关注 spark技术分享,
撸spark源码 玩spark最佳实践

SparkPlan Contract — Physical Operators in Physical Query Plan of Structured Query

SparkPlan Contract — Physical Operators in Physical Query Plan of Structured Query

SparkPlan is the contract of physical operators to build a physical query plan (aka query execution plan).

SparkPlan contract requires that a concrete physical operator implements doExecute method.

doExecute allows a physical operator to describe a distributed computation (that is a runtime representation of the operator in particular and a structured query in general) as an RDD of internal binary rows, i.e. RDD[InternalRow], and thus execute.

Table 1. SparkPlan’s Extension Hooks
Name Description

doExecuteBroadcast

By default reports a UnsupportedOperationException.

Executed exclusively as part of executeBroadcast to return the result of a structured query as a broadcast variable.

doPrepare

Prepares a physical operator for execution

Executed exclusively as part of prepare and is supposed to set some state up before executing a query (e.g. BroadcastExchangeExec to broadcast a relation asynchronously or SubqueryExec to execute a child operator)

requiredChildDistribution

The required partition requirements (aka child output distributions) of the input data, i.e. how children physical operators’ output is split across partitions.

Defaults to a UnspecifiedDistribution for all of the child operators.

Used exclusively when EnsureRequirements physical query plan optimization is executed (and enforces partition requirements).

requiredChildOrdering

Specifies required sort ordering for each partition requirement (from children operators)

Defaults to no sort ordering for all of the physical operator’s children.

Used exclusively when EnsureRequirements physical query plan optimization is executed (and enforces partition requirements).

SparkPlan is a recursive data structure in Spark SQL’s Catalyst tree manipulation framework and as such represents a single physical operator in a physical execution query plan as well as a physical execution query plan itself (i.e. a tree of physical operators in a query plan of a structured query).

spark sql SparkPlan webui physical plan.png
Figure 1. Physical Plan of Structured Query (i.e. Tree of SparkPlans)
Note
A structured query can be expressed using Spark SQL’s high-level Dataset API for Scala, Java, Python, R or good ol’ SQL.

A SparkPlan physical operator is a Catalyst tree node that may have zero or more child physical operators.

Note
A structured query is basically a single SparkPlan physical operator with child physical operators.
Note
Spark SQL uses Catalyst tree manipulation framework to compose nodes to build a tree of (logical or physical) operators that, in this particular case, is composing SparkPlan physical operator nodes to build the physical execution plan tree of a structured query.

The entry point to Physical Operator Execution Pipeline is execute.

When executed, SparkPlan executes the internal query implementation in a named scope (for visualization purposes, e.g. web UI) that triggers prepare of the children physical operators first followed by prepareSubqueries and finally doPrepare methods. After subqueries have finished, doExecute method is eventually triggered.

spark sql SparkPlan execute.png
Figure 2. SparkPlan’s Execution (execute Method)

The result of executing a SparkPlan is an RDD of internal binary rows, i.e. RDD[InternalRow].

Note
Executing a structured query is simply a translation of the higher-level Dataset-based description to an RDD-based runtime representation that Spark will in the end execute (once an Dataset action is used).
Caution
FIXME Picture between Spark SQL’s Dataset ⇒ Spark Core’s RDD

SparkPlan has access to the owning SparkContext.

Note

execute is called when QueryExecution is requested for the RDD that is Spark Core’s physical execution plan (as a RDD lineage) that triggers query execution (i.e. physical planning, but not execution of the plan) and could be considered execution of a structured query.

The could part above refers to the fact that the final execution of a structured query happens only when a RDD action is executed on the RDD of a structured query. And hence the need for Spark SQL’s high-level Dataset API in which the Dataset operators simply execute a RDD action on the corresponding RDD. Easy, isn’t it?

Tip

Use explain operator to see the execution plan of a structured query.

You may also access the execution plan of a Dataset using its queryExecution property.

The SparkPlan contract assumes that concrete physical operators define doExecute method (with optional hooks like doPrepare) which is executed when the physical operator is executed.

spark sql SparkPlan execute pipeline.png
Figure 3. SparkPlan.execute — Physical Operator Execution Pipeline

SparkPlan has the following final methods that prepare execution environment and pass calls to corresponding methods (that constitute SparkPlan Contract).

Table 2. SparkPlan’s Final Methods
Name Description

execute

“Executes” a physical operator (and its children) that triggers physical query planning and in the end generates an RDD of internal binary rows (i.e. RDD[InternalRow]).

Used mostly when QueryExecution is requested for the RDD-based runtime representation of a structured query (that describes a distributed computation using Spark Core’s RDD).

Internally, execute first prepares the physical operator for execution and eventually requests it to doExecute.

Note
Executing doExecute in a named scope happens only after the operator is prepared for execution followed by waiting for any subqueries to finish.

executeQuery

Executes a physical operator in a single RDD scope, i.e. all RDDs created during execution of the physical operator have the same scope.

executeQuery executes the input query after the following methods (in order):

Note

executeQuery is used when:

prepare

Prepares a physical operator for execution

prepare is used mainly when a physical operator is requested to execute a structured query

prepare is also used recursively for every child physical operator (down the physical plan) and when a physical operator is requested to prepare subqueries.

Note
prepare is idempotent, i.e. can be called multiple times with no change to the final result. It uses prepared internal flag to execute the physical operator once only.

