 1s
timeoutDuration: 5s
instances:
backendA:
baseConfig: default
limitForPeriod: 1
重新执行,结果如下:
开始执行前: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
开始执行前: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
RateLimiterEvent{type=SUCCESSFUL_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T18:25:18.894+08:00[Asia/Shanghai]}
count的值 = 0
RateLimiterEvent{type=SUCCESSFUL_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T18:25:18.894+08:00[Asia/Shanghai]}
count的值 = 1
RateLimiterEvent{type=SUCCESSFUL_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T18:25:19.706+08:00[Asia/Shanghai]}
count的值 = 2
RateLimiterEvent{type=SUCCESSFUL_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T18:25:19.706+08:00[Asia/Shanghai]}
count的值 = 3
RateLimiterEvent{type=SUCCESSFUL_ACQUIRE, rateLimiterName='backendA', creationTime=2019-07-10T18:25:20.703+08:00[Asia/Shanghai]}
count的值 = 4
服务正常运行,获取用户列表
服务正常运行,获取用户列表
服务正常运行,获取用户列表
服务正常运行,获取用户列表
执行结束: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
执行结束: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
执行结束: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
执行结束: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
服务正常运行,获取用户列表
执行结束: , metrics[ availablePermissions=2, numberOfWaitingThreads=0 ]
可以看出,几个服务都被放入并正常执行了,即使上个服务还没完成,依然可以放入,只与时间有关,而与线程无关。