11

I'm writing on an upcoming blog post of mine on ranking and aggregate window functions, specifically the Segment and Sequence Project iterators. The way I understand it is that Segment identifies rows in a stream that constitute the end/beginning of a group, so the following query:

SELECT ROW_NUMBER() OVER (PARTITION BY someGroup ORDER BY someOrder)

Will use Segment to tell when a row belongs to a different group other than the previous row. The Sequence Project iterator then does the actual row number calculation, based on the output of the Segment iterator's output.

But the following query, using that logic, shouldn't have to include a Segment, because there's no partition expression.

SELECT ROW_NUMBER() OVER (ORDER BY someGroup, someOrder)

However, When I try this hypothesis both these queries use a Segment operator. The only difference is that the second query does not need a GroupBy on the Segment. Doesn't that eliminate the need for a Segment in the first place?

Example

CREATE TABLE dbo.someTable (
    someGroup   int NOT NULL,
    someOrder   int NOT NULL,
    someValue   numeric(8, 2) NOT NULL,
    PRIMARY KEY CLUSTERED (someGroup, someOrder)
);

--- Query 1:
SELECT ROW_NUMBER() OVER (PARTITION BY someGroup ORDER BY someOrder)
FROM dbo.someTable;

--- Query 2:
SELECT ROW_NUMBER() OVER (ORDER BY someGroup, someOrder)
FROM dbo.someTable;
  • 1
    Although there is no partition expression, I guess you are still technically splitting the result set into partitions, albeit only one in this case? – Mark Sinkinson Dec 2 '14 at 13:22
  • The QP shows an empty <GroupBy /> so the segment really does nothing, almost, it does output the segment column to the Sequence Project operator. The reason for the segment operator to be there could be that the Sequence Project operator needs that value to do its work. – Mikael Eriksson Dec 2 '14 at 13:23
  • That's my theory, too. But the optimizer usually does away with these kinds of unneccessary operators, imho.. – Daniel Hutmacher Dec 2 '14 at 13:29
12

I found this 6 year old blog post mentioning the same behavior.

It looks like ROW_NUMBER() always includes a segment operator, whether PARTITION BY is used or not. If I had to guess I would say this is because it makes creating a query plan easier on the engine.

If the segment is needed in most cases, and in the cases where it's not needed it's essentially a zero-cost non-operation, it's a lot simpler to just always include it in the plan when a windowing function is used.

11

According to the showplan.xsd for the execution plan, GroupBy appears without minOccurs or maxOccurs attributes which therefore default to [1..1] making the element compulsory, not necessarily content. The child element ColumnReference of type (ColumnReferenceType) has minOccurs 0 and maxOccurs unbounded [0..*], making it optional, hence the allowed empty element. If you manually attempt to remove the GroupBy and force the plan you get the expected error:

Msg 6965, Level 16, State 1, Line 29
XML Validation: Invalid content. Expected element(s): '{http://schemas.microsoft.com/sqlserver/2004/07/showplan}GroupBy','{http://schemas.microsoft.com/sqlserver/2004/07/showplan}DefinedValues','{http://schemas.microsoft.com/sqlserver/2004/07/showplan}InternalInfo'. Found: element '{http://schemas.microsoft.com/sqlserver/2004/07/showplan}SegmentColumn' instead. Location: /*:ShowPlanXML[1]/*:BatchSequence[1]/*:Batch[1]/*:Statements[1]/*:StmtSimple[1]/*:QueryPlan[1]/*:RelOp[1]/*:SequenceProject[1]/*:RelOp[1]/*:Segment[1]/*:SegmentColumn[1].

Interestingly I found you can manually remove the Segment operator to get a valid plan for forcing which looks like this:

enter image description here

However when you run with that plan (using OPTION ( USE PLAN ... ) ) the Segment Operator magically reappears. Just goes to show the optimizer only takes the XML plans as a rough guide.

My test rig:

USE tempdb
GO
SET NOCOUNT ON
GO
IF OBJECT_ID('dbo.someTable') IS NOT NULL DROP TABLE dbo.someTable
GO
CREATE TABLE dbo.someTable (
    someGroup   int NOT NULL,
    someOrder   int NOT NULL,
    someValue   numeric(8, 2) NOT NULL,
    PRIMARY KEY CLUSTERED (someGroup, someOrder)
);
GO

-- Generate some dummy data
;WITH cte AS (
SELECT TOP 1000 ROW_NUMBER() OVER ( ORDER BY ( SELECT 1 ) ) rn
FROM master.sys.columns c1
    CROSS JOIN master.sys.columns c2
    CROSS JOIN master.sys.columns c3
)
INSERT INTO dbo.someTable ( someGroup, someOrder, someValue )
SELECT rn % 333, rn % 444, rn % 55
FROM cte
GO


