作者：京东物流 杨建民

一、Sentinel简介

Sentinel 以流量为切入点，从流量控制、熔断降级、系统负载保护等多个维度保护服务的稳定性。

Sentinel 具有以下特征:

• 丰富的应用场景：秒杀（即突发流量控制在系统容量可以承受的范围）、消息削峰填谷、集群流量控制、实时熔断下游不可用应用等。
• 完备的实时监控：Sentinel 同时提供实时的监控功能。您可以在控制台中看到接入应用的单台机器秒级数据，甚至 500 台以下规模的集群的汇总运行情况。
• 广泛的开源生态：Sentinel 提供开箱即用的与其它开源框架/库的整合模块，例如与 Spring Cloud、Apache Dubbo、gRPC、Quarkus 的整合。您只需要引入相应的依赖并进行简单的配置即可快速地接入 Sentinel。同时 Sentinel 提供 Java/Go/C++ 等多语言的原生实现。
• 完善的 SPI 扩展机制：Sentinel 提供简单易用、完善的 SPI 扩展接口。您可以通过实现扩展接口来快速地定制逻辑。例如定制规则管理、适配动态数据源等

有关Sentinel的详细介绍以及和Hystrix的区别可以自行网上检索，推荐一篇文章：https://mp.weixin.qq.com/s/Q7Xv8cypQFrrOQhbd9BOXw

本次主要使用了Sentinel的降级、限流、系统负载保护功能

二、Sentinel关键技术源码解析

无论是限流、降级、负载等控制手段，大致流程如下：

•StatisticSlot 则用于记录、统计不同维度的 runtime 指标监控信息

•责任链依次触发后续 slot 的 entry 方法，如 SystemSlot、FlowSlot、DegradeSlot 等的规则校验；

•当后续的 slot 通过，没有抛出 BlockException 异常，说明该资源被成功调用，则增加执行线程数和通过的请求数等信息。

关于数据统计，主要会牵扯到 ArrayMetric、BucketLeapArray、MetricBucket、WindowWrap 等类。

项目结构

以下主要分析core包里的内容

2.1注解入口

2.1.1 Entry、Context、Node

SphU门面类的方法出参都是Entry，Entry可以理解为每次进入资源的一个凭证，如果调用SphO.entry()或者SphU.entry()能获取Entry对象，代表获取了凭证，没有被限流，否则抛出一个BlockException。

Entry中持有本次对资源调用的相关信息：

•createTime：创建该Entry的时间戳。

•curNode：Entry当前是在哪个节点。

•orginNode：Entry的调用源节点。

•resourceWrapper：Entry关联的资源信息。

Entry是一个抽象类，CtEntry是Entry的实现，CtEntry持有Context和调用链的信息

Context的源码注释如下，

1This class holds metadata of current invocation
2
3

Node的源码注释

1Holds real-time statistics for resources
2
3

Node中保存了对资源的实时数据的统计，Sentinel中的限流或者降级等功能就是通过Node中的数据进行判断的。Node是一个接口，里面定义了各种操作request、exception、rt、qps、thread的方法。

在细看Node实现时，不难发现LongAddr的使用，关于LongAddr和DoubleAddr都是java8 java.util.concurrent.atomic里的内容，感兴趣的小伙伴可以再深入研究一下，这两个是高并发下计数功能非常优秀的数据结构，实际应用场景里需要计数时可以考虑使用。

关于Node的介绍后续还会深入，此处大致先提一下这个概念。

2.2 初始化

2.2.1 Context初始化

在初始化slot责任链部分前，还执行了context的初始化，里面涉及几个重要概念，需要解释一下：

可以发现在Context初始化的过程中，会把EntranceNode加入到Root子节点中（实际Root本身是一个特殊的EntranceNode），并把EntranceNode放到contextNameNodeMap中。

之前简单提到过Node，是用来统计数据用的，不同Node功能如下：

•Node：用于完成数据统计的接口

•StatisticNode：统计节点，是Node接口的实现类，用于完成数据统计

•EntranceNode：入口节点，一个Context会有一个入口节点，用于统计当前Context的总体流量数据

•DefaultNode：默认节点，用于统计一个资源在当前Context中的流量数据

•ClusterNode：集群节点，用于统计一个资源在所有Context中的总体流量数据

1protected static Context trueEnter(String name, String origin) {
2        Context context = contextHolder.get();
3        if (context == null) {
4            Map<String, DefaultNode> localCacheNameMap = contextNameNodeMap;
5            DefaultNode node = localCacheNameMap.get(name);
6            if (node == null) {
7                if (localCacheNameMap.size() > Constants.MAX_CONTEXT_NAME_SIZE) {
8                    setNullContext();
9                    return NULL_CONTEXT;
10                } else {
11                    LOCK.lock();
12                    try {
13                        node = contextNameNodeMap.get(name);
14                        if (node == null) {
15                            if (contextNameNodeMap.size() > Constants.MAX_CONTEXT_NAME_SIZE) {
16                                setNullContext();
17                                return NULL_CONTEXT;
18                            } else {
19                                node = new EntranceNode(new StringResourceWrapper(name, EntryType.IN), null);
20                                // Add entrance node.
21                                Constants.ROOT.addChild(node);
22
23                                Map<String, DefaultNode> newMap = new HashMap<>(contextNameNodeMap.size() + 1);
24                                newMap.putAll(contextNameNodeMap);
25                                newMap.put(name, node);
26                                contextNameNodeMap = newMap;
27                            }
28                        }
29                    } finally {
30                        LOCK.unlock();
31                    }
32                }
33            }
34            context = new Context(node, name);
35            context.setOrigin(origin);
36            contextHolder.set(context);
37        }
38
39        return context;
40    }
41
42

2.2.2 通过SpiLoader默认初始化8个slot

每个slot的主要职责如下：

•NodeSelectorSlot 负责收集资源的路径，并将这些资源的调用路径，以树状结构存储起来，用于根据调用路径来限流降级；

•ClusterBuilderSlot 则用于存储资源的统计信息以及调用者信息，例如该资源的 RT, QPS, thread count 等等，这些信息将用作为多维度限流，降级的依据；

•StatisticSlot 则用于记录、统计不同纬度的 runtime 指标监控信息；

•FlowSlot 则用于根据预设的限流规则以及前面 slot 统计的状态，来进行流量控制；

•AuthoritySlot 则根据配置的黑白名单和调用来源信息，来做黑白名单控制；

•DegradeSlot 则通过统计信息以及预设的规则，来做熔断降级；

•SystemSlot 则通过系统的状态，例如 集群QPS、线程数、RT、负载 等，来控制总的入口流量；

2.3 StatisticSlot

2.3.1 Node

深入看一下Node，因为统计信息都在里面，后面不论是限流、熔断、负载保护等都是结合规则+统计信息判断是否要执行

从Node的源码注释看，它会持有资源维度的实时统计数据，以下是接口里的方法定义，可以看到totalRequest、totalPass、totalSuccess、blockRequest、totalException、passQps等很多request、qps、thread的相关方法：