Internally, prepare calls doPrepare of its children before prepareSubqueries and doPrepare.

executeBroadcast

Table 3. Physical Query Operators / Specialized SparkPlans
Name Description

BinaryExecNode

Binary physical operator with two child left and right physical operators

LeafExecNode

Leaf physical operator with no children

By default, the set of all attributes that are produced is exactly the set of attributes that are output.

UnaryExecNode

Unary physical operator with one child physical operator

Note
The naming convention for physical operators in Spark’s source code is to have their names end with the Exec prefix, e.g. DebugExec or LocalTableScanExec that is however removed when the operator is displayed, e.g. in web UI.
Table 4. SparkPlan’s Internal Properties (e.g. Registries, Counters and Flags)
Name Description

prepared

Flag that controls that prepare is executed only once.

subexpressionEliminationEnabled

Flag that controls whether the subexpression elimination optimization is enabled or not.

Used when the following physical operators are requested to execute (i.e. describe a distributed computation as an RDD of internal rows):

Caution
FIXME SparkPlan is Serializable. Why? Is this because Dataset.cache persists executed query plans?

Decoding Byte Arrays Back to UnsafeRows — decodeUnsafeRows Method

Caution
FIXME

Compressing Partitions of UnsafeRows (to Byte Arrays) After Executing Physical Operator — getByteArrayRdd Internal Method

Caution
FIXME

resetMetrics Method

resetMetrics takes metrics and request them to reset.

Note
resetMetrics is used when…​FIXME

prepareSubqueries Method

Caution
FIXME

executeToIterator Method

Caution
FIXME

executeCollectIterator Method

executeCollectIterator…​FIXME

Note
executeCollectIterator is used when…​FIXME

Preparing SparkPlan for Query Execution — executeQuery Final Method

executeQuery executes the input query in a named scope (i.e. so that all RDDs created will have the same scope for visualization like web UI).

Internally, executeQuery calls prepare and waitForSubqueries followed by executing query.

Note
executeQuery is executed as part of execute, executeBroadcast and when CodegenSupport-enabled physical operator produces a Java source code.

Broadcasting Result of Structured Query — executeBroadcast Final Method

executeBroadcast returns the result of a structured query as a broadcast variable.

Internally, executeBroadcast calls doExecuteBroadcast inside executeQuery.

Note
executeBroadcast is called in BroadcastHashJoinExec, BroadcastNestedLoopJoinExec and ReusedExchangeExec physical operators.

Performance Metrics — metrics Method

metrics returns the SQLMetrics by their names.

By default, metrics contains no SQLMetrics (i.e. Map.empty).

Note
metrics is used when…​FIXME

Taking First N UnsafeRows — executeTake Method

executeTake gives an array of up to n first internal rows.

spark sql SparkPlan executeTake.png
Figure 4. SparkPlan’s executeTake takes 5 elements

Internally, executeTake gets an RDD of byte array of n unsafe rows and scans the RDD partitions one by one until n is reached or all partitions were processed.

executeTake runs Spark jobs that take all the elements from requested number of partitions, starting from the 0th partition and increasing their number by spark.sql.limit.scaleUpFactor property (but minimum twice as many).

Note
executeTake uses SparkContext.runJob to run a Spark job.

In the end, executeTake decodes the unsafe rows.

Note
executeTake gives an empty collection when n is 0 (and no Spark job is executed).
Note
executeTake may take and decode more unsafe rows than really needed since all unsafe rows from a partition are read (if the partition is included in the scan).

Note

executeTake is used when:

Executing Physical Operator and Collecting Results — executeCollect Method

executeCollect executes the physical operator and compresses partitions of UnsafeRows as byte arrays (that yields a RDD[(Long, Array[Byte])] and so no real Spark jobs may have been submitted).

executeCollect runs a Spark job to collect the elements of the RDD and for every pair in the result (of a count and bytes per partition) decodes the byte arrays back to UnsafeRows and stores the decoded arrays together as the final Array[InternalRow].

Note
executeCollect runs a Spark job using Spark Core’s RDD.collect operator.
Note
executeCollect returns Array[InternalRow], i.e. keeps the internal representation of rows unchanged and does not convert rows to JVM types.
Note

executeCollect is used when:

executeCollectPublic Method

executeCollectPublic…​FIXME

Note
executeCollectPublic is used when…​FIXME

newPredicate Method

newPredicate…​FIXME

Note
newPredicate is used when…​FIXME

Waiting for Subqueries to Finish — waitForSubqueries Method

waitForSubqueries requests every ExecSubqueryExpression in runningSubqueries to updateResult.

Note
waitForSubqueries is used exclusively when a physical operator is requested to prepare itself for query execution (when it is executed or requested to executeBroadcast).

Output Data Partitioning Requirements — outputPartitioning Method

outputPartitioning specifies the output data partitioning requirements, i.e. a hint for the Spark Physical Optimizer for the number of partitions the output of the physical operator should be split across.

outputPartitioning defaults to a UnknownPartitioning (with 0 partitions).

Note

outputPartitioning is used when:

Output Data Ordering Requirements — outputOrdering Method

outputOrdering specifies the output data ordering requirements of the physical operator, i.e. a hint for the Spark Physical Optimizer for the sorting (ordering) of the data (within and across partitions).

outputOrdering defaults to no ordering (Nil).

Note

outputOrdering is used when:

赞(0) 打赏
未经允许不得转载:spark技术分享 » SparkPlan Contract — Physical Operators in Physical Query Plan of Structured Query
分享到: 更多 (0)

关注公众号:spark技术分享

联系我们联系我们

觉得文章有用就打赏一下文章作者

支付宝扫一扫打赏

微信扫一扫打赏