SQL Server Performance With Large Query

Question

Hi everyone I have a couple of queries for some reports in which each query is pulling Data from 35+ tables. Each Table has almost 100K records. All the Queries are Union ALL for Example

;With CTE
AS
(
Select col1, col2, col3 FROM Table1 WHERE Some_Condition
UNION ALL 
Select col1, col2, col3 FROM Table2 WHERE Some_Condition
UNION ALL 
Select col1, col2, col3 FROM Table3 WHERE Some_Condition
UNION ALL 
Select col1, col2, col3 FROM Table4 WHERE Some_Condition
.
.
. And so on 
)
SELECT col1, col2, col3 FROM CTE
ORDER BY col3 DESC

So far I have only tested this query on Dev Server and I can see It takes its time to get the results. All of these 35+ tables are not related with each other and this is the only way I can think of to get all the Desired Data in result set.

1. Is there a better way to do this kind of query ??
   

2. If this is the only way to go for this kind of query how can I improve the performance for this Query by making any changes if possible??
   

My Opinion
I Dont mind having a few dirty-reads in this report. I was thinking of using Query hints with nolock or Transaction Isolation Level set to READ UNCOMMITED.

Will any of this help ???

Edit
Every Table has 5-10 Bit columns and a Corresponding Date column to each Bit Column and my condition for each SELECT Statement is something like

WHERE BitColumn = 1 AND DateColumn IS NULL 

Suggestion By Peers

Filtered Index

CREATE NONCLUSTERED INDEX IX_Table_Column
ON TableName(BitColumn)
WHERE BitColum = 1

Filtered Index with Included Column

CREATE NONCLUSTERED INDEX fIX_IX_Table_Column
ON TableName(BitColumn)
INCLUDE (DateColumn)
WHERE DateColumn IS NULL

Is this the best way to go ? or any suggestions please ???

Answer 1

There are lots of things that can be done to make it faster. If I assume you need to do these UNIONs, then you can speed up the query by :

1. Caching the results, for example,
   • Can you create an indexed view from the whole statement ? Or there are lots of different WHERE conditions, so there'd be lots of indexed views ? But know that this will slow down modifications (INSERT, etc.) for those tables
   • Can you cache it in a different way ? Maybe in the mid layer ?
   • Can it be recalculated in advance ?
2. Make a covering index. Leading columns are columns form WHERE and then all other columns from the query as included columns
   • Note that a covering index can be also filtered but filtered index isn't used if the WHERE in the query will have variables / parameters and they can potentially have the value that is not covered by the filtered index (i.e., the row isn't covered)
3. ORDER BY will cause sort
   • If you can cache it, then it's fine - no sort will be needed (it's cached sorted)
   • Otherwise, sort is CPU bound (and I/O bound if not in memory). To speed it up, do you use fast collation ? The performance difference between the slowest and fastest collation can be even 3 times. For example, SQL_EBCDIC280_CP1_CS_AS, SQL_Latin1_General_CP1251_CS_AS, SQL_Latin1_General_CP1_CI_AS are one of the fastest collations. However, it's hard to make recommendations if I don't know the collation characteristics you need
4. Network
   • 'network packet size' for the connection that does the SELECT should be the maximum value possible - 32,767 bytes if the result set (number of rows) will be big. This can be set on the client side, e.g., if you use .NET and SqlConnection in the connection string. This will minimize CPU overhead when sending data from the SQL Server and will improve performance on both side - client and server. This can boost performance even by tens of percents if the network was the bottleneck
   • Use shared memory endpoint if the client is on the SQL Server; otherwise TCP/IP for the best performance
5. General things
   • As you said, using isolation level read uncommmitted will improve the performance

...

Probably you can't do changes beyond rewriting the query, etc. but just in case, adding more memory in case it isn't sufficient now, or using SQL Server 2014 in memory features :-), ... would surely help.

There are way too many things that could be tuned but it's hard to point out the key ones if the question isn't very specific.

Hope this helps a bit

Answer 2

well you haven't give any statistics or sample run time of any execution so it is not possible to guess what is slow and is it really slow. how much data is in the result set? it might be just retrieving 100K rows as in result is just taking its time. if the result set of 10000 rows is taking 5 minute, yes definitely something can be looked at. so if you have sample query, number of rows in result and how much time it took for couple of execution with different where conditions, post that. it will help us to compare results.