1/**
2 * Holds real-time statistics for resources.
3 *
4 * @author qinan.qn
5 * @author leyou
6 * @author Eric Zhao
7 */
8public interface Node extends OccupySupport, DebugSupport {
9    long totalRequest();
10    long totalPass();
11    long totalSuccess();
12    long blockRequest();
13    long totalException();
14    double passQps();
15    double blockQps();
16    double totalQps();
17    double successQps();
18    ……
19}
20
21

2.3.2 StatisticNode

我们先从最基础的StatisticNode开始看，源码给出的定位是：

1The statistic node keep three kinds of real-time statistics metrics：
2metrics in second level ({@code rollingCounterInSecond})
3metrics in minute level ({@code rollingCounterInMinute})
4thread count
5
6

StatisticNode只有四个属性，除了之前提到过的LongAddr类型的curThreadNum外，还有两个属性是Metric对象，通过入参已经属性命名可以看出，一个用于秒级，一个用于分钟级统计。接下来我们就要看看Metric

1// StatisticNode持有两个Metric，一个秒级一个分钟级，由入参可知，秒级统计划分了两个时间窗口，窗口程度是500ms
2private transient volatile Metric rollingCounterInSecond = new ArrayMetric(SampleCountProperty.SAMPLE_COUNT,
3    IntervalProperty.INTERVAL);
4
5// 分钟级统计划分了60个时间窗口，窗口长度是1000ms
6private transient Metric rollingCounterInMinute = new ArrayMetric(60, 60 * 1000, false);
7
8/**
9 * The counter for thread count.
10 */
11private LongAdder curThreadNum = new LongAdder();
12
13/**
14 * The last timestamp when metrics were fetched.
15 */
16private long lastFetchTime = -1;
17
18

ArrayMetric只有一个属性LeapArray<MetricBucket>，其余都是用于统计的方法，LeapArray是sentinel中统计最基本的数据结构，这里有必要详细看一下，总体就是根据timeMillis去获取一个bucket，分为：没有创建、有直接返回、被废弃后的reset三种场景。

1//以分钟级的统计属性为例，看一下时间窗口初始化过程
2private transient Metric rollingCounterInMinute = new ArrayMetric(60, 60 * 1000, false);
3
4
5public LeapArray(int sampleCount, int intervalInMs) {
6        AssertUtil.isTrue(sampleCount > 0, "bucket count is invalid: " + sampleCount);
7        AssertUtil.isTrue(intervalInMs > 0, "total time interval of the sliding window should be positive");
8        AssertUtil.isTrue(intervalInMs % sampleCount == 0, "time span needs to be evenly divided");
9        // windowLengthInMs = 60*1000 / 60 = 1000 滑动窗口时间长度，可见sentinel默认将单位时间分为了60个滑动窗口进行数据统计
10        this.windowLengthInMs = intervalInMs / sampleCount;
11        // 60*1000
12        this.intervalInMs = intervalInMs;
13        // 60
14        this.intervalInSecond = intervalInMs / 1000.0;
15        // 60
16        this.sampleCount = sampleCount;
17        // 数组长度60
18        this.array = new AtomicReferenceArray<>(sampleCount);
19    }
20
21/**
22     * Get bucket item at provided timestamp.
23     *
24     * @param timeMillis a valid timestamp in milliseconds
25     * @return current bucket item at provided timestamp if the time is valid; null if time is invalid
26     */
27    public WindowWrap<T> currentWindow(long timeMillis) {
28        if (timeMillis < 0) {
29            return null;
30        }
31        // 根据当前时间戳算一个数组索引
32        int idx = calculateTimeIdx(timeMillis);
33        // Calculate current bucket start time.
34        // timeMillis % 1000
35        long windowStart = calculateWindowStart(timeMillis);
36
37        /*
38         * Get bucket item at given time from the array.
39         *
40         * (1) Bucket is absent, then just create a new bucket and CAS update to circular array.
41         * (2) Bucket is up-to-date, then just return the bucket.
42         * (3) Bucket is deprecated, then reset current bucket.
43         */
44        while (true) {
45            WindowWrap<T> old = array.get(idx);
46            if (old == null) {
47                /*
48                 *     B0       B1      B2    NULL      B4
49                 * ||_______|_______|_______|_______|_______||___
50                 * 200     400     600     800     1000    1200  timestamp
51                 *                             ^
52                 *                          time=888
53                 *            bucket is empty, so create new and update
54                 *
55                 * If the old bucket is absent, then we create a new bucket at {@code windowStart},
56                 * then try to update circular array via a CAS operation. Only one thread can
57                 * succeed to update, while other threads yield its time slice.
58                 */
59                // newEmptyBucket 方法重写，秒级和分钟级统计对象实现不同
60                WindowWrap<T> window = new WindowWrap<T>(windowLengthInMs, windowStart, newEmptyBucket(timeMillis));
61                if (array.compareAndSet(idx, null, window)) {
62                    // Successfully updated, return the created bucket.
63                    return window;
64                } else {
65                    // Contention failed, the thread will yield its time slice to wait for bucket available.
66                    Thread.yield();
67                }
68            } else if (windowStart == old.windowStart()) {
69                /*
70                 *     B0       B1      B2     B3      B4
71                 * ||_______|_______|_______|_______|_______||___
72                 * 200     400     600     800     1000    1200  timestamp
73                 *                             ^
74                 *                          time=888
75                 *            startTime of Bucket 3: 800, so it's up-to-date
76                 *
77                 * If current {@code windowStart} is equal to the start timestamp of old bucket,
78                 * that means the time is within the bucket, so directly return the bucket.
79                 */
80                return old;
81            } else if (windowStart > old.windowStart()) {
82                /*
83                 *   (old)
84                 *             B0       B1      B2    NULL      B4
85                 * |_______||_______|_______|_______|_______|_______||___
86                 * ...    1200     1400    1600    1800    2000    2200  timestamp
87                 *                              ^
88                 *                           time=1676
89                 *          startTime of Bucket 2: 400, deprecated, should be reset
90                 *
91                 * If the start timestamp of old bucket is behind provided time, that means
92                 * the bucket is deprecated. We have to reset the bucket to current {@code windowStart}.
93                 * Note that the reset and clean-up operations are hard to be atomic,
94                 * so we need a update lock to guarantee the correctness of bucket update.
95                 *
96                 * The update lock is conditional (tiny scope) and will take effect only when
97                 * bucket is deprecated, so in most cases it won't lead to performance loss.
98                 */
99                if (updateLock.tryLock()) {
100                    try {
101                        // Successfully get the update lock, now we reset the bucket.
102                        return resetWindowTo(old, windowStart);
103                    } finally {
104                        updateLock.unlock();
105                    }
106                } else {
107                    // Contention failed, the thread will yield its time slice to wait for bucket available.
108                    Thread.yield();
109                }
110            } else if (windowStart < old.windowStart()) {
111                // Should not go through here, as the provided time is already behind.
112                return new WindowWrap<T>(windowLengthInMs, windowStart, newEmptyBucket(timeMillis));
113            }
114        }
115    }
116// 持有一个时间窗口对象的数据，会根据当前时间戳除以时间窗口长度然后散列到数组中
117private int calculateTimeIdx(/*@Valid*/ long timeMillis) {
118        long timeId = timeMillis / windowLengthInMs;
119        // Calculate current index so we can map the timestamp to the leap array.
120        return (int)(timeId % array.length());
121    }
122
123

