Sort order specified in primary key, yet sorting is executed on SELECT

For a non partitioned table I get the following plan

Plan 1

There is a single seek predicate on Seek Keys[1]: Prefix: DeviceId, SensorId = (3819, 53), Start: Date < 1339225010.

Meaning that SQL Server can perform an equality seek on the first two columns and then begin a range seek starting at 1339225010 and ordered FORWARD (as the index is defined with [Date] DESC)

The TOP operator will stop requesting more rows from the seek after the first row is emitted.

When I create the partition scheme and function

AS RANGE LEFT FOR VALUES (1000, 1339225009 ,1339225010 , 1339225011);

And populate the table with the following data

INSERT INTO [dbo].[SensorValues]    
/*500 rows matching date and SensorId, DeviceId predicate*/
SELECT TOP (500) 3819,53,1, ROW_NUMBER() OVER (ORDER BY (SELECT 0))           
FROM master..spt_values
/*700 rows matching date but not SensorId, DeviceId predicate*/
SELECT TOP (700) 3819,52,1, ROW_NUMBER() OVER (ORDER BY (SELECT 0))           
FROM master..spt_values
/*1100 rows matching SensorId, DeviceId predicate but not date */
SELECT TOP (1100) 3819,53,1, ROW_NUMBER() OVER (ORDER BY (SELECT 0)) + 1339225011      
FROM master..spt_values

The plan on SQL Server 2008 looks as follows.

Plan 2

The actual number of rows emitted from the seek is 500. The plan shows seek predicates

Seek Keys[1]: Start: PtnId1000 <= 2, End: PtnId1000 >= 1, 
Seek Keys[2]: Prefix: DeviceId, SensorId = (3819, 53), Start: Date < 1339225010

Indicating it is using the skip scan approach described here

the query optimizer is extended so that a seek or scan operation with one condition can be done on PartitionID (as the logical leading column) and possibly other index key columns, and then a second-level seek, with a different condition, can be done on one or more additional columns, for each distinct value that meets the qualification for the first-level seek operation.

This plan is a serial plan and so for the specific query you have it seems that if SQL Server ensured that it processed the partitions in descending order of date that the original plan with the TOP would still work and it could stop processing after the first matching row was found rather than continuing on and outputting the remaining 499 matches.

In fact the plan on 2005 looks like it does take that approach

Plan on 2005

I'm not sure if it is straight forward to get the same plan on 2008 or maybe it would need an OUTER APPLY on sys.partition_range_values to simulate it.

A lot of people believe that a clustered index guarantees a sort order on output. But that's not what it does; it guarantees a storage order on disk.

See, for example, this blog post, a follow-up, and this longer discussion.

I'm speculating that the SORT is needed because of parallel plan. I base this on some dim and distant blog article: but I found this on MSDN which may or may not justify this

So, try with MAXDOP 1 and see what happens...

Also hinted at in @sql kiwi's blog post on Simple Talk under "Exchange Operator" I think. And "DOP dependence" here