Normalization, The Sequel

If there’s one thing that SQL Server is really good at, it’s relationships. After all, a relational database management system without the relationships is nothing more than a place to store your stuff.

Last week we briefly looked at a denormalized table, and then I suggested that breaking it up into five separate tables would be a good idea. So good, in fact, that it took me more than 2,000 words to explain just the first table in our highly contrived example.

Assuming you have read through all those words, let’s attempt a much more condensed look at the other four tables. If you recall, we had:

  • Transactions
  • Products
  • Customers
  • Salespersons
  • Stores

We tackled the Stores table first because everything is backwards when we design databases.

For the next three tables, I’m going to just show you how I would design them, without going into detail. Take a close look at Salespersons, though (which we’ll look at first) because it will give you a clue about how we link all the tables together finally in the Transaction table.

Then take a look at … PaymentTypes? ProductTypes? Colours? Categories? Sizes? Uh … What’s going on here? Where did all those tables come from? Luckily, T-SQL allows comments, which you’ll see below.

CREATE TABLE [Salespersons] (
[SalespersonID] SMALLINT NOT NULL IDENTITY(1,1),
[StoreID] SMALLINT NOT NULL,
[FirstName] NVARCHAR(255) NOT NULL,
[LastName] NVARCHAR(255) NOT NULL,
[EmailAddress] NVARCHAR(512) NOT NULL
CONSTRAINT [PK_Salespersons] PRIMARY KEY CLUSTERED ([SalespersonID] ASC)
);

-- List of possible payment types, (e.g. credit card, cheque)
CREATE TABLE [PaymentTypes] (
[PaymentTypeID] TINYINT NOT NULL IDENTITY(1,1),
[Description] VARCHAR(255) NOT NULL,
CONSTRAINT [PK_PaymentTypes] PRIMARY KEY CLUSTERED ([PaymentTypeID] ASC)
);

CREATE TABLE [Customers] (
[CustomerID] BIGINT NOT NULL IDENTITY(1,1),
[FirstName] NVARCHAR(255) NOT NULL,
[LastName] NVARCHAR(255) NOT NULL,
[EmailAddress] NVARCHAR(512) NOT NULL,
[Telephone] VARCHAR(25) NOT NULL,
[PaymentTypeID] TINYINT NOT NULL,
CONSTRAINT [PK_Customers] PRIMARY KEY CLUSTERED ([CustomerID] ASC)
);

-- List of possible product types (e.g. iPhone, iPhone cover, iPod)
CREATE TABLE [ProductTypes] (
[ProductTypeID] TINYINT NOT NULL IDENTITY(1,1),
[Description] VARCHAR(255) NOT NULL,
CONSTRAINT [PK_ProductTypes] PRIMARY KEY CLUSTERED ([ProductTypeID] ASC)
);

-- List of possible colours
CREATE TABLE [Colours] (
[ColourID] TINYINT NOT NULL IDENTITY(1,1),
[Description] VARCHAR(255) NOT NULL,
CONSTRAINT [PK_Colours] PRIMARY KEY CLUSTERED ([ColourID] ASC)
);

-- List of possible categories (5.5", 4.7", 4", 3.5")
-- This replaces "size", since we might use Size to denote storage
CREATE TABLE [Categories] (
[CategoryID] TINYINT NOT NULL IDENTITY(1,1),
[Description] VARCHAR(255) NOT NULL,
CONSTRAINT [PK_Categories] PRIMARY KEY CLUSTERED ([CategoryID] ASC)
);

-- List of possible sizes ("8GB", "16GB", "32GB", etc.)
-- Can also be used for other product types like laptops
CREATE TABLE [Sizes] (
[SizeID] TINYINT NOT NULL IDENTITY(1,1),
[Description] VARCHAR(255) NOT NULL,
CONSTRAINT [PK_Sizes] PRIMARY KEY CLUSTERED ([SizeID] ASC)
);

CREATE TABLE [Products] (
[ProductID] TINYINT NOT NULL IDENTITY(1,1),
[ProductTypeID] TINYINT NOT NULL,
[ColourID] TINYINT NOT NULL,
[CategoryID] TINYINT NOT NULL,
[SizeID] TINYINT NOT NULL,
[SellingPrice] SMALLMONEY NOT NULL,
CONSTRAINT [PK_Products] PRIMARY KEY CLUSTERED ([ProductID] ASC)
);

Those tables popped out of nowhere, didn’t they? Welcome to the world of normalization. To design a database properly, we come to the realisation that we can simplify the data input even more, reducing repeated values, and finding the most unique way of representing data. Products comprise product types. A single list of colours can be reused in various places. Payment types can be used for all sorts of transactional data.