WindowWrap持有了windowLengthInMs, windowStart和LeapArray（分钟统计实现是BucketLeapArray，秒级统计实现是OccupiableBucketLeapArray），对于分钟级别的统计，MetricBucket维护了一个longAddr数组和一个配置的minRT

1/**
2 * The fundamental data structure for metric statistics in a time span.
3 *
4 * @author jialiang.linjl
5 * @author Eric Zhao
6 * @see LeapArray
7 */
8public class BucketLeapArray extends LeapArray<MetricBucket> {
9
10    public BucketLeapArray(int sampleCount, int intervalInMs) {
11        super(sampleCount, intervalInMs);
12    }
13
14    @Override
15    public MetricBucket newEmptyBucket(long time) {
16        return new MetricBucket();
17    }
18
19    @Override
20    protected WindowWrap<MetricBucket> resetWindowTo(WindowWrap<MetricBucket> w, long startTime) {
21        // Update the start time and reset value.
22        w.resetTo(startTime);
23        w.value().reset();
24        return w;
25    }
26}
27
28

对于秒级统计，QPS=20场景下，如何准确统计的问题，此处用到了另外一个LeapArry实现FutureBucketLeapArray，至于秒级统计如何保证没有统计误差，读者可以再研究一下FutureBucketLeapArray的上下文就好。

2.4 FlowSlot

2.4.1 常见限流算法

介绍sentinel限流实现前，先介绍一下常见限流算法，基本分为三种：计数器、漏斗、令牌桶。

计数器算法

顾名思义，计数器算法就是统计某个时间段内的请求，每单位时间加1，然后与配置的限流值（最大QPS）进行比较，如果超出则触发限流。但是这种算法不能做到“平滑限流”，以1s为单位时间，100QPS为限流值为例，如下图，会出现某时段超出限流值的情况

因此在单纯计数器算法上，又出现了滑动窗口计数器算法，我们将统计时间细分，比如将1s统计时长分为5个时间窗口，通过滚动统计所有时间窗口的QPS作为系统实际的QPS的方式，就能解决上述临界统计问题，后续我们看sentinel源码时也能看到类似操作。

漏斗算法

不论流量有多大都会先到漏桶中，然后以均匀的速度流出。如何在代码中实现这个匀速呢？比如我们想让匀速为100q/s，那么我们可以得到每流出一个流量需要消耗10ms，类似一个队列，每隔10ms从队列头部取出流量进行放行，而我们的队列也就是漏桶，当流量大于队列的长度的时候，我们就可以拒绝超出的部分。

漏斗算法同样的也有一定的缺点：无法应对突发流量。比如一瞬间来了100个请求，在漏桶算法中只能一个一个的过去，当最后一个请求流出的时候时间已经过了一秒了，所以漏斗算法比较适合请求到达比较均匀，需要严格控制请求速率的场景。

令牌桶算法

令牌桶算法和漏斗算法比较类似，区别是令牌桶存放的是令牌数量不是请求数量，令牌桶可以根据自身需求多样性得管理令牌的生产和消耗，可以解决突发流量的问题。

2.4.2 单机限流模式

接下来我们看一下Sentinel中的限流实现，相比上述基本限流算法，Sentinel限流的第一个特性就是引入“资源”的概念，可以细粒度多样性的支持特定资源、关联资源、指定链路的限流。

FlowSlot的主要逻辑都在FlowRuleChecker里，介绍之前，我们先看一下Sentinel关于规则的模型描述，下图分别是限流、访问控制规则、系统保护规则（Linux负载）、降级规则

1    /**
2     * 流量控制两种模式 
3     *   0: thread count（当调用该api的线程数达到阈值的时候，进行限流）
4     *   1: QPS（当调用该api的QPS达到阈值的时候，进行限流）
5     */
6    private int grade = RuleConstant.FLOW_GRADE_QPS;
7
8    /**
9     * 流量控制阈值，值含义与grade有关
10     */
11    private double count;
12
13    /**
14     * 调用关系限流策略（可以支持关联资源或指定链路的多样性限流需求）
15     *  直接（api 达到限流条件时，直接限流）
16     *  关联（当关联的资源达到限流阈值时，就限流自己）
17     *  链路（只记录指定链路上的流量）
18     * {@link RuleConstant#STRATEGY_DIRECT} for direct flow control (by origin);
19     * {@link RuleConstant#STRATEGY_RELATE} for relevant flow control (with relevant resource);
20     * {@link RuleConstant#STRATEGY_CHAIN} for chain flow control (by entrance resource).
21     */
22    private int strategy = RuleConstant.STRATEGY_DIRECT;
23
24    /**
25     * Reference resource in flow control with relevant resource or context.
26     */
27    private String refResource;
28
29    /**
30     * 流控效果：
31     * 0. default(reject directly),直接拒绝，抛异常FlowException
32     * 1. warm up, 慢启动模式（根据coldFactor（冷加载因子，默认3）的值，从阈值/coldFactor，经过预热时长，才达到设置的QPS阈值）
33     * 2. rate limiter  排队等待
34     * 3. warm up + rate limiter
35     */
36    private int controlBehavior = RuleConstant.CONTROL_BEHAVIOR_DEFAULT;
37
38    private int warmUpPeriodSec = 10;
39
40    /**
41     * Max queueing time in rate limiter behavior.
42     */
43    private int maxQueueingTimeMs = 500;
44
45    /**
46    *  是否集群限流，默认为否
47    */
48    private boolean clusterMode;
49    /**
50     * Flow rule config for cluster mode.
51     */
52    private ClusterFlowConfig clusterConfig;
53
54    /**
55     * The traffic shaping (throttling) controller.
56     */
57    private TrafficShapingController controller;
58
59

