Skip to content

EnsureRequirements Physical Optimization

[[apply]] EnsureRequirements is a physical query optimization (aka physical query preparation rule or simply preparation rule) that QueryExecution uses to optimize the physical plan of a structured query by transforming the following physical operators (up the plan tree):

. Removes two adjacent[ShuffleExchangeExec] physical operators if the child partitioning scheme guarantees the parent's partitioning

. For other non-ShuffleExchangeExec physical operators, <> (possibly adding new physical operators, e.g.[BroadcastExchangeExec] and[ShuffleExchangeExec] for distribution or[SortExec] for sorting)

Technically, EnsureRequirements is just a catalyst/[Catalyst rule] for transforming[physical query plans], i.e. Rule[SparkPlan].

EnsureRequirements is part of preparations batch of physical query plan rules and is executed when QueryExecution is requested for the optimized physical query plan (i.e. in executedPlan phase of a query execution).

[[conf]] EnsureRequirements takes a SQLConf when created.

[source, scala]

val q = ??? // FIXME val sparkPlan = q.queryExecution.sparkPlan

import val plan = EnsureRequirements(spark.sessionState.conf).apply(sparkPlan)

=== [[createPartitioning]] createPartitioning Internal Method


=== [[defaultNumPreShufflePartitions]] defaultNumPreShufflePartitions Internal Method


Enforcing Partition Requirements (Distribution and Ordering) of Physical Operator

  operator: SparkPlan): SparkPlan

ensureDistributionAndOrdering takes the following from the input physical operator:

  •[required partition requirements] for the children

  •[required sort ordering] per the required partition requirements per child

  • child physical plans

NOTE: The number of requirements for partitions and their sort ordering has to match the number and the order of the child physical plans.

ensureDistributionAndOrdering matches the operator's required partition requirements of children (requiredChildDistributions) to the children's[output partitioning] and (in that order):

. If the child satisfies the requested distribution, the child is left unchanged

. For BroadcastDistribution, the child becomes the child of[BroadcastExchangeExec] unary operator for[broadcast hash joins]

. Any other pair of child and distribution leads to[ShuffleExchangeExec] unary physical operator (with proper <> for distribution and with spark.sql.shuffle.partitions number of partitions)

NOTE:[ShuffleExchangeExec] can appear in the physical plan when the children's output partitioning cannot satisfy the physical operator's required child distribution.

If the input operator has multiple children and specifies child output distributions, then the children's[output partitionings] have to be compatible.

If the children's output partitionings are not all compatible, then...FIXME

ensureDistributionAndOrdering <> (only when Adaptive Query Execution is enabled which is not by default).

NOTE: At this point in ensureDistributionAndOrdering the required child distributions are already handled.

ensureDistributionAndOrdering matches the operator's required sort ordering of children (requiredChildOrderings) to the children's[output partitioning] and if the orderings do not match,[SortExec] unary physical operator is created as a new child.

In the end, ensureDistributionAndOrdering sets the new children for the input operator.

ensureDistributionAndOrdering is used when EnsureRequirements physical optimization is executed.

=== [[reorderJoinPredicates]] reorderJoinPredicates Internal Method

[source, scala]

reorderJoinPredicates(plan: SparkPlan): SparkPlan


NOTE: reorderJoinPredicates is used when...FIXME