We end up with a lot of tables when we normalize a database, and this is perfectly normal. When we want to read information out of the system the way senior management wants, we must join all these tables together.

The only way to join tables together in a safe and meaningful way is with foreign key relationships, where one table’s primary key is referenced in another table, with a matching data type.

The SalesPersons table has a StoreID column. As it stands, there’s no relationship between Stores and SalesPersons until we create the relationship using T-SQL.

ALTER TABLE [SalesPersons]
ADD CONSTRAINT FK_SalesPersons_Stores FOREIGN KEY (StoreID)
REFERENCES [Stores] (StoreID);

  • Line 1: Inform SQL Server that we are altering an existing table
  • Line 2: By adding a foreign key constraint (i.e. limiting what can go into the StoreID column)
  • Line 3: By forcing it to use the values from the Stores table’s StoreID column (i.e. the primary key).

In a relationship diagram, it looks like this (in SQL Server Management Studio’s Database Diagram tool):

The yellow key in each table is the Primary Key (StoreID and SalespersonID respectively). There is a StoreID column in both tables with the same data type (SMALLINT). The foreign key (FK) does not have to match the name of the primary key (PK), but it makes things a lot easier to have the same name for both sides of a relationship in large databases, so it’s a good habit.

Notice the direction of the relationship (FK_Salespersons_Stores) in the picture, with the yellow key on the table with the Primary Key. The name of the relationship is also sensible. To the casual eye, this says that there’s a Foreign Key constraint in the Salespersons table that points to the Primary Key in the Stores table.

Now we see why data types are so important with relational data. A relationship is not even possible between two tables if the data type is not the same in both key columns.

With this constraint enabled, whenever we insert data into the Salespersons table, we have to make sure that whatever we put into the StoreID column must already exist in the Stores table.

Let’s do the rest of the relationships so far, and then we’ll look at the Transactions table.

ALTER TABLE [Customers]
ADD CONSTRAINT FK_Customers_PaymentTypes FOREIGN KEY (PaymentTypeID)
REFERENCES [PaymentTypes] (PaymentTypeID);

ALTER TABLE [Products]
ADD CONSTRAINT FK_Products_ProductTypes FOREIGN KEY (ProductTypeID)
REFERENCES [ProductTypes] (ProductTypeID);

ALTER TABLE [Products]
ADD CONSTRAINT FK_Products_Colours FOREIGN KEY (ColourID)
REFERENCES [Colours] (ColourID);

ALTER TABLE [Products]
ADD CONSTRAINT FK_Products_Categories FOREIGN KEY (CategoryID)
REFERENCES [Categories] (CategoryID);

ALTER TABLE [Products]
ADD CONSTRAINT FK_Products_Sizes FOREIGN KEY (SizeID)
REFERENCES [Sizes] (SizeID);

As we can see now, we have more than one foreign key relationship in the Products table, to ProductTypes, Colours, Categories, and Sizes, which is a clue to how the Transactions table will look.

CREATE TABLE [Transactions] (
[TransactionID] BIGINT NOT NULL IDENTITY(1,1),
[TransactionDate] DATETIME2(3) NOT NULL,
[ProductID] TINYINT NOT NULL,
[DiscountPercent] DECIMAL(4,2) NOT NULL DEFAULT(0),
[SalesPersonID] SMALLINT NOT NULL,
[CustomerID] BIGINT NOT NULL,
[HasAppleCare] BIT NOT NULL DEFAULT(0),
CONSTRAINT [PK_Transactions] PRIMARY KEY CLUSTERED ([TransactionID] ASC)
);

Let’s assume we’ve created all the relationships as well, so that we’re left with the following relationships (click to enlarge):

Ten tables, compared to our original one denormalized table, is a significant increase in number of tables. However, let’s compare the data usage for adding transactions.

If we were to populate each row of each table with enough data to provide the same number of transactions as last week’s two purchases, it would look like this.

