A partitioned table is really more like a collection of individual tables stitched together. So your in example of clustering by IncidentKey
and partition by IncidentDate
, say that the partitioning function splits the tables into two partitions so that 1/1/2010 is in partition 1 and 7/1/2010 is partition two. The data will be layed out on disk as:
Partition 1:
IncidentKey Date
ABC123 1/1/2010
ABC123 1/1/2011
XYZ999 1/1/2010
Partition 2:
IncidentKey Date
ABC123 7/1/2010
XYZ999 7/1/2010
At a low level there really are two, distinct rowsets. Is the query processor that gives the illusion of a single table by creating plans that seek, scan and update all rowsets together, as one.
Any row in any non-clustered index will have have the clustered index key to which it corresponds, say ABC123,7/1/2010
. Since the clustered index key always contains the partitioning key column, the engine will always know in what partition (rowset) of the clustered index to search for this value (in this case, in partition 2).
Now whenever you're dealing with partitioning you must consider if your NC indexes will be aligned (NC index is partitioned exactly the same as the clustered index) or non-aligned (NC index is non-partitioned, or partitioned differently from clustered index). Non-aligned indexes are more flexible, but they have some drawbacks:
Using aligned indexes solves these issues, but brings its own set of problems, because this physical, storage design, option ripples into the data model:
- aligned indexes mean unique constrains can no longer be created/enforced (except for the partitioning column)
- all foreign keys referencing the partitioned table must include the partitioning key in the relation (since the partitioning key is, due to alignment, in every index), and this in turn requires that all tables referencing the partitioned table contain partitioning key column value. Think Orders->OrderDetails, if Orders have OrderID but is partitioned by OrderDate, then OrderDetails must contain not only OrderID, but also OrderDate, in order to properly declare the foreign key constraint.
These effects I found seldom called out at the beginning of a project that deploys partitioning, but they exists and have serious consequences.
If you think aligned indexes are a rare or extreme case, then consider this: in many cases the cornerstone of ETL and partitioning solutions is the fast switch in of staging tables. Switch in operations require aligned indexes.
Oh, one more thing: all my argument about foreign keys and the ripple effect of adding the partitioning column value to other tables applies equally to joins.
The clustered index is in fact the table. On the assumption that your primary key is clustered then I would create a clustered primary key with page level compression rather than trying to do it in two steps.
-- Add primary key
ALTER TABLE dbo.TableName
ADD CONSTRAINT PK_TableName
PRIMARY KEY CLUSTERED (<Columns>)
WITH (DATA_COMPRESSION = PAGE)
;
I would also copy about 100k rows to a temporary (temporary physical not #temporary) table and run some tests. Try running compression first, clustered key first, try doing them as one step. See what runs fastest. I would guess it will be one step personally :).
Best Answer
The main question here is, do you want to save space & memory usage at the cost of more cpu cycles? This is something we will not be able to answer for you.
Since you are able to compress the clustered indexes, no other restrictions noted in Considerations for When You Use Row and Page Compression are possible limitations.
You could always run the sp_estimate_data_compression_savings procedure to calculate the gains you could get from compressing the nonclustered indexes. See here for a solution for every table in your data warehouse.
Speaking of data warehouses, you could also take a look at (clustered / nonclustered) columnstore indexes and see if they are a fit for your (fact) tables.