Hcfyk

Consider this table:

create table Books
(
 Id bigint not null primary key identity(1, 1),
 UniqueToken varchar(100) not null,
 [Text] nvarchar(max) not null
)

Let's imagine that we have over 100,000 books in this table.

Now we're given a 10,000 books data to insert into this table, some of which are duplicate. So we need to filter duplicates first, and then insert new books.

One way to check for the duplicates is this way:

select UniqueToken
from Books 
where UniqueToken in 
(
 'first unique token',
 'second unique token'
 -- 10,000 items here
)

Does the existence of Text column affect this query's performance? If so, how can we optimized it?

P.S.
I have the same structure, for some other data. And it's not performing well. A friend told me that I should break my table into two tables as follow:

create table BookUniqueTokens 
(
 Id bigint not null primary key identity(1, 1),
 UniqueToken varchar(100)
)

create table Books
(
 Id bigint not null primary key,
 [Text] nvarchar(max)
)

And I have to do my duplicate finding algorithm on the first table only, and then insert data into both of them. This way he claimed performance gets way better, because tables are physically separate. He claimed that [Text] column affects any select query on the UniqueToken column.

Examples

Consider your query with 8 filter predicates in your IN clause on a dataset of 10K records.

select UniqueToken
from Books 
where UniqueToken in 
(
 'Unique token 1',
 'Unique token 2',
 'Unique token 3',
 'Unique token 4',
 'Unique token 5',
 'Unique token 6',
 'Unique token 9999',
 'Unique token 5000'
 -- 10,000 items here
);

A clustered index scan is used, there are no other indexes present on this test table

With a data size of 216 Bytes.

You should also note how even with 8 records, the OR filters are stacking up.

The reads that happened on this table:

Credits to statisticsparser.

When you include the Text column in the select part of your query, the actual data size changes drastically:

select UniqueToken,Text
from Books 
where UniqueToken in 
(
 'Unique token 1',
 'Unique token 2',
 'Unique token 3',
 'Unique token 4',
 'Unique token 5',
 'Unique token 6',
 'Unique token 9999',
 'Unique token 5000'
 -- 10,000 items here
);

Again, the Clustered index Scan with a residual predicate:

But with a dataset of 32KB.

As there are almost 1000 lob logical reads:

Now, when we create the two tables in question, and fill them up with the same 10k records

Executing the same select without Text. Remember that we had 99 Logical reads when using the Books Table.

select UniqueToken
from BookUniqueTokens 
where UniqueToken in 
(
 'Unique token 1',
 'Unique token 2',
 'Unique token 3',
 'Unique token 4',
 'Unique token 5',
 'Unique token 6',
 'Unique token 9999',
 'Unique token 5000'
 -- 10,000 items here
)

The reads on BookUniqueTokens are lower, 67 instead of 99.

We can track that back to the pages in the original Books table and the pages in the new table without the Text.

Original Books table:

New BookUniqueTokens table

So, all the pages + (2 overhead pages?) are read from the clustered index.

Why is there a difference, and why is the difference not bigger? After all the datasize difference is huge (Lob data <> No Lob data)

Books Data space

BooksWithText Data space

The reason for this is ROW_OVERFLOW_DATA.

When data gets bigger than 8kb the data is stored as ROW_OVERFLOW_DATA on different pages.

Ok, if lob data is stored on different pages, why are the page sizes of these two clustered indexes not the same?

Due to the 24 byte pointer added to the Clustered index to track each of these pages.
After all, sql server needs to know where it can find the lob data.

Source

To answer your questions

He claimed that [Text] column affects any select query on the
UniqueToken column.

And

Does the existence of Text column affect this query's performance? If
so, how can we optimized it?

If the data stored is actually Lob Data, and the Query provided in the answer is used

It does bring some overhead due to the 24 byte pointers.

Depending on the executions / min not being crazy high, I would say that this is negligible, even with 100K records.

Remember that this overhead only happens if an index that includes Text is used, such as the clustered index.

But, what if the clustered index scan is used, and the lob data does not exceed 8kb?

If the data does not exceed 8kb, and you have no index on UniqueToken,the overhead could be bigger . even when not selecting the Text column.

Logical reads on 10k records when Text is only 137 characters long (all records):

Table 'Books2'. Scan count 1, logical reads 419

Due to all this extra data being on the clustered index pages.

Again, an index on UniqueToken (Without including the Text column) will resolve this issue.

As pointed out by @David Browne - Microsoft, you could also store the data off row, as to not add this overhead on the Clustered index when not selecting this Text Column.

Also, if you do want the text stored off-row, you can force that
without using a separate table. Just set the 'large value types out of
row' option with sp_tableoption.
docs.microsoft.com/en-us/sql/relational-databases

TL;DR

Based on the query given, indexing UniqueToken without including TEXT should resolve your troubles.
Additionally, I would use a temporary table or table type to do the filtering instead of the IN statement.

EDIT:

yes there is a nonclustered index on UniqueToken

Your example query is not touching the Text column, and based on the query this should be a covering index.

If we test this on the three tables we previously used (UniqueToken + Lob data, Solely UniqueToken, UniqueToken + 137 Char data in nvarchar(max) column)

CREATE INDEX [IX_Books_UniqueToken] ON Books(UniqueToken);
CREATE INDEX [IX_BookUniqueTokens_UniqueToken] ON BookUniqueTokens(UniqueToken);
CREATE INDEX [IX_Books2_UniqueToken] ON Books2(UniqueToken);

The reads remain the same for these three tables, because the nonclustered index is used.