Stores table: 172 bytes

  • StoreID: 2 bytes (IDENTITY value as SMALLINT)
  • StoreCode: 0 bytes (NULL)
  • StoreName: 30 bytes (“Chinook Centre” is 14 characters long including the space, so with Unicode that becomes 28 bytes, plus 2 for the overhead of using NVARCHAR)
  • Address: 86 bytes (“6455 Macleod Trail SW, Calgary, AB T2H 0K8” is 42 characters, so 84 with Unicode, plus 2 bytes for NVARCHAR overhead)
  • ManagerName: 22 bytes (“Bob Bobson” in Unicode, plus 2 bytes NVARCHAR overhead)
  • ManagerEmail: 32 bytes (“[email protected]” in Unicode, plus 2 bytes NVARCHAR overhead)

Salespersons table: 70 bytes

  • SalespersonID: 2 bytes
  • StoreID: 2 bytes
  • FirstName: 14 bytes (“Thandi” in Unicode + 2 bytes)
  • LastName: 14 bytes (“Funaki” in Unicode + 2 bytes)
  • EmailAddress: 38 bytes (“[email protected]” in Unicode + 2 bytes)

PaymentTypes table: 14 bytes

  • PaymentTypeID: 1 byte
  • Description: 13 bytes (“Credit Card” + 2 bytes)

ProductTypes table: 23 bytes

  • ProductTypeID: 1 byte
  • Description: 8 bytes (“iPhone” + 2 bytes)
  • ProductTypeID: 1 byte
  • Description: 13 bytes (“iPhone Case” + 2 bytes)

Colours table: 21 bytes

  • ColourID: 1 byte
  • Description: 7 bytes (“Black” + 2 bytes)
  • ColourID: 1 byte
  • Description: 6 bytes (“Blue” + 2 bytes)
  • ColourID: 1 byte
  • Description: 5 bytes (“Red” + 2 bytes)

Categories table: 7 bytes

  • CategoryID: 1 byte
  • Description: 6 bytes (“5.5″” + 2 bytes)

Sizes table: 16 bytes

  • SizeID: 1 byte
  • Description: 7 bytes (“128GB” + 2 bytes)
  • SizeID: 1 byte
  • Description: 7 bytes (“256GB” + 2 bytes)

Products table: 27 bytes (3 products at 9 bytes each)

  • ProductID: 1 byte
  • ProductTypeID: 1 byte
  • ColourID: 1 byte
  • CategoryID: 1 byte
  • SizeID: 1 byte
  • SellingPrice: 4 bytes

Customers table: 192 bytes

  • CustomerID: 8 bytes
  • FirstName: 10 bytes (“I.M.” in Unicode + 2 bytes)
  • LastName: 18 bytes (“Customer” in Unicode + 2 bytes)
  • EmailAddress: 42 bytes (“[email protected]” in Unicode + 2 bytes)
  • Telephone: 16 bytes (“(403) 555-1212” + 2 bytes)
  • PaymentTypeID: 1 byte
  • CustomerID: 8 bytes
  • FirstName: 10 bytes (“U.R.” in Unicode + 2 bytes)
  • LastName: 18 bytes (“Customer” in Unicode + 2 bytes)
  • EmailAddress: 44 bytes (“[email protected]” in Unicode + 2 bytes)
  • Telephone: 16 bytes (“(403) 665-0011” + 2 bytes)
  • PaymentTypeID: 1 byte

Transactions table: 96 bytes (32 bytes per transaction)

  • TransactionID: 8 bytes
  • TransactionDate: 7 bytes
  • ProductID: 1 byte
  • DiscountPercent: 5 bytes
  • SalesPersonID: 2 bytes
  • CustomerID: 8 bytes
  • HasAppleCare: 1 bit (expands to 1 byte)

GRAND TOTAL: 638 bytes to represent all three transactions

The denormalized version, for which we can see the original example below, works out as follows. Recall we said last week that each column was NVARCHAR(4000), or possibly even NVARCHAR(MAX).

A Wide Table
A Wide Table Appears – click to enlarge

At our most generous, we would need 1,166 bytes to record these three transactions. That’s almost double the data required, just for these three. Plus, the data has no foreign key relationships, so we cannot be sure that whatever is being added to the denormalized table is valid.

As time goes on, the normalized tables will grow at a much lower rate proportionally. Consider what a denormalized Transactions table would look like with an average row size of 388 bytes, for ten million rows (3.6GB).

Compare that to a normalized database, with ten million transactions for 8 million customers. Even assuming we have a hundred products, with twenty colours, and 30 product types, we would see only around 1GB of space required to store the same data.