接着我们继续分析FlowRuleChecker

canPassCheck第一步会好看limitApp，这个是结合访问授权限制规则使用的，默认是所有。

1private static boolean passLocalCheck(FlowRule rule, Context context, DefaultNode node, int acquireCount,
2                                          boolean prioritized) {
3        // 根据策略选择Node来进行统计（可以是本身Node、关联的Node、指定的链路）
4        Node selectedNode = selectNodeByRequesterAndStrategy(rule, context, node);
5        if (selectedNode == null) {
6            return true;
7        }
8
9        return rule.getRater().canPass(selectedNode, acquireCount, prioritized);
10    }
11
12
13static Node selectNodeByRequesterAndStrategy(/*@NonNull*/ FlowRule rule, Context context, DefaultNode node) {
14        // limitApp是访问控制使用的，默认是default，不限制来源
15        String limitApp = rule.getLimitApp();
16        // 拿到限流策略
17        int strategy = rule.getStrategy();
18        String origin = context.getOrigin();
19        // 基于调用来源做鉴权
20        if (limitApp.equals(origin) && filterOrigin(origin)) {
21            if (strategy == RuleConstant.STRATEGY_DIRECT) {
22                // Matches limit origin, return origin statistic node.
23                return context.getOriginNode();
24            }
25            // 
26            return selectReferenceNode(rule, context, node);
27        } else if (RuleConstant.LIMIT_APP_DEFAULT.equals(limitApp)) {
28            if (strategy == RuleConstant.STRATEGY_DIRECT) {
29                // Return the cluster node.
30                return node.getClusterNode();
31            }
32
33            return selectReferenceNode(rule, context, node);
34        } else if (RuleConstant.LIMIT_APP_OTHER.equals(limitApp)
35            && FlowRuleManager.isOtherOrigin(origin, rule.getResource())) {
36            if (strategy == RuleConstant.STRATEGY_DIRECT) {
37                return context.getOriginNode();
38            }
39
40            return selectReferenceNode(rule, context, node);
41        }
42
43        return null;
44    }
45
46static Node selectReferenceNode(FlowRule rule, Context context, DefaultNode node) {
47        String refResource = rule.getRefResource();
48        int strategy = rule.getStrategy();
49
50        if (StringUtil.isEmpty(refResource)) {
51            return null;
52        }
53
54        if (strategy == RuleConstant.STRATEGY_RELATE) {
55            return ClusterBuilderSlot.getClusterNode(refResource);
56        }
57
58        if (strategy == RuleConstant.STRATEGY_CHAIN) {
59            if (!refResource.equals(context.getName())) {
60                return null;
61            }
62            return node;
63        }
64        // No node.
65        return null;
66    }
67
68// 此代码是load限流规则时根据规则初始化流量整形控制器的逻辑，rule.getRater()返回TrafficShapingController
69private static TrafficShapingController generateRater(/*@Valid*/ FlowRule rule) {
70        if (rule.getGrade() == RuleConstant.FLOW_GRADE_QPS) {
71            switch (rule.getControlBehavior()) {
72                // 预热模式返回WarmUpController
73                case RuleConstant.CONTROL_BEHAVIOR_WARM_UP:
74                    return new WarmUpController(rule.getCount(), rule.getWarmUpPeriodSec(),
75                            ColdFactorProperty.coldFactor);
76                // 排队模式返回ThrottlingController
77                case RuleConstant.CONTROL_BEHAVIOR_RATE_LIMITER:
78                    return new ThrottlingController(rule.getMaxQueueingTimeMs(), rule.getCount());
79                // 预热+排队模式返回WarmUpRateLimiterController
80                case RuleConstant.CONTROL_BEHAVIOR_WARM_UP_RATE_LIMITER:
81                    return new WarmUpRateLimiterController(rule.getCount(), rule.getWarmUpPeriodSec(),
82                            rule.getMaxQueueingTimeMs(), ColdFactorProperty.coldFactor);
83                case RuleConstant.CONTROL_BEHAVIOR_DEFAULT:
84                default:
85                    // Default mode or unknown mode: default traffic shaping controller (fast-reject).
86            }
87        }
88        // 默认是DefaultController
89        return new DefaultController(rule.getCount(), rule.getGrade());
90    }
91
92

Sentinel单机限流算法

上面我们看到根据限流规则controlBehavior属性（流控效果），会初始化以下实现：

•DefaultController：是一个非常典型的滑动窗口计数器算法实现，将当前统计的qps和请求进来的qps进行求和，小于限流值则通过，大于则计算一个等待时间，稍后再试

•ThrottlingController：是漏斗算法的实现，实现思路已经在源码片段中加了备注

•WarmUpController：实现参考了Guava的带预热的RateLimiter，区别是Guava侧重于请求间隔，类似前面提到的令牌桶，而Sentinel更关注于请求数，和令牌桶算法有点类似

•WarmUpRateLimiterController：低水位使用预热算法，高水位使用滑动窗口计数器算法排队。

DefaultController

1    @Override
2    public boolean canPass(Node node, int acquireCount, boolean prioritized) {
3        int curCount = avgUsedTokens(node);
4        if (curCount + acquireCount > count) {
5            if (prioritized && grade == RuleConstant.FLOW_GRADE_QPS) {
6                long currentTime;
7                long waitInMs;
8                currentTime = TimeUtil.currentTimeMillis();
9                waitInMs = node.tryOccupyNext(currentTime, acquireCount, count);
10                if (waitInMs < OccupyTimeoutProperty.getOccupyTimeout()) {
11                    node.addWaitingRequest(currentTime + waitInMs, acquireCount);
12                    node.addOccupiedPass(acquireCount);
13                    sleep(waitInMs);
14
15                    // PriorityWaitException indicates that the request will pass after waiting for {@link @waitInMs}.
16                    throw new PriorityWaitException(waitInMs);
17                }
18            }
19            return false;
20        }
21        return true;
22    }
23
24