Table 'Books'. Scan count 8, logical reads 16, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Table 'BookUniqueTokens'. Scan count 8, logical reads 16, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Table 'Books2'. Scan count 8, logical reads 16, physical reads 5, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Additional details

by @David Browne - Microsoft

Also, if you do want the text stored off-row, you can force that
without using a separate table. Just set the 'large value types out of
row' option with sp_tableoption.
docs.microsoft.com/en-us/sql/relational-databases/

Remember that you have to rebuild your indexes for this to take effect on already populated data.

By @Erik Darling

On

• MAX Data Types Do WHAT?

Filtering on Lob data sucks.

• Memory Grants and Data Size

Your memory grants might go through the roof when using bigger datatypes, impacting performance.

Technically speaking, anything that takes up more space on a data page such that the data requires more data pages will decrease performance, even if by such a small amount that it cannot be easily measured. But more data pages means more operations required to read more pages, and more memory required to hold more data pages, and so on.

So, if you are scanning a heap or index, then the presence of an NVARCHAR(MAX) column can impact performance, even if not selecting it. For example, if you have 5000 - 7000 bytes per row, then in the schema shown in the question, that would be stored in-row, creating the need for more data pages. Yet 8100 bytes (roughly) or more will guarantee that the data is stored off-row with just a pointer to the LOB page(s), so that wouldn't be so bad.

But, in your case, since you mentioned having a non-clustered index on UniqueToken, then it really shouldn't matter nearly as much (or even at all), if there is an NVARCHAR(MAX) column with 5000-7000 bytes (causing 1 page per row) because the query should be looking at the index which only has the Id and UniqueToken columns in it. And, the operation should be doing a seek instead of a scan, so not reading all data pages in the index.

Final consideration: unless you have really old hardware (meaning, no RAM and/or other processes hogging disk/CPU/RAM, in which case most queries would be affected, not just this one), then 100,000 rows is not a lot of rows. In fact, it's not even close to a lot of rows. 1 million rows wouldn't even be a lot of rows to make a huge difference here.

So, assuming that your query is indeed using the nonclustered index, then I think we should look somewhere besides the NVARCHAR(MAX) column for the issue. This is not to say that sometimes splitting a table into two tables isn't the best choice, it's just doubtful that it would help here given the info provided.

The three places I would look at for improvement are:

1. Explicit schema names: This is minor, but still, always prefix schema-based objects with their schema name. So you should be using dbo.Books instead of just Books. Not only will this help in cases where multiple schemas are being used and different users have different default schemas, but it also reduces some locking that happens when the schema is not explicitly stated and SQL Server needs to check a few places for it.



2. 
   

   The IN list: These are convenient, but not known for scalability. IN lists expand to an OR condition for each item in the list. Meaning:

   
   
   

   where UniqueToken in 
   (
    'first unique token',
    'second unique token'
    -- 10,000 items here
   )
   

   
   
   

   becomes:

   
   
   

   where UniqueToken = 'first unique token'
   OR UniqueToken = 'second unique token'
   -- 10,000 items here (9,998 more OR conditions)
   

   
   
   

   As you add more items to the list, you get more OR conditions.

   
   
   

   Instead of building an IN list dynamically, create a local temporary table and create the list of INSERT statements. Also, wrap all of them in a transaction to avoid the transaction overhead that would otherwise occur per each INSERT (hence reducing 10,000 transactions down to 1):

   
   
   

   CREATE TABLE #UniqueTokens
   (
    UniqueToken VARCHAR(100) NOT NULL
    COLLATE Latin1_General_100_BIN2
    PRIMARY KEY
   );
   
   BEGIN TRAN;
   
   ..dynamically generated INSERT INTO #UniqueTokens (UniqueToken) VALUES ('...');
   
   COMMIT TRAN;
   

   
   
   

   Now that you have that list loaded, you can use it as follows to get the same set of duplicate tokens:

   
   
   

   SELECT bk.[UniqueToken]
   FROM dbo.Books bk
   INNER JOIN #UniqueTokens tmp
    ON tmp.[UniqueToken] = bk.[UniqueToken];
   

   
   
   

   Or, given that you want to know which of the 10,000 new entries you can load, you really want the list of non-duplicate tokens so that you can insert those, right? In which case you would do the following:

   
   
   

   SELECT tmp.[UniqueToken]
   FROM #UniqueTokens tmp
   WHERE NOT EXISTS(SELECT *
    FROM dbo.Books bk
    WHERE bk.[UniqueToken] = tmp.[UniqueToken]);
   

   
   


3. 
   

   String comparison: If there is no specific need for case-insensitive and/or accent-insensitive comparisons with regards to UniqueToken, and assuming that the database in which you created this table (that does not use the COLLATE clause for the [UniqueToken] column) does not have a binary default collation, then you can improve performance of the matching of UniqueToken values by using a binary comparison. Non-binary comparisons need to create a sort key for each value, and that sort key is based on linguistic rules for a specific culture (i.e. Latin1_General, French, Hebrew, Syriac, etc.). That is a lot of extra processing if the values simply need to be exactly the same. So do the following:

   
   
   
   1. Drop the nonclustered index on UniqueToken
      
   
   
   2. Alter the UniqueToken colummn to be VARCHAR(100) NOT NULL COLLATE Latin1_General_100_BIN2 (just like in the temp table shown above)
   
   
   3. Recreate the nonclustered index on UniqueToken