We know Apple as being one of the most successful technology companies in terms of sales, so extrapolating to 1 billion transactions, we’d be comparing 361GB (for the denormalized table) with less than half that (178GB) if every single customer was unique and only ever bought one item.

Aside from the staggering savings in storage requirements, normalization gives us sanity checks with data validation by using foreign key constraints. Along with proper data type choices, we have an easy way to design a database properly from the start.

Sure, it takes longer to start, but the benefits outweigh the costs almost immediately. Less storage, less memory to read the data, less space for backups, less time to run maintenance tasks, and so on.

Normalization matters.

Next week, we talk briefly about bits and bytes, and then we will start writing queries. Stay tuned.

Find me on Twitter to discuss your favourite normalization story at @bornsql.

A First Look At Normalization

Phew! There’s a lot to take in with data types, collation, precision, scale, length, and Unicode, and we’re just getting warmed up. This week’s post is over 2,000 words long!

Over the last three weeks, we’ve gone fairly deep into data types, and now we are going to see how they come into play with normalization.

If we go back to the first post in this series, I mentioned normalization, and then apparently I forgot about it in the next two posts. What you didn’t see is that I was talking about it all along.

Continue reading “A First Look At Normalization”

Fundamentals of Data Types

Last week, we discussed storing text in a database. This week we will dive deeper into data types.

When storing data in our database, we want to make sure that it’s stored accurately and that we only use the required amount of space.

This is because when we access the data later, we want to make sure any calculations are accurate; plus reading the data takes up memory, and we want to be as efficient as we can with memory usage.

There are seven data type categories in SQL Server:

  • exact numerics
  • approximate numerics
  • date and time
  • character strings
  • Unicode character strings
  • binary strings
  • other

When we want to use these data types for our columns, we need to declare them. Some require a length, some require a precision and scale, and some can be declared without a length at all. For example:

No Length (implied in data type):
DECLARE @age AS TINYINT;

Explicit Length (length is supplied):
DECLARE @firstName AS VARCHAR(255);

Precision and Scale:
DECLARE @interestRate AS DECIMAL(9,3);

Let’s talk a bit about precision and scale, because those values between the brackets may not work the way we think they do.

Precision and Scale

Data types with decimal places are defined by what we call fixed precision and scale. Let’s look at an example:

123,456.789

In the above number, we see a six-digit number (ignoring the thousand separator) followed by a decimal point, and then a fraction represented by three decimal places. This number has a scale of 3 (the digits after the decimal point) and a precision of 9 (the digits for the entire value, on both sides of the decimal point). We would declare this value as DECIMAL(9,3).

This is confusing at first glance, because we have to declare it “backwards”, with the precision first, and then the scale. It may be easier to think of the precision in the same way we think of a character string’s length.

Date and time data types can also have decimal places, and SQL Server supports times accurate to the nearest 100 nanoseconds. The most accurate datetime is DATETIME2(7), where 7 decimal places are reserved for the time.

Before SQL Server 2008, we used DATETIME, which is only accurate to the nearest 3 milliseconds, and uses 8 bytes. A drop-in replacement for this is DATETIME2(3), using 3 decimal places, and accurate to the nearest millisecond. It only needs 7 bytes per column.

Be mindful that, as higher precision and scale are required, a column’s storage requirement increases. Accuracy is a trade-off with disk space and memory, so we may find ourselves using floating point values everywhere.

However, in cases where accuracy is required, always stick to exact numerics. Financial calculations, for example, should always use DECIMAL and MONEY data types.

Exact Numerics

Exact Numerics are exact, because any value that is stored is the exact same value that is retrieved later. These are the most common types found in a database, and INT is the most prevalent.

Exact numerics are split up into integers (BIGINT, INT, SMALLINT, TINYINT, BIT) and decimals (NUMERIC, DECIMAL, MONEY, SMALLMONEY). Decimals have decimal places (defined by precision and scale), while integers do not.

Integers have fixed sizes (see table below), so we don’t need to specify a length when declaring this data type.

Type Bytes Range
BIGINT 8 bytes -2^63 to 2^63-1
INT 4 bytes -2^31 to 2^31-1
SMALLINT 2 bytes -2^15 to 2^15-1
TINYINT 1 byte 0 to 255
BIT 1 bit 0 to 1
  • BIT is often used for storing Boolean values, where 1 = True and 0 = False.
  • Yes, BIGINT can store numbers as large as 2 to the power of 63 minus 1. That’s 19 digits wide, with a value of 9,223,372,036,854,775,807, or 9.2 quintillion.