ThrottlingController

1 public ThrottlingController(int queueingTimeoutMs, double maxCountPerStat) {
2        this(queueingTimeoutMs, maxCountPerStat, 1000);
3    }
4
5    public ThrottlingController(int queueingTimeoutMs, double maxCountPerStat, int statDurationMs) {
6        AssertUtil.assertTrue(statDurationMs > 0, "statDurationMs should be positive");
7        AssertUtil.assertTrue(maxCountPerStat >= 0, "maxCountPerStat should be >= 0");
8        AssertUtil.assertTrue(queueingTimeoutMs >= 0, "queueingTimeoutMs should be >= 0");
9        this.maxQueueingTimeMs = queueingTimeoutMs;
10        this.count = maxCountPerStat;
11        this.statDurationMs = statDurationMs;
12        // Use nanoSeconds when durationMs%count != 0 or count/durationMs> 1 (to be accurate)
13        // 可见配置限流值count大于1000时useNanoSeconds会是true否则是false
14        if (maxCountPerStat > 0) {
15            this.useNanoSeconds = statDurationMs % Math.round(maxCountPerStat) != 0 || maxCountPerStat / statDurationMs > 1;
16        } else {
17            this.useNanoSeconds = false;
18        }
19    }
20
21    @Override
22    public boolean canPass(Node node, int acquireCount) {
23        return canPass(node, acquireCount, false);
24    }
25
26    private boolean checkPassUsingNanoSeconds(int acquireCount, double maxCountPerStat) {
27        final long maxQueueingTimeNs = maxQueueingTimeMs * MS_TO_NS_OFFSET;
28        long currentTime = System.nanoTime();
29        // Calculate the interval between every two requests.
30        final long costTimeNs = Math.round(1.0d * MS_TO_NS_OFFSET * statDurationMs * acquireCount / maxCountPerStat);
31
32        // Expected pass time of this request.
33        long expectedTime = costTimeNs + latestPassedTime.get();
34
35        if (expectedTime <= currentTime) {
36            // Contention may exist here, but it's okay.
37            latestPassedTime.set(currentTime);
38            return true;
39        } else {
40            final long curNanos = System.nanoTime();
41            // Calculate the time to wait.
42            long waitTime = costTimeNs + latestPassedTime.get() - curNanos;
43            if (waitTime > maxQueueingTimeNs) {
44                return false;
45            }
46
47            long oldTime = latestPassedTime.addAndGet(costTimeNs);
48            waitTime = oldTime - curNanos;
49            if (waitTime > maxQueueingTimeNs) {
50                latestPassedTime.addAndGet(-costTimeNs);
51                return false;
52            }
53            // in race condition waitTime may <= 0
54            if (waitTime > 0) {
55                sleepNanos(waitTime);
56            }
57            return true;
58        }
59    }
60    
61    // 漏斗算法具体实现
62    private boolean checkPassUsingCachedMs(int acquireCount, double maxCountPerStat) {
63        long currentTime = TimeUtil.currentTimeMillis();
64        // 计算两次请求的间隔（分为秒级和纳秒级）
65        long costTime = Math.round(1.0d * statDurationMs * acquireCount / maxCountPerStat);
66
67        // 请求的期望的时间
68        long expectedTime = costTime + latestPassedTime.get();
69
70        if (expectedTime <= currentTime) {
71            // latestPassedTime是AtomicLong类型，支持volatile语义
72            latestPassedTime.set(currentTime);
73            return true;
74        } else {
75            // 计算等待时间
76            long waitTime = costTime + latestPassedTime.get() - TimeUtil.currentTimeMillis();
77            // 如果大于最大排队时间，则触发限流
78            if (waitTime > maxQueueingTimeMs) {
79                return false;
80            }
81            
82            long oldTime = latestPassedTime.addAndGet(costTime);
83            waitTime = oldTime - TimeUtil.currentTimeMillis();
84            if (waitTime > maxQueueingTimeMs) {
85                latestPassedTime.addAndGet(-costTime);
86                return false;
87            }
88            // in race condition waitTime may <= 0
89            if (waitTime > 0) {
90                sleepMs(waitTime);
91            }
92            return true;
93        }
94    }
95
96    @Override
97    public boolean canPass(Node node, int acquireCount, boolean prioritized) {
98        // Pass when acquire count is less or equal than 0.
99        if (acquireCount <= 0) {
100            return true;
101        }
102        // Reject when count is less or equal than 0.
103        // Otherwise, the costTime will be max of long and waitTime will overflow in some cases.
104        if (count <= 0) {
105            return false;
106        }
107        if (useNanoSeconds) {
108            return checkPassUsingNanoSeconds(acquireCount, this.count);
109        } else {
110            return checkPassUsingCachedMs(acquireCount, this.count);
111        }
112    }
113
114    private void sleepMs(long ms) {
115        try {
116            Thread.sleep(ms);
117        } catch (InterruptedException e) {
118        }
119    }
120
121    private void sleepNanos(long ns) {
122        LockSupport.parkNanos(ns);
123    }
124
125

1long costTime = Math.round(1.0d * statDurationMs * acquireCount / maxCountPerStat);
2
3

由上述计算两次请求间隔的公式我们可以发现，当maxCountPerStat（规则配置的限流值QPS）超过1000后，就无法准确计算出匀速排队模式下的请求间隔时长，因此对应前面介绍的，当规则配置限流值超过1000QPS后，会采用checkPassUsingNanoSeconds，小于1000QPS会采用checkPassUsingCachedMs，对比一下checkPassUsingNanoSeconds和checkPassUsingCachedMs，可以发现主体思路没变，只是统计维度从毫秒换算成了纳秒，因此只看checkPassUsingCachedMs实现就可以

WarmUpController

1 
2@Override
3    public boolean canPass(Node node, int acquireCount, boolean prioritized) {
4        long passQps = (long) node.passQps();
5
6        long previousQps = (long) node.previousPassQps();
7        syncToken(previousQps);
8
9        // 开始计算它的斜率
10        // 如果进入了警戒线，开始调整他的qps
11        long restToken = storedTokens.get();
12        if (restToken >= warningToken) {
13            long aboveToken = restToken - warningToken;
14            // 消耗的速度要比warning快，但是要比慢
15            // current interval = restToken*slope+1/count
16            double warningQps = Math.nextUp(1.0 / (aboveToken * slope + 1.0 / count));
17            if (passQps + acquireCount <= warningQps) {
18                return true;
19            }
20        } else {
21            if (passQps + acquireCount <= count) {
22                return true;
23            }
24        }
25
26        return false;
27    }
28
29protected void syncToken(long passQps) {
30        long currentTime = TimeUtil.currentTimeMillis();
31        currentTime = currentTime - currentTime % 1000;
32        long oldLastFillTime = lastFilledTime.get();
33        if (currentTime <= oldLastFillTime) {
34            return;
35        }
36
37        long oldValue = storedTokens.get();
38        long newValue = coolDownTokens(currentTime, passQps);
39
40        if (storedTokens.compareAndSet(oldValue, newValue)) {
41            long currentValue = storedTokens.addAndGet(0 - passQps);
42            if (currentValue < 0) {
43                storedTokens.set(0L);
44            }
45            lastFilledTime.set(currentTime);
46        }
47
48    }
49
50private long coolDownTokens(long currentTime, long passQps) {
51        long oldValue = storedTokens.get();
52        long newValue = oldValue;
53
54        // 添加令牌的判断前提条件:
55        // 当令牌的消耗程度远远低于警戒线的时候
56        if (oldValue < warningToken) {
57            newValue = (long)(oldValue + (currentTime - lastFilledTime.get()) * count / 1000);
58        } else if (oldValue > warningToken) {
59            if (passQps < (int)count / coldFactor) {
60                newValue = (long)(oldValue + (currentTime - lastFilledTime.get()) * count / 1000);
61            }
62        }
63        return Math.min(newValue, maxToken);
64    }
65
66