BTW, do not use CTE just use regular inner and outer query select. make sure the Temp DB is configured properly. LDF and MDF is not default configured for 10% increase. by certain try and error you will come to know how much log and temp DB is increased for verity of range queries and based on that you should set the initial and increment size of the MDF and LDF of temp DB. for the Covered filter index the include column should be col1, col2 and co3 not column Date unless Date is also in select list.

how frequently the data in original 35 tables get updated? if max once per day or if they all get updates almost same time then Indexed-Views can be a possible solution. but if original tables getting updates more than once a day or they gets updates anytime and no where they are in same line then do no think about Indexed-View.

if disk space is not an issue as a last resort try and test performance using trigger on each 35 table. create new table to hold final results as you are expecting from this select query. create insert/update/delete trigger on each 35 table where you check the conditions inside trigger and if yes then only copy the same insert/update/delete to new table. yes you will need a column in new table that identifies which data coming from which table. because Date is Null-Able column you do not get full advantage of Index on that Column as "mostly you are looking for WHERE Date is NULL". in the new Table only query you always do is where Date is NULL then do not even bother to create that column just create BIT columns and other col1, col2, col3 etc... if you give real example of your query and explain the actual tables, other details can be workout later.

Answer 3

The query hints or the Isolation Level are only going to help you in case of any blocking occurs. If you dont mind dirty reads and there are locks during the execution it could be a good idea.

The key question is how many data fits the Where clausule you need to use (WHERE BitColumn = 1 AND DateColumn IS NULL) If the subset filtered by that is small compared with the total number of rows, then use an index on both columns, BitColum and DateColumn, including the columns in the select clausule to avoid "Page Lookup" operations in your query plan.

CREATE NONCLUSTERED INDEX IX_[Choose an IndexName]
ON TableName(BitColumn, DateColumn)
INCLUDE (col1, col2, col3)

Of course the space needed for that covered-filtered index depends on the datatype of the fields involved and the number of rows that satisfy WHERE BitColumn = 1 AND DateColumn IS NULL.

After that I recomend to use a View instead of a CTE:

CREATE VIEW [Choose a ViewName]
AS
(
Select col1, col2, col3 FROM Table1 WHERE Some_Condition
UNION ALL 
Select col1, col2, col3 FROM Table2 WHERE Some_Condition
.
.
.
)

By doing that, your query plan should look like 35 small index scans, but if most of the data satisfies the where clausule of your index, the performance is going to be similar to scan the 35 source tables and the solution won't worth it.

But You say "Every Table has 5-10 Bit columns and a Corresponding Date column.." then I think is not going to be a good idea to make an index per bit colum. If you need to filter by using diferent BitColums and Different DateColums, use a compute column in your table:

ALTER TABLE Table1 ADD ComputedFilterFlag AS
CAST(
    CASE WHEN BitColum1 = 1 AND DateColumn1 IS NULL THEN 1 ELSE 0 END +
    CASE WHEN BitColum2 = 1 AND DateColumn2 IS NULL THEN 2 ELSE 0 END +
    CASE WHEN BitColum3 = 1 AND DateColumn3 IS NULL THEN 4 ELSE 0 END
AS tinyint)

I recomend you use the value 2^(X-1) for conditionX(BitColumnX=1 and DateColumnX IS NOT NULL). It is going to allow you to filter by using any combination of that criteria. By using value 3 you can locate all rows that accomplish: Bit1, Date1 and Bit2, Date2 condition. Any condition combination has its corresponding ComputedFilterFlag value because the ComputedFilterFlag acts as a bitmap of conditions. If you heve less than 8 diferents filters you should use tinyint to save space in the index and decrease the IO operations needed.

Then use an Index over ComputedFilterFlag colum:

CREATE NONCLUSTERED INDEX IX_[Choose an IndexName]
ON TableName(ComputedFilterFlag)
INCLUDE (col1, col2, col3)

And create the view:

CREATE VIEW [Choose a ViewName]
AS
(
Select col1, col2, col3 FROM Table1 WHERE ComputedFilterFlag IN [Choose the Target Filter Value set]--(1, 3, 5, 7)
UNION ALL 
Select col1, col2, col3 FROM Table2 WHERE ComputedFilterFlag IN [Choose the Target Filter Value set]--(1, 3, 5, 7)
.
.
.
)

By doing that, your index coveres all the conditions and your query plan should look like 35 small index seeks.

But this is a tricky solution, may be a refactoring in your table schema could produce simpler and faster results.

Answer 4

You'll never get real time results from a union all query over many tables but I can tell you how I got a little speed out of a similar situation. Hopefully this will help you out.

You can actually run all of them at once with a little bit coding and ingenuity.

You create a global temporary table instead of a common table expression and don't put any keys on the global temporary table it will just slow things down. Then you start all the individual queries which insert into the global temporary table. I've done this a hundred or so times manually and it's faster than a union query because you get a query running on each cpu core. The tricky part is the mechanism to determine when the individual queries have finished your on your own for that piece hence I do these manually.