Decimals may vary depending on the precision and scale, so we have to specify those in the declaration.

Type Bytes Range
DECIMAL 5 to 17 bytes Depends on precision and scale.
38 digits is the longest possible precision.
NUMERIC
  • DECIMAL and NUMERIC are synonyms and can be used interchangeably. Read more about this data type, and how precision and scale affects bytes used, here.

Although the MONEY and SMALLMONEY data types do have decimal places, they don’t require the precision and scale in the declaration because these are actually synonyms for DECIMAL(19,4) and DECIMAL(10,4) respectively. Think of these data types for convenience more than anything.

Type Bytes Range
MONEY 8 bytes -922,337,203,685,477.5808 to 922,337,203,685,477.5807
SMALLMONEY 4 bytes -214,748.3648 to 214,748.3647

Approximate Numerics

Approximate Numerics mean that the value stored is only approximate. Floating point numbers would be classified as approximate numerics, and these comprise FLOAT and REAL.

Declaring a FLOAT requires a length, which represents the number of bits used to store the mantissa. REAL is a synonym of FLOAT(24).

The mantissa means the significant digits of a number in scientific notation, which is how floating point numbers are represented. The default is FLOAT(53). Generally, we stick to the defaults, and use REAL if we want to save space, forgoing some accuracy of the larger FLOAT(53).

Type Bytes Range
FLOAT 4 or 8 bytes -1.79E+308 to -2.23E-308, 0 (zero),
and 2.23E-308 to 1.79E+308
REAL 4 bytes -3.40E+38 to -1.18E-38, 0 (zero),
and 1.18E-38 to 3.40E+38

Date and Time

Date and time data types are slightly more complex. For storing dates (with no time), we use DATE. We store times (with no dates) using TIME. For storing both date and time in the same column, we can use DATETIME2, DATETIME, or SMALLDATETIME. Finally, we can even store timezone-aware values comprising a date and time and timezone offset, using DATETIMEOFFSET.

DATETIME2, TIME, and DATETIMEOFFSET take a length in their declarations, otherwise they default to 7 (accurate to the nearest 100 nanoseconds).

Character Strings

As we saw last week, characters can be fixed-length (CHAR) or variable-length (VARCHAR), and can support special Unicode character types (NCHAR and NVARCHAR respectively). Collation should also be taken into account.

Length can be 1 to 8000 for CHAR and VARCHAR, or 1 to 4000 for NCHAR and NVARCHAR. For storing values larger than that, see the Large Objects section below.

Binary Strings

Sometimes we want to store binary content in a database. This might be a JPEG image, a Word document, an SSL certificate file, or anything that could traditionally be saved on the file system. SQL Server provides the BINARY and VARBINARY data types for this (and IMAGE for backward compatibility).

Length can be 1 to 8000 for BINARY and VARBINARY. For storing values larger than that, see the Large Object section below.

Large Objects

SQL Server 2008 introduced a new MAX length for several data types, including CHAR, NCHAR, VARCHAR, NVARCHAR, BINARY and VARBINARY.

(The XML data type uses MAX under the covers as well.)

This new specification allows up to 2 GB of data to be stored in a column with that declared length. We should take care not to use 2 GB when inserting data into these columns, but it provides greater flexibility when inserting more than 8000 bytes into one of these columns.

Other Data Types

SQL Server supports other types of data, which fall outside the scope of text and numerics. These include CURSOR, TABLE, XML, UNIQUEIDENTIFIER, TIMESTAMP (not to be confused with the date and time types), HIERARCHYID, SQL_VARIANT, and Spatial Types (GEOGRAPHY and GEOMETRY).

Next week, we will see how normalization and data types work together, now that we have a good overview of the different data types in a database.

If you have any thoughts or comments, please find me on Twitter at @bornsql.

Data Types and Collation

Last week we started with a very simple definition of a database: a discrete set of information, with a specific structure and order to it.

We briefly looked at normalization, which is a way to store as little of the information as possible, so that it stays unique.

We will cover more normalization as we move forward through this series, but first we will talk about how the information, or data, is stored. (This does affect normalization and relationships, even if that is not immediately clear.)

For this week’s discussion, we need to consider a spreadsheet application, like Microsoft Excel or Google Sheets.

Columns and Rows

In a spreadsheet, we have columns and rows. Usually we will also have a header row, so we can distinguish between each column.

Each column, in turn, may be formatted a certain way so that we can easily see what kind of information is in that column. For instance, we may want to format a column as a currency, with leading symbol and the two decimal places at the end.

We may left-align text values, and we may decide that numbers have no decimal places and are right-aligned. Dates and times get their own formatting.

If we were to compare this structure to that of a database, we can imagine that each sheet is a table, and each column and row is a column and row in the table.

In some applications like Microsoft Access, we may hear different terminology for columns and rows, namely fields and records. However, in SQL Server, we maintain the same convention as Excel and call them columns and rows.

Because SQL Server doesn’t care about how our data looks, we have to specify those traits when we create the table. Whether creating from scratch or from an import process through an external application (Excel, for example), we need to specify the data type for each column.

There are several key reasons why we want to do this.

In the case of numbers that will be summarized in some way (sum, average, minimum, maximum, mean, mode), we want SQL Server’s database engine to treat these as numbers internally so that it doesn’t have to convert anything, which in turn makes the calculations much faster.

The same goes for dates, times, and datetimes (where both the date and time is in one column) because the database engine understands date and time calculations, provided the data types are correct.

Text values are also very important but for a fundamentally different reason. While computers understand numbers, it’s humans that understand text.

We will focus the rest of this week’s discussion on storing strings in a database.

Collation

Imagine we are developing a database for international customers, and we need to support accented characters or an entirely different alphabet. Database systems use a catch-all term for this, and that is collation.

When we install SQL Server, we are asked what the “default” is, then we are presented with some arcane terminology which may be confusing, so we leave the defaults and click Next.

Collation has to do with how data is sorted, and thus the order in which we see it when data is returned.

Note that collation only affects text columns.

The Windows regional settings, for the user installing SQL Server, will affect the default collation of a SQL Server installation. If we were to install SQL Server on a machine that is configured with U.S. regional settings, it will have a very different default collation than a server that is set for Canada or Finland.

The default SQL Server collation for US regional settings (SQL_Latin1_General_CP1) may need to be changed to match what is required for the user databases that will be running on a server.

The above values mean the following:

  • General – the sort order follows 0-9, A-Z;
  • CP1 – code-page 1, the US English default;
  • Case Insensitivity and Accent Sensitivity are implied (see below).

When not using US English, or the Latin alphabet, we need to be aware that the data’s sort order is taken into account.

Even more confusingly, some vendor products require a specific collation for their database. For example, Microsoft’s own SharePoint database uses the collation Latin1_General_CI_AS_KS_WS:

  • CICase Insensitive – no difference between upper and lower case when sorting data;
  • ASAccent Sensitive – distinguishes between accented characters, for instance, the Afrikaans words “sê” and “se” are considered different;
  • KSKana Sensitive – distinguishes between different Japanese character sets;
  • WSWidth Sensitive – distinguishes between characters that can be expressed by both single- or double-byte characters.

(Read up more about collation options here.)

Text Data Types

Now that we have a very basic grasp of collation, let’s look at text data types.

We tend to use only four text data types in SQL Server these days:

CHAR(n), NCHAR(n), VARCHAR(n), and NVARCHAR(n), where n may be a number between 1 and 8,000 or the keyword MAX.

Why 8,000?

For historic reasons, SQL Server set their data page size (the amount of storage available on each data page, including headers and footers) to 8KB many years ago. This means that the largest amount of data we can store on a single page is 8,192 bytes. Once we take away the header and the slot array at the end, we are left with slightly more than 8,000 bytes for our data.

When we store a text value, we need to decide if the characters can be expressed in a single byte or as double-byte characters (also known as Unicode, using two bytes per character). Alphabets like Kanji, Chinese (Simplified or Traditional), and Turkish, will require double-byte characters, for each character in their alphabet.

(Some code pages need more than two bytes for a character. That is outside of the scope of this discussion.)

So CHAR or VARCHAR uses one byte per character, while NCHAR and NVARCHAR uses two bytes per character (the N represents Unicode).

Thus, the longest a CHAR or VARCHAR string can be is 8000, while the longest an NCHAR or NVARCHAR string can be is 4000 (at two bytes per character).