2.4.3 集群限流

passClusterCheck方法(因为clusterService找不到会降级到非集群限流)

1private static boolean passClusterCheck(FlowRule rule, Context context, DefaultNode node, int acquireCount,
2                                            boolean prioritized) {
3        try {
4            // 获取当前节点是Token Client还是Token Server
5            TokenService clusterService = pickClusterService();
6            if (clusterService == null) {
7                return fallbackToLocalOrPass(rule, context, node, acquireCount, prioritized);
8            }
9            long flowId = rule.getClusterConfig().getFlowId();
10            // 根据获取的flowId通过TokenService进行申请token。从上面可知，它可能是TokenClient调用的，也可能是ToeknServer调用的。分别对应的类是DefaultClusterTokenClient和DefaultTokenService
11            TokenResult result = clusterService.requestToken(flowId, acquireCount, prioritized);
12            return applyTokenResult(result, rule, context, node, acquireCount, prioritized);
13            // If client is absent, then fallback to local mode.
14        } catch (Throwable ex) {
15            RecordLog.warn("[FlowRuleChecker] Request cluster token unexpected failed", ex);
16        }
17        // Fallback to local flow control when token client or server for this rule is not available.
18        // If fallback is not enabled, then directly pass.
19        return fallbackToLocalOrPass(rule, context, node, acquireCount, prioritized);
20    }
21
22//获取当前节点是Token Client还是Token Server。
23//1） 如果当前节点的角色是Client，返回的TokenService为DefaultClusterTokenClient；
24//2）如果当前节点的角色是Server，则默认返回的TokenService为DefaultTokenService。
25private static TokenService pickClusterService() {
26        if (ClusterStateManager.isClient()) {
27            return TokenClientProvider.getClient();
28        }
29        if (ClusterStateManager.isServer()) {
30            return EmbeddedClusterTokenServerProvider.getServer();
31        }
32        return null;
33    }
34
35

集群限流模式

Sentinel 集群限流服务端有两种启动方式：

•嵌入模式（Embedded）适合应用级别的限流，部署简单，但对应用性能有影响

•独立模式（Alone）适合全局限流，需要独立部署

考虑到文章篇幅，集群限流有机会再展开详细介绍。

集群限流模式降级

1private static boolean passClusterCheck(FlowRule rule, Context context, DefaultNode node, int acquireCount,
2                                            boolean prioritized) {
3        try {
4            TokenService clusterService = pickClusterService();
5            if (clusterService == null) {
6                return fallbackToLocalOrPass(rule, context, node, acquireCount, prioritized);
7            }
8            long flowId = rule.getClusterConfig().getFlowId();
9            TokenResult result = clusterService.requestToken(flowId, acquireCount, prioritized);
10            return applyTokenResult(result, rule, context, node, acquireCount, prioritized);
11            // If client is absent, then fallback to local mode.
12        } catch (Throwable ex) {
13            RecordLog.warn("[FlowRuleChecker] Request cluster token unexpected failed", ex);
14        }
15        // Fallback to local flow control when token client or server for this rule is not available.
16        // If fallback is not enabled, then directly pass.
17        // 可以看到如果集群限流有异常，会降级到单机限流模式，如果配置不允许降级，那么直接会跳过此次校验
18        return fallbackToLocalOrPass(rule, context, node, acquireCount, prioritized);
19    }
20
21

2.5 DegradeSlot

CircuitBreaker

大神对断路器的解释：https://martinfowler.com/bliki/CircuitBreaker.html

首先就看到了根据资源名称获取断路器列表，Sentinel的断路器有两个实现：RT模式使用ResponseTimeCircuitBreaker、异常模式使用ExceptionCircuitBreaker

1public interface CircuitBreaker {
2
3    /**
4     * Get the associated circuit breaking rule.
5     *
6     * @return associated circuit breaking rule
7     */
8    DegradeRule getRule();
9
10    /**
11     * Acquires permission of an invocation only if it is available at the time of invoking.
12     *
13     * @param context context of current invocation
14     * @return {@code true} if permission was acquired and {@code false} otherwise
15     */
16    boolean tryPass(Context context);
17
18    /**
19     * Get current state of the circuit breaker.
20     *
21     * @return current state of the circuit breaker
22     */
23    State currentState();
24
25    /**
26     * <p>Record a completed request with the context and handle state transformation of the circuit breaker.</p>
27     * <p>Called when a <strong>passed</strong> invocation finished.</p>
28     *
29     * @param context context of current invocation
30     */
31    void onRequestComplete(Context context);
32
33    /**
34     * Circuit breaker state.
35     */
36    enum State {
37        /**
38         * In {@code OPEN} state, all requests will be rejected until the next recovery time point.
39         */
40        OPEN,
41        /**
42         * In {@code HALF_OPEN} state, the circuit breaker will allow a "probe" invocation.
43         * If the invocation is abnormal according to the strategy (e.g. it's slow), the circuit breaker
44         * will re-transform to the {@code OPEN} state and wait for the next recovery time point;
45         * otherwise the resource will be regarded as "recovered" and the circuit breaker
46         * will cease cutting off requests and transform to {@code CLOSED} state.
47         */
48        HALF_OPEN,
49        /**
50         * In {@code CLOSED} state, all requests are permitted. When current metric value exceeds the threshold,
51         * the circuit breaker will transform to {@code OPEN} state.
52         */
53        CLOSED
54    }
55}
56
57

