一. 前言
在presto中,partition join和broadcast join的差异主要在执行计划的方式上。如果join的一侧直接以REPLICATE exchange的方式将数据传送到另外一侧,则为broadcast join,如下图1的执行计划所示。 如果join 两侧的数据经过REPARTITION后再进行exchange到下游执行join操作,则为partition join,如下图2。
图1: broadcast join执行计划样例
图2: partition join执行计划样例
在presto中,可以通过设置join_distribution_type参数强制使用任意一种join方式。
在presto中,生成partition join和broadcast join的执行计划在AddExchanges::visitJoin中实现的。
public PlanWithProperties visitJoin(JoinNode node, PreferredProperties preferredProperties)
{
....
if (distributionType == JoinNode.DistributionType.REPLICATED) {
// 生成broadcast join执行计划
planReplicatedJoin(node, left);
}
else {
// 生成partition join执行计划
return planPartitionedJoin(node, leftSymbols, rightSymbols);
}
}
二. broadcast join
private PlanWithProperties planReplicatedJoin(JoinNode node, PlanWithProperties left)
{
PlanWithProperties right = node.getRight().accept(this, PreferredProperties.any());
// 如果左表已经是单节点运行的任务了(比如coordaintor的aggregate), 则右表添加ExchangeNode.Type.GATHER的exchange 汇聚到左表执行join操作
if (left.getProperties().isSingleNode()) {
if (!right.getProperties().isSingleNode() ||
(!isColocatedJoinEnabled(session) && hasMultipleSources(left.getNode(), right.getNode()))) {
right = withDerivedProperties(
gatheringExchange(idAllocator.getNextId(), REMOTE, right.getNode()),
right.getProperties());
}
}
else {
// 右表直接以ExchangeNode.Type.REPLICATE的形式复制到左表,实现数据广播
right = withDerivedProperties(
replicatedExchange(idAllocator.getNextId(), REMOTE, right.getNode()),
right.getProperties());
}
return buildJoin(node, left, right, JoinNode.DistributionType.REPLICATED);
}
三. partition join
partition join的实现比broadcast join复杂一些,因为涉及到以哪一列进行hash的问题。其实现的过程如以下代码所示:
private PlanWithProperties planPartitionedJoin(JoinNode node, List<Symbol> leftSymbols, List<Symbol> rightSymbols, PlanWithProperties left)
{
SetMultimap<Symbol, Symbol> rightToLeft = createMapping(rightSymbols, leftSymbols);
SetMultimap<Symbol, Symbol> leftToRight = createMapping(leftSymbols, rightSymbols);
PlanWithProperties right;
// 左表的join列且是分区列
if (left.getProperties().isNodePartitionedOn(leftSymbols) && !left.getProperties().isSingleNode()) {
// 找到左表join列对应的右表列,对该列进行hash Partitioning
Partitioning rightPartitioning = left.getProperties().translate(createTranslator(leftToRight)).getNodePartitioning().get();
right = node.getRight().accept(this, PreferredProperties.partitioned(rightPartitioning));
if (!right.getProperties().isCompatibleTablePartitioningWith(left.getProperties(), rightToLeft::get, metadata, session)) {
// 如下一行代码会添加一个ExchangeNode.Type.REPARTITION的exchage,hash的列为右表的join列,从而实现对右表进行REPARTITION到下游进行join操作
right = withDerivedProperties(
partitionedExchange(idAllocator.getNextId(), REMOTE, right.getNode(), new PartitioningScheme(rightPartitioning, right.getNode().getOutputSymbols())),
right.getProperties());
}
}
else {
right = node.getRight().accept(this, PreferredProperties.partitioned(ImmutableSet.copyOf(rightSymbols)));
// 如下的代码用意和上边相同,只是操作的左表,尝试将左表进行REPARTITION到下游进行join操作
if (right.getProperties().isNodePartitionedOn(rightSymbols) && !right.getProperties().isSingleNode()) {
Partitioning leftPartitioning = right.getProperties().translate(createTranslator(rightToLeft)).getNodePartitioning().get();
left = withDerivedProperties(
partitionedExchange(idAllocator.getNextId(), REMOTE, left.getNode(), new PartitioningScheme(leftPartitioning, left.getNode().getOutputSymbols())),
left.getProperties());
}
else {
// 如下是处理无分区列的情况,左右表都按照自己Join列进行Hash partition,然后exchange到下游执行join 操作
left = withDerivedProperties(
partitionedExchange(idAllocator.getNextId(), REMOTE, left.getNode(), leftSymbols, Optional.empty()),
left.getProperties());
right = withDerivedProperties(
partitionedExchange(idAllocator.getNextId(), REMOTE, right.getNode(), rightSymbols, Optional.empty()),
right.getProperties());
}
}
....
return buildJoin(node, left, right, JoinNode.DistributionType.PARTITIONED);
}