MAX Data Type

In SQL Server 2008, several new data types were introduced, including the MAX data type for strings and binary data. The underlying storage mechanism was changed to allow columns longer than 8,000 bytes, where these would be stored in another section of the database file under certain conditions.

The MAX data type allows up to 2 GB (more than two billion bytes) for every row that column is used.

So we have to consider three distinct things when deciding how we store text: collation, Unicode, and string length.

Because my readers are terribly intelligent, you’ve already deduced that the VAR in VARCHAR means “variable length”, and you’d be correct.

We use VARCHAR (and its Unicode equivalent NVARCHAR) for columns that will contain strings with variable lengths, including names, addresses, phone numbers, product names, etc. In fact, along with INT (meaning a 4-byte integer), VARCHAR is probably the most common data type in any database today.

CHAR (and NCHAR), on the other hand, are fixed-length data types. We use this type for string lengths that are unlikely to change. For example, IMEI numbers, Vehicle Identification Numbers, social security numbers (where the dash forms part of the number), product codes, serial numbers with leading zeroes, and so on. The point here is that the length is fixed.

So why don’t we just use VARCHAR instead of CHAR everywhere?

Let’s start with why VARCHAR was introduced in the first place, and why we would use it instead of CHAR.

For columns with unpredictably long strings, we don’t want to reserve all 8,000 bytes per row for a string that may only take up 2,000 bytes—and end up wasting 6,000 (not to mention the storage required for a MAX column)—so we switch to VARCHAR, and each row only uses as many bytes as it needs.

However, SQL Server needs to keep track of the length of a VARCHAR column for each row in a table. There is a small overhead of a few bytes per row for every VARCHAR for SQL Server to keep track of this length. The reason we don’t replace CHAR and NCHAR outright, is ironically to save space.

It doesn’t make sense for a table containing millions or billions of rows to use VARCHAR for fixed-length columns because we would be adding on another few bytes per row as unnecessary overhead. Adding just one byte per million rows is roughly 1 MB of storage.

Extrapolating that extra byte to the memory required to hold it, maintenance plans when updating indexes and statistics, backups, replicated databases, and so on, we are now looking at extra megabytes, and possibly gigabytes, for the sake of convenience.

We must make sure that we pick the correct character type for storing strings, beyond just the length of the string. Both CHAR and VARCHAR have their place.

While we did spend most of this discussion on collations and text, we’ve only scratched the surface.

Next week, we will discuss how to pick the right data type for your columns, with concrete examples. This matters a lot with how numbers are stored.

If you have any feedback, find me on Twitter at @bornsql.

Testing for Object Existence: CREATE OR ALTER

For the longest time, T-SQL writers have had to wrestle with ways of testing for an object’s existence so that it can either be dropped and recreated, or modified as needed.

Last week we covered the new DROP ... IF EXISTS syntax. This week goes into how we handle changes to objects.

We’ve spent many hours of our lives fighting with an object existence check so that ALTER commands don’t fail.

The common implementation pattern now is to CREATE a dummy object with the appropriate name and then use ALTER to write the actual code. That way, future changes can just be done by using the ALTER keyword.

We can see this in a famous example: sp_WhoIsActive, an extremely popular stored procedure written by Adam Machanic, which I highly recommend installing when setting up a new SQL Server instance.

USE master;
GO

IF NOT EXISTS (SELECT * FROM INFORMATION_SCHEMA.ROUTINES
WHERE ROUTINE_NAME = 'sp_WhoIsActive')
EXEC ('CREATE PROC dbo.sp_WhoIsActive AS
SELECT ''stub version, to be replaced''');
GO
...
ALTER PROC dbo.sp_WhoIsActive (...);

Look at all that unnecessary code. It’s messy; it’s prone to errors. There are so many ways to do it, which makes it inconsistent, meaning that automatically searching through a code base isn’t reliable. Additionally, if we miss a single quotation mark, the entire script fails.

In a word: Ugh!

What happens if we forget to run the stub first? What happens if we have an existing object and run the CREATE accidentally, then we have to manually change it to an ALTER

(Side note: I spent a good few minutes one day a few years ago not remembering the ALTER keyword.)

SQL Server 2016 Service Pack 1 has finally added in a feature that many DBAs and database developers have been clamouring for: CREATE OR ALTER.