以ExceptionCircuitBreaker为例看一下具体实现

1public class ExceptionCircuitBreaker extends AbstractCircuitBreaker {
2    
3    // 异常模式有两种，异常率和异常数
4    private final int strategy;
5    // 最小请求数
6    private final int minRequestAmount;
7    // 阈值
8    private final double threshold;
9    
10    // LeapArray是sentinel统计数据非常重要的一个结构，主要封装了时间窗口相关的操作
11    private final LeapArray<SimpleErrorCounter> stat;
12
13    public ExceptionCircuitBreaker(DegradeRule rule) {
14        this(rule, new SimpleErrorCounterLeapArray(1, rule.getStatIntervalMs()));
15    }
16
17    ExceptionCircuitBreaker(DegradeRule rule, LeapArray<SimpleErrorCounter> stat) {
18        super(rule);
19        this.strategy = rule.getGrade();
20        boolean modeOk = strategy == DEGRADE_GRADE_EXCEPTION_RATIO || strategy == DEGRADE_GRADE_EXCEPTION_COUNT;
21        AssertUtil.isTrue(modeOk, "rule strategy should be error-ratio or error-count");
22        AssertUtil.notNull(stat, "stat cannot be null");
23        this.minRequestAmount = rule.getMinRequestAmount();
24        this.threshold = rule.getCount();
25        this.stat = stat;
26    }
27
28    @Override
29    protected void resetStat() {
30        // Reset current bucket (bucket count = 1).
31        stat.currentWindow().value().reset();
32    }
33
34    
35    @Override
36    public void onRequestComplete(Context context) {
37        Entry entry = context.getCurEntry();
38        if (entry == null) {
39            return;
40        }
41        Throwable error = entry.getError();
42        SimpleErrorCounter counter = stat.currentWindow().value();
43        if (error != null) {
44            counter.getErrorCount().add(1);
45        }
46        counter.getTotalCount().add(1);
47
48        handleStateChangeWhenThresholdExceeded(error);
49    }
50
51    private void handleStateChangeWhenThresholdExceeded(Throwable error) {
52        if (currentState.get() == State.OPEN) {
53            return;
54        }
55        
56        if (currentState.get() == State.HALF_OPEN) {
57            // In detecting request
58            if (error == null) {
59                fromHalfOpenToClose();
60            } else {
61                fromHalfOpenToOpen(1.0d);
62            }
63            return;
64        }
65        
66        List<SimpleErrorCounter> counters = stat.values();
67        long errCount = 0;
68        long totalCount = 0;
69        for (SimpleErrorCounter counter : counters) {
70            
71 += counter.errorCount.sum();
72            totalCount += counter.totalCount.sum();
73        }
74        if (totalCount < minRequestAmount) {
75            return;
76        }
77        double curCount = errCount;
78        if (strategy == DEGRADE_GRADE_EXCEPTION_RATIO) {
79            // Use errorRatio
80            curCount = errCount * 1.0d / totalCount;
81        }
82        if (curCount > threshold) {
83            transformToOpen(curCount);
84        }
85    }
86
87    static class SimpleErrorCounter {
88        private LongAdder errorCount;
89        private LongAdder totalCount;
90
91        public SimpleErrorCounter() {
92            this.errorCount = new LongAdder();
93            this.totalCount = new LongAdder();
94        }
95
96        public LongAdder getErrorCount() {
97            return errorCount;
98        }
99
100        public LongAdder getTotalCount() {
101            return totalCount;
102        }
103
104        public SimpleErrorCounter reset() {
105            errorCount.reset();
106            totalCount.reset();
107            return this;
108        }
109
110        @Override
111        public String toString() {
112            return "SimpleErrorCounter{" +
113                "errorCount=" + errorCount +
114                ", totalCount=" + totalCount +
115                '}';
116        }
117    }
118
119    static class SimpleErrorCounterLeapArray extends LeapArray<SimpleErrorCounter> {
120
121        public SimpleErrorCounterLeapArray(int sampleCount, int intervalInMs) {
122            super(sampleCount, intervalInMs);
123        }
124
125        @Override
126        public SimpleErrorCounter newEmptyBucket(long timeMillis) {
127            return new SimpleErrorCounter();
128        }
129
130        @Override
131        protected WindowWrap<SimpleErrorCounter> resetWindowTo(WindowWrap<SimpleErrorCounter> w, long startTime) {
132            // Update the start time and reset value.
133            w.resetTo(startTime);
134            w.value().reset();
135            return w;
136        }
137    }
138}
139
140

2.6 SystemSlot

校验逻辑主要集中在com.alibaba.csp.sentinel.slots.system.SystemRuleManager#checkSystem，以下是片段，可以看到，作为负载保护规则校验，实现了集群的QPS、线程、RT（响应时间）、系统负载的控制，除系统负载以外，其余统计都是依赖StatisticSlot实现，系统负载是通过SystemRuleManager定时调度SystemStatusListener，通过OperatingSystemMXBean去获取

1/**
2     * Apply {@link SystemRule} to the resource. Only inbound traffic will be checked.
3     *
4     * @param resourceWrapper the resource.
5     * @throws BlockException when any system rule's threshold is exceeded.
6     */
7    public static void checkSystem(ResourceWrapper resourceWrapper, int count) throws BlockException {
8        if (resourceWrapper == null) {
9            return;
10        }
11        // Ensure the checking switch is on.
12        if (!checkSystemStatus.get()) {
13            return;
14        }
15
16        // for inbound traffic only
17        if (resourceWrapper.getEntryType() != EntryType.IN) {
18            return;
19        }
20
21        // total qps 此处是拿到某个资源在集群中的QPS总和，相关概念可以会看初始化关于Node的介绍
22        double currentQps = Constants.ENTRY_NODE.passQps();
23        if (currentQps + count > qps) {
24            throw new SystemBlockException(resourceWrapper.getName(), "qps");
25        }
26
27        // total thread 
28        int currentThread = Constants.ENTRY_NODE.curThreadNum();
29        if (currentThread > maxThread) {
30            throw new SystemBlockException(resourceWrapper.getName(), "thread");
31        }
32
33        double rt = Constants.ENTRY_NODE.avgRt();
34        if (rt > maxRt) {
35            throw new SystemBlockException(resourceWrapper.getName(), "rt");
36        }
37
38        // load. BBR algorithm.
39        if (highestSystemLoadIsSet && getCurrentSystemAvgLoad() > highestSystemLoad) {
40            if (!checkBbr(currentThread)) {
41                throw new SystemBlockException(resourceWrapper.getName(), "load");
42            }
43        }
44
45        // cpu usage
46        if (highestCpuUsageIsSet && getCurrentCpuUsage() > highestCpuUsage) {
47            throw new SystemBlockException(resourceWrapper.getName(), "cpu");
48        }
49    }
50
51    private static boolean checkBbr(int currentThread) {
52        if (currentThread > 1 &&
53            currentThread > Constants.ENTRY_NODE.maxSuccessQps() * Constants.ENTRY_NODE.minRt() / 1000) {
54            return false;
55        }
56        return true;
57    }
58
59    public static double getCurrentSystemAvgLoad() {
60        return statusListener.getSystemAverageLoad();
61    }
62
63    public static double getCurrentCpuUsage() {
64        return statusListener.getCpuUsage();
65    }
66
67

