降級(jí)方式

image.png

Alibaba Sentinel 支持多種降級(jí)方式：

1. 根據(jù)響應(yīng)時(shí)間：判斷單位時(shí)間內(nèi)平均響應(yīng)時(shí)間是否達(dá)到閾值参淹；
2. 根據(jù)異常比例：判斷單位時(shí)間內(nèi)宁炫，異常數(shù)量和異常比例是否達(dá)到閾值；
3. 根據(jù)異常數(shù)量：判斷單位時(shí)間內(nèi)異常數(shù)量是否達(dá)到閾值州既；

一旦觸發(fā)熔斷谜洽，熔斷開(kāi)關(guān)將會(huì)打開(kāi)，這時(shí)將拒絕所有請(qǐng)求易桃，拒絕時(shí)間為設(shè)置的降級(jí)時(shí)間間隔褥琐。通過(guò)源碼我們可以發(fā)現(xiàn)，Sentinel直接使用的是ScheduledExecutorService開(kāi)啟的一個(gè)延遲任務(wù)來(lái)實(shí)現(xiàn)降級(jí)時(shí)間間隔晤郑。

如：響應(yīng)時(shí)間達(dá)到閾值敌呈，并且熔斷時(shí)間間隔配置為5S，這時(shí)熔斷開(kāi)關(guān)會(huì)打開(kāi)造寝，并且在5S內(nèi)拒絕所有請(qǐng)求磕洪，當(dāng)5S后熔斷開(kāi)關(guān)再次關(guān)閉，這時(shí)會(huì)放行請(qǐng)求诫龙，如果放行請(qǐng)求后又觸發(fā)了熔斷析显，那么又需要等5S鐘。

熔斷降級(jí)規(guī)則說(shuō)明

熔斷降級(jí)規(guī)則（DegradeRule）包含下面幾個(gè)重要的屬性：

最新參數(shù)請(qǐng)看官網(wǎng) 尺上。

源碼

public class DegradeRule extends AbstractRule {

    @SuppressWarnings("PMD.ThreadPoolCreationRule")
    private static ScheduledExecutorService pool = Executors.newScheduledThreadPool(
        Runtime.getRuntime().availableProcessors(), new NamedThreadFactory("sentinel-degrade-reset-task", true));

    public DegradeRule() {}

    public DegradeRule(String resourceName) {
        setResource(resourceName);
    }

    /**
     * RT threshold or exception ratio threshold count.
     */
    private double count;

    /**
     * Degrade recover timeout (in seconds) when degradation occurs.
     */
    private int timeWindow;

    /**
     * Degrade strategy (0: average RT, 1: exception ratio, 2: exception count).
     */
    private int grade = RuleConstant.DEGRADE_GRADE_RT;

    /**
     * Minimum number of consecutive slow requests that can trigger RT circuit breaking.
     *
     * @since 1.7.0
     */
    private int rtSlowRequestAmount = RuleConstant.DEGRADE_DEFAULT_SLOW_REQUEST_AMOUNT;

    /**
     * Minimum number of requests (in an active statistic time span) that can trigger circuit breaking.
     *
     * @since 1.7.0
     */
    private int minRequestAmount = RuleConstant.DEGRADE_DEFAULT_MIN_REQUEST_AMOUNT;

    ...

    // Internal implementation (will be deprecated and moved outside).

    private AtomicLong passCount = new AtomicLong(0);
    private final AtomicBoolean cut = new AtomicBoolean(false);

    @Override
    public boolean passCheck(Context context, DefaultNode node, int acquireCount, Object... args) {
        // 判斷熔斷開(kāi)關(guān)的狀態(tài)
        if (cut.get()) {
            return false;
        }

        ClusterNode clusterNode = ClusterBuilderSlot.getClusterNode(this.getResource());
        if (clusterNode == null) {
            return true;
        }

        if (grade == RuleConstant.DEGRADE_GRADE_RT) {  // 基于RT的熔斷模式
            // 獲取平均RT
            double rt = clusterNode.avgRt();
            // 判斷平均RT是否達(dá)到閾值材蛛，如果沒(méi)有就放行
            if (rt < this.count) {
                passCount.set(0);
                return true;
            }

            // 如果RT達(dá)到閾值，還需要判斷單位時(shí)間內(nèi)請(qǐng)求量是否達(dá)到閾值（默認(rèn)是5）
            if (passCount.incrementAndGet() < rtSlowRequestAmount) {
                return true;
            }
        } else if (grade == RuleConstant.DEGRADE_GRADE_EXCEPTION_RATIO) { // 基于異常比例的熔斷策略
            double exception = clusterNode.exceptionQps();
            double success = clusterNode.successQps();
            double total = clusterNode.totalQps();
            // If total amount is less than minRequestAmount, the request will pass.
            // 判斷單位時(shí)間內(nèi)請(qǐng)求數(shù)量是否發(fā)到熔斷閾值（默認(rèn)5）
            if (total < minRequestAmount) {
                return true;
            }

            // In the same aligned statistic time window,
            // "success" (aka. completed count) = exception count + non-exception count (realSuccess)
            double realSuccess = success - exception;
            // 判斷單位時(shí)間內(nèi)異常請(qǐng)求數(shù)量是否達(dá)到閾值
            if (realSuccess <= 0 && exception < minRequestAmount) {
                return true;
            }
            // 判斷異常比例是否達(dá)到閾值
            if (exception / success < count) {
                return true;
            }
        } else if (grade == RuleConstant.DEGRADE_GRADE_EXCEPTION_COUNT) { // 基于異常數(shù)量的熔斷策略
            double exception = clusterNode.totalException();
            // 判斷異常數(shù)量是否達(dá)到閾值
            if (exception < count) {
                return true;
            }
        }

        if (cut.compareAndSet(false, true)) {
            ResetTask resetTask = new ResetTask(this);
            pool.schedule(resetTask, timeWindow, TimeUnit.SECONDS);
        }

        return false;
    }

    private static final class ResetTask implements Runnable {

        private DegradeRule rule;

        ResetTask(DegradeRule rule) {
            this.rule = rule;
        }

        @Override
        public void run() {
            rule.passCount.set(0);
            rule.cut.set(false);
        }
    }
}

基于RT的熔斷流程圖：

image.png

總結(jié)

1. 有一個(gè)值得借鑒的地方怎抛，使用ScheduledExecutorService來(lái)實(shí)現(xiàn)延遲任務(wù)的執(zhí)行卑吭。
2. 基于RT的熔斷，框架是基于所有請(qǐng)求的平均響應(yīng)時(shí)間來(lái)實(shí)現(xiàn)的马绝，這種方式不會(huì)產(chǎn)生上下文切換豆赏。還有一種簡(jiǎn)單的方式，這種方式采用FutureTask機(jī)制迹淌，但是會(huì)產(chǎn)生上下文切換河绽，如：

ExecutorService executorService = Executors.newFixedThreadPool(10);
FutureTask<String> futureTask = new FutureTask<String>(() -> {
    // 業(yè)務(wù)邏輯
    return "處理結(jié)果";
});
executorService.submit(futureTask);
futureTask.get(5, TimeUnit.SECONDS);