In the above example, the entire IF NOT EXISTS section can be replaced with:

USE master;
GO

CREATE OR ALTER PROC dbo.sp_WhoIsActive (...);

While this has been a long time coming, causing many anguished cries from people writing T-SQL scripts over the years, we are going to love using this new, small, yet significant, syntax.

Share your best ALTER story with me on Twitter, at @bornsql.

Testing for Object Existence: DROP … IF EXISTS

For the longest time, T-SQL writers have had to wrestle with ways of testing for an object’s existence so that it can either be dropped and recreated, or modified as needed.

This is especially common in the case of temp tables or table variables. If the object already exists when our script runs, the script will fail and leave our workflow in an inconsistent state.

Consider this script:

IF OBJECT_ID('tempdb..#doTheThing') IS NOT NULL
DROP TABLE #doTheThing;

CREATE TABLE #doTheThing (
ImportantColumn1 BIGINT,
ImportantColumn2 TINYINT,
ImportantColumn3 NVARCHAR(255)
);

With SQL Server 2016, we can now do this:

DROP TABLE IF EXISTS #doTheThing;

CREATE TABLE #doTheThing (
ImportantColumn1 BIGINT,
ImportantColumn2 TINYINT,
ImportantColumn3 NVARCHAR(255)
);

DROP ... IF EXISTS can be used on many objects, including DATABASE, FUNCTION, INDEX, PROCEDURE, ROLE, SCHEMA, SEQUENCE, SYNONYM, TABLE, TRIGGER, TYPE, USER and VIEW.

Share your object existence check nightmares with me on Twitter at @bornsql.

Temporal Tables and Hidden Period Columns

In my November 2015 post, An Introduction to Temporal Tables in SQL Server 2016 using a DeLorean, I wrote:

The HIDDEN property is optional and will hide these columns from a standard SELECT statement for backward compatibility with our application and queries. You cannot apply the HIDDEN property to an existing column.

It turns out that this is no longer true. You can apply the HIDDEN property to an existing period column.

Let’s assume you have a temporal table containing two visible period columns, StartDate and EndDate, which you’d like to hide from a typical SELECT statement.

Using an ALTER TABLE ... ALTER COLUMN statement, simply place the ADD HIDDEN syntax after the period column name(s).

ALTER TABLE [dbo].[Account] ALTER COLUMN [StartDate] ADD HIDDEN;
ALTER TABLE [dbo].[Account] ALTER COLUMN [EndDate] ADD HIDDEN;

You can also remove this flag if you wish, using DROP HIDDEN:

ALTER TABLE [dbo].[Account] ALTER COLUMN [StartDate] DROP HIDDEN;
ALTER TABLE [dbo].[Account] ALTER COLUMN [EndDate] DROP HIDDEN;

This is a great improvement to an already fantastic feature of SQL Server 2016. Thanks to Borko Novakovic for this tip.

If you have any more temporal table tricks you want to share, find me on Twitter at @bornsql.

Temporal Tables and History Retention

I’m a huge fan of Temporal Tables in SQL Server 2016. I first wrote about them, in a four-part series in November 2015, before SQL Server was even released. I don’t always get this excited about new features.

However, it has some limitations. As part of this week’s T-SQL Tuesday, hosted by the attractive and humble Brent Ozar, I have discovered a Microsoft Connect item I feel very strongly about.

Adam Machanic, the creator of an indispensable tool, sp_WhoIsActive, has created a Connect item entitled Temporal Tables: Improve History Retention of Dropped Columns.

As my readers know, temporal tables have to have the same schema as their base tables (the number and order of columns, and their respective data types, have to match).

Where this breaks down is when a table structure has changed on the base table. The history table also needs to take those changes into account, which could potentially result in data loss or redundant columns in the base table.

Adam suggests allowing columns which no longer appear in the base table to be retained in the history table and marked as nullable (or hidden), and should only appear when performing a point-in-time query by referring to the column(s) explicitly.

I have voted for this suggestion, and at the time of writing, it has 16 upvotes. I encourage you to add your voice to this suggestion.

If you have any other suggestions, or wish to discuss temporal tables, please contact me on Twitter at @bornsql .

Updated Max Server Memory Script

Earlier this year I released a free T-SQL script that will calculate the correct amount of RAM you should allocate to SQL Server, assuming it is a standalone instance.

After attending the PASS Summit in Seattle in October, I visited the SQL Server Tiger team’s GitHub repository and discovered something similar, but