1public class SystemStatusListener implements Runnable {
2
3    volatile double currentLoad = -1;
4    volatile double currentCpuUsage = -1;
5
6    volatile String reason = StringUtil.EMPTY;
7
8    volatile long processCpuTime = 0;
9    volatile long processUpTime = 0;
10
11    public double getSystemAverageLoad() {
12        return currentLoad;
13    }
14
15    public double getCpuUsage() {
16        return currentCpuUsage;
17    }
18
19    @Override
20    public void run() {
21        try {
22            OperatingSystemMXBean osBean = ManagementFactory.getPlatformMXBean(OperatingSystemMXBean.class);
23            currentLoad = osBean.getSystemLoadAverage();
24
25            /*
26             * Java Doc copied from {@link OperatingSystemMXBean#getSystemCpuLoad()}:</br>
27             * Returns the "recent cpu usage" for the whole system. This value is a double in the [0.0,1.0] interval.
28             * A value of 0.0 means that all CPUs were idle during the recent period of time observed, while a value
29             * of 1.0 means that all CPUs were actively running 100% of the time during the recent period being
30             * observed. All values between 0.0 and 1.0 are possible depending of the activities going on in the
31             * system. If the system recent cpu usage is not available, the method returns a negative value.
32             */
33            double systemCpuUsage = osBean.getSystemCpuLoad();
34
35            // calculate process cpu usage to support application running in container environment
36            RuntimeMXBean runtimeBean = ManagementFactory.getPlatformMXBean(RuntimeMXBean.class);
37            long newProcessCpuTime = osBean.getProcessCpuTime();
38            long newProcessUpTime = runtimeBean.getUptime();
39            int cpuCores = osBean.getAvailableProcessors();
40            long processCpuTimeDiffInMs = TimeUnit.NANOSECONDS
41                    .toMillis(newProcessCpuTime - processCpuTime);
42            long processUpTimeDiffInMs = newProcessUpTime - processUpTime;
43            double processCpuUsage = (double) processCpuTimeDiffInMs / processUpTimeDiffInMs / cpuCores;
44            processCpuTime = newProcessCpuTime;
45            processUpTime = newProcessUpTime;
46
47            currentCpuUsage = Math.max(processCpuUsage, systemCpuUsage);
48
49            if (currentLoad > SystemRuleManager.getSystemLoadThreshold()) {
50                writeSystemStatusLog();
51            }
52        } catch (Throwable e) {
53            RecordLog.warn("[SystemStatusListener] Failed to get system metrics from JMX", e);
54        }
55    }
56
57    private void writeSystemStatusLog() {
58        StringBuilder sb = new StringBuilder();
59        sb.append("Load exceeds the threshold: ");
60        sb.append("load:").append(String.format("%.4f", currentLoad)).append("; ");
61        sb.append("cpuUsage:").append(String.format("%.4f", currentCpuUsage)).append("; ");
62        sb.append("qps:").append(String.format("%.4f", Constants.ENTRY_NODE.passQps())).append("; ");
63        sb.append("rt:").append(String.format("%.4f", Constants.ENTRY_NODE.avgRt())).append("; ");
64        sb.append("thread:").append(Constants.ENTRY_NODE.curThreadNum()).append("; ");
65        sb.append("success:").append(String.format("%.4f", Constants.ENTRY_NODE.successQps())).append("; ");
66        sb.append("minRt:").append(String.format("%.2f", Constants.ENTRY_NODE.minRt())).append("; ");
67        sb.append("maxSuccess:").append(String.format("%.2f", Constants.ENTRY_NODE.maxSuccessQps())).append("; ");
68        RecordLog.info(sb.toString());
69    }
70}
71
72

三、京东版最佳实践

3.1 使用方式

Sentinel使用方式本身非常简单，就是一个注解，但是要考虑规则加载和规则持久化的方式，现有的方式有：

•使用Sentinel-dashboard功能：使用面板接入需要维护一个配置规则的管理端，考虑到偏后端的系统需要额外维护一个面板成本较大，如果是像RPC框架这种本身有管理端的接入可以考虑次方案。

•中间件（如：zookepper、nacos、eureka、redis等）：Sentinel源码extension包里提供了类似的实现，如下图

结合京东实际，我实现了一个规则热部署的Sentinel组件，实现方式类似zookeeper的方式，将规则记录到ducc的一个key上，在spring容器启动时做第一次规则加载和监听器注册，组件也做一了一些规则读取，校验、实例化不同规则对象的工作

插件使用方式：注解+配置

第一步 引入组件

1<dependency>
2    <groupId>com.jd.ldop.tools</groupId>
3    <artifactId>sentinel-tools</artifactId>
4    <version>1.0.0-SNAPSHOT</version>
5</dependency>
6
7

第二步 初始化sentinelProcess

支持ducc、本地文件读取、直接写入三种方式规则写入方式

目前支持限流规则、熔断降级规则两种模式，系统负载保护模式待开发和验证

1<!-- 基于sentinel的降级、限流、熔断组件 -->
2    <bean id="sentinelProcess" class="com.jd.ldop.sentinel.SentinelProcess">
3        <property name="ruleResourceWrappers">
4            <list>
5                <ref bean="degradeRule"/>
6            </list>
7        </property>
8    </bean>
9
10    <!-- 降级或限流规则配置 -->
11    <bean id="degradeRule" class="com.jd.ldop.sentinel.dto.RuleResourceWrapper">
12        <constructor-arg index="0" value="ducc.degradeRule"/>
13        <constructor-arg index="1" value="0"/>
14        <constructor-arg index="2" value="0"/>
15    </bean>
16
17

ducc上配置如下：

第三步 定义资源和关联类型

通过@SentinelResource可以直接在任意位置定义资源名以及对应的熔断降级或者限流方式、回调方法等，同时也可以指定关联类型，支持直接、关联、指定链路三种

1    @Override
2    @SentinelResource(value = "modifyGetWaybillState", fallback = "executeDegrade")
3    public ExecutionResult<List<Integer>> execute(@NotNull Model imodel) {
4        // 业务逻辑处理
5    }
6
7    public ExecutionResult<List<Integer>> executeDegrade(@NotNull Model imodel) {
8        // 降级业务逻辑处理
9    }
10
11

3.2 应用场景

组件支持任意的业务降级、限流、负载保护

四、Sentinel压测数据

4.1 压测目标

调用量：1.2W/m

应用机器内存稳定在50%以内

机器规格： 8C16G50G磁盘*2

Sentinel降级规则：

count=350-------慢调用临界阈值350ms

timeWindow=180------熔断时间窗口180s

grade=0-----降级模式 慢调用

statIntervalMs=60000------统计时长1min

4.2 压测结果

应用机器监控：