浅谈Flink批模式Adaptive Hash Join

Flink批Hash Join递归超限问题

随着Flink流批一体能力的迅速发展以及Flink SQL易用性的提升，越来越多的厂商开始将Flink作为离线批处理引擎使用。在我们使用Flink进行大规模join操作时，也许会发生如下的异常，导致任务失败：

Hash join exceeded maximum number of recursions, without reducing partitions enough to be memory resident.

字面意思即为Hash Join的递归次数超出限制。Flink批模式下的join算法有两种，即Hybrid Hash Join和Sort-Merge Join。顾名思义，Hybrid Hash Join就是Simple Hash Join和Grace Hash Join两种算法的结合(关于它们，看官可参考这篇文章)。引用一张Flink官方博客中的手绘图来说明。

Flink的Hybrid Hash Join在build阶段会积极地利用TaskManager的托管内存，并将内存无法容纳的哈希分区spill到磁盘中。在probe阶段，当内存中的哈希分区处理完成后，会释放掉对应的MemorySegment，并将先前溢写到磁盘的分区读入，以提升probe效率。特别注意，如果溢写分区对空闲的托管内存而言仍然过大(特别是存在数据倾斜的情况时)，就会将其递归拆分成更小的分区，原理如下图所示。

当然，递归拆分也不能是无限制的。在Blink Runtime中，如果递归拆分3次仍然不能满足内存需求，就会抛出前文所述的异常了。

笔者在今年7月ApacheCon Asia 2022流处理专场的分享内容里谈到了这个问题，并且将其归咎于Flink SQL的CBO优化器的代价模型不太科学，导致其十分偏向选择Hash Join。由于修改的难度很大，所以暂时的workaround就是在任务失败后，自动设置table.exec.disabled-operators参数来禁用掉ShuffleHashJoin算子，从而强制使用Sort-Merge Join。

当然这仍然不算优雅的解决方法，接下来简要看看Flink 1.16版本中提出的更好一点的方案：Adaptive Hash Join。

Adaptive Hash Join的实现

所谓adaptive(自适应)，就是指Hash Join递归超限时，不必让任务失败，而是将这些大分区自动转为Sort-Merge Join来处理。

Blink Runtime中的哈希表有两种，即BinaryHashTable(key的类型为BinaryRowData)和LongHybridHashTable(key的类型为Long)。以前者为例，查看其prepareNextPartition()方法，该方法负责递归地取得下一个要处理的哈希分区。

private boolean prepareNextPartition() throws IOException {
        // finalize and cleanup the partitions of the current table
        // ......

        // there are pending partitions
        final BinaryHashPartition p = this.partitionsPending.get(0);
        // ......

        final int nextRecursionLevel = p.getRecursionLevel() + 1;
        if (nextRecursionLevel == 2) {
            LOG.info("Recursive hash join: partition number is " + p.getPartitionNumber());
        } else if (nextRecursionLevel > MAX_RECURSION_DEPTH) {
            LOG.info(
                    "Partition number [{}] recursive level more than {}, process the partition using SortMergeJoin later.",
                    p.getPartitionNumber(),
                    MAX_RECURSION_DEPTH);
            // if the partition has spilled to disk more than three times, process it by sort merge
            // join later
            this.partitionsPendingForSMJ.add(p);
            // also need to remove it from pending list
            this.partitionsPending.remove(0);
            // recursively get the next partition
            return prepareNextPartition();
        }

        // build the next table; memory must be allocated after this call
        buildTableFromSpilledPartition(p, nextRecursionLevel);

        // set the probe side
        setPartitionProbeReader(p);

        // unregister the pending partition
        this.partitionsPending.remove(0);
        this.currentRecursionDepth = p.getRecursionLevel() + 1;

        // recursively get the next
        return nextMatching();
    }

注意当递归深度超过MAX_RECURSION_DEPTH(常量定义即为3)时，会将分区直接放入一个名为partitionsPendingForSMJ的容器中，等待做Sort-Merge Join。另外，在该方法调用的buildTableFromSpilledPartition()方法(对溢写分区执行build操作)开头，去掉了对递归超限的判断，也就是说Hash join exceeded maximum number of recursions异常已经成为历史。

那么等待做Sort-Merge Join的分区是如何被处理的？查看Blink Runtime中的HashJoinOperator算子，在构造该算子时，需要比原来多传入一个SortMergeJoinFunction的实例：

private final SortMergeJoinFunction sortMergeJoinFunction;

SortMergeJoinFunction实际上是将旧版的SortMergeJoinOperator处理逻辑抽离出来的类，算法本身没有任何变化。然后从哈希表中读取前述的partitionsPendingForSMJ容器，对每个分区的build侧和probe侧分别执行Sort-Merge Join操作即可。

/**
     * If here also exists partitions which spilled to disk more than three time when hash join end,
     * means that the key in these partitions is very skewed, so fallback to sort merge join
     * algorithm to process it.
     */
    private void fallbackSMJProcessPartition() throws Exception {
        if (!table.getPartitionsPendingForSMJ().isEmpty()) {
            // release memory to MemoryManager first that is used to sort merge join operator
            table.releaseMemoryCacheForSMJ();
            // initialize sort merge join operator
            LOG.info("Fallback to sort merge join to process spilled partitions.");
            initialSortMergeJoinFunction();
            fallbackSMJ = true;

            for (BinaryHashPartition p : table.getPartitionsPendingForSMJ()) {
                // process build side
                RowIterator<BinaryRowData> buildSideIter =
                        table.getSpilledPartitionBuildSideIter(p);
                while (buildSideIter.advanceNext()) {
                    processSortMergeJoinElement1(buildSideIter.getRow());
                }

                // process probe side
                ProbeIterator probeIter = table.getSpilledPartitionProbeSideIter(p);
                BinaryRowData probeNext;
                while ((probeNext = probeIter.next()) != null) {
                    processSortMergeJoinElement2(probeNext);
                }
            }

            // close the HashTable
            closeHashTable();

            // finish build and probe
            sortMergeJoinFunction.endInput(1);
            sortMergeJoinFunction.endInput(2);
            LOG.info("Finish sort merge join for spilled partitions.");
        }
    }

    private void initialSortMergeJoinFunction() throws Exception {
        sortMergeJoinFunction.open(
                true,
                this.getContainingTask(),
                this.getOperatorConfig(),
                (StreamRecordCollector) this.collector,
                this.computeMemorySize(),
                this.getRuntimeContext(),
                this.getMetricGroup());
    }

    private void processSortMergeJoinElement1(RowData rowData) throws Exception {
        if (leftIsBuild) {
            sortMergeJoinFunction.processElement1(rowData);
        } else {
            sortMergeJoinFunction.processElement2(rowData);
        }
    }

    private void processSortMergeJoinElement2(RowData rowData) throws Exception {
        if (leftIsBuild) {
            sortMergeJoinFunction.processElement2(rowData);
        } else {
            sortMergeJoinFunction.processElement1(rowData);
        }
    }

与BinaryHashTable不同，LongHybridHashTable的join逻辑全部是代码生成的，在对应的生成器LongHashJoinGenerator中，可以看到与上文类似的代码，看官可以自行找来读读。