-- Try and force the plan
SELECT ROW_NUMBER() OVER (ORDER BY someGroup, someOrder)
FROM dbo.someTable
OPTION ( USE PLAN N'<?xml version="1.0" encoding="utf-16"?>
<ShowPlanXML xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xsd="http://www.w3.org/2001/XMLSchema" Version="1.2" Build="12.0.2000.8" xmlns="http://schemas.microsoft.com/sqlserver/2004/07/showplan">
  <BatchSequence>
    <Batch>
      <Statements>
        <StmtSimple StatementCompId="1" StatementEstRows="1000" StatementId="1" StatementOptmLevel="TRIVIAL" CardinalityEstimationModelVersion="120" StatementSubTreeCost="0.00596348" StatementText="SELECT ROW_NUMBER() OVER (ORDER BY someGroup, someOrder)&#xD;&#xA;FROM dbo.someTable" StatementType="SELECT" QueryHash="0x193176312402B8E7" QueryPlanHash="0x77F1D72C455025A4" RetrievedFromCache="true">
          <StatementSetOptions ANSI_NULLS="true" ANSI_PADDING="true" ANSI_WARNINGS="true" ARITHABORT="true" CONCAT_NULL_YIELDS_NULL="true" NUMERIC_ROUNDABORT="false" QUOTED_IDENTIFIER="true" />
          <QueryPlan DegreeOfParallelism="1" CachedPlanSize="16" CompileTime="0" CompileCPU="0" CompileMemory="88">
            <OptimizerHardwareDependentProperties EstimatedAvailableMemoryGrant="131072" EstimatedPagesCached="65536" EstimatedAvailableDegreeOfParallelism="4" />
            <RelOp AvgRowSize="15" EstimateCPU="8E-05" EstimateIO="0" EstimateRebinds="0" EstimateRewinds="0" EstimatedExecutionMode="Row" EstimateRows="1000" LogicalOp="Compute Scalar" NodeId="0" Parallel="false" PhysicalOp="Sequence Project" EstimatedTotalSubtreeCost="0.00596348">
              <OutputList>
                <ColumnReference Column="Expr1002" />
              </OutputList>
              <SequenceProject>
                <DefinedValues>
                  <DefinedValue>
                    <ColumnReference Column="Expr1002" />
                    <ScalarOperator ScalarString="row_number">
                      <Sequence FunctionName="row_number" />
                    </ScalarOperator>
                  </DefinedValue>
                </DefinedValues>

                <!-- Segment operator completely removed from plan -->
                <!--<RelOp AvgRowSize="15" EstimateCPU="2E-05" EstimateIO="0" EstimateRebinds="0" EstimateRewinds="0" EstimatedExecutionMode="Row" EstimateRows="1000" LogicalOp="Segment" NodeId="1" Parallel="false" PhysicalOp="Segment" EstimatedTotalSubtreeCost="0.00588348">
                  <OutputList>
                    <ColumnReference Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Column="someGroup" />
                    <ColumnReference Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Column="someOrder" />
                    <ColumnReference Column="Segment1003" />
                  </OutputList>
                  <Segment>
                    <GroupBy />
                    <SegmentColumn>
                      <ColumnReference Column="Segment1003" />
                    </SegmentColumn>-->


                    <RelOp AvgRowSize="15" EstimateCPU="0.001257" EstimateIO="0.00460648" EstimateRebinds="0" EstimateRewinds="0" EstimatedExecutionMode="Row" EstimateRows="1000" LogicalOp="Clustered Index Scan" NodeId="0" Parallel="false" PhysicalOp="Clustered Index Scan" EstimatedTotalSubtreeCost="0.00586348" TableCardinality="1000">
                      <OutputList>
                        <ColumnReference Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Column="someGroup" />
                        <ColumnReference Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Column="someOrder" />
                      </OutputList>
                      <IndexScan Ordered="true" ScanDirection="FORWARD" ForcedIndex="false" ForceSeek="false" ForceScan="false" NoExpandHint="false" Storage="RowStore">
                        <DefinedValues>
                          <DefinedValue>
                            <ColumnReference Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Column="someGroup" />
                          </DefinedValue>
                          <DefinedValue>
                            <ColumnReference Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Column="someOrder" />
                          </DefinedValue>
                        </DefinedValues>
                        <Object Database="[tempdb]" Schema="[dbo]" Table="[someTable]" Index="[PK__someTabl__7CD03C8950FF62C1]" IndexKind="Clustered" Storage="RowStore" />
                      </IndexScan>
                    </RelOp>

                <!--</Segment>
                </RelOp>-->
              </SequenceProject>
            </RelOp>

          </QueryPlan>
        </StmtSimple>
      </Statements>
    </Batch>
  </BatchSequence>
</ShowPlanXML>' )

Chop out the XML plan from the test rig and save it as a .sqlplan to view the plan minus the Segment.

PS I wouldn't spend too much time chopping around SQL plans manually as if you know me you would know I regard it as time-eating busy work and something I would never do. Oh hang on!? : )

  • You have far too much time on your hands...Nice work! – Mark Sinkinson Dec 2 '14 at 16:55
  • Agree with Mark. I'm learning stuff I didn't even think to ask for. Thanks! :) – Daniel Hutmacher Dec 2 '14 at 19:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.