Question

我需要记录请求的一些属性，例如请求ID和区域设置，但是当使用parallelStream时，似乎MDC的ThreadLocal会松散信息。

我已经分析了在创建parallelStream时在线程之间传递MDC上下文的解决方案，但是它看起来很脏，并且我也有很多parallelStream的用法。

还有其他方法吗？

谢谢

Answer 1

我发现的唯一解决方案是将上下文复制到流外部的最终变量中，并将其应用于每个迭代：

Map<String, String> contextMap = MDC.getCopyOfContextMap();
Stream.iterate(0, i -> i + 1).parallel()
    .peek(i -> MDC.setContextMap(contextMap))
    // ...logic...
    // in case you're using a filter, you need to use a predicate and combine it with a clear step:
    filter(yourPredicate.or(i -> {
                MDC.clear();
                return false;
            }))
    // clear right before terminal operation
    .peek(i -> MDC.clear())
    .findFirst();

// since the initial thread is also used within the stream and the context is cleared there, 
// we need to set it again to its initial state
MDC.setContextMap(contextMap);

该解决方案的成本为1）每100次迭代需要几微秒，并且2）可读性较差，但都可以接受：

这是一个比较IntStream.range(0, 100).parallel().sum()（=基准）和使用该MDC复制逻辑的相同流的基准：

Benchmark               Mode  Cnt   Score   Error   Units
MDC_CopyTest.baseline  thrpt    5   0,038 ± 0,005  ops/us
MDC_CopyTest.withMdc   thrpt    5   0,024 ± 0,001  ops/us
MDC_CopyTest.baseline   avgt    5  28,239 ± 1,308   us/op
MDC_CopyTest.withMdc    avgt    5  40,178 ± 0,761   us/op

为了提高可读性，可以将其包装到一个小的帮助程序类中：

public class MDCCopyHelper {
    private Map<String, String> contextMap = MDC.getCopyOfContextMap();

    public void set(Object... any) {
        MDC.setContextMap(contextMap);
    }

    public void clear(Object... any) {
        MDC.clear();
    }

    public boolean clearAndFail() {
        MDC.clear();
        return false;
    }
}

然后，流式代码看上去会更好一些：

MDCCopyHelper mdcHelper = new MDCCopyHelper();
Optional<Integer> findFirst = Stream.iterate(0, i -> i + 1)
        .parallel()
        .peek(mdcHelper::set)
        // ...logic...
        // filters predicates should be combined with clear step
        .filter(yourPredicate.or(mdcHelper::clearAndFail))
        // before terminal call:
        .peek(mdcHelper::clear)
        .findFirst();
mdcHelper.set();

Answer 2

我的解决方案是包装那些功能接口。类似于静态代理模式。
例如

public static void main(String[] args) {
    System.err.println(Thread.currentThread().getName());
    String traceId = "100";
    MDC.put("id", traceId);
    System.err.println("------------------------");
    Stream.of(1, 2, 3, 4)
          .parallel()
          .forEach((num -> {
              System.err.println(Thread.currentThread().getName()+" "+ traceId.equals(MDC.get("id")));
          }));
    System.err.println("------------------------");
    Stream.of(1, 2, 3, 4)
          .parallel()
          // the key is the TraceableConsumer
          .forEach(new TraceableConsumer<>(num -> {
              System.err.println(Thread.currentThread().getName() + " " + traceId.equals(MDC.get("id")));
          }));
}

public class TraceableConsumer<T> extends MDCTraceable implements Consumer<T> {

    private final Consumer<T> target;

    public TraceableConsumer(Consumer<T> target) {
        this.target = target;
    }

    @Override
    public void accept(T t) {
        setMDC();
        target.accept(t);
    }
}

public abstract class MDCTraceable {

    private final Long id;

    private final Long userId;

    public MDCTraceable() {
        id = Optional.ofNullable(MDC.get("id")).map(Long::parseLong).orElse(0L);
        userId = Optional.ofNullable(MDC.get("userId")).map(Long::parseLong).orElse(0L);
    }

    public void setMDC(){
        MDC.put("id", id.toString());
        MDC.put("userId", userId.toString());
    }

    public void cleanMDC(){
        MDC.clear();
    }
}

如何在Java中将MDC与parallelStream一起使用和登录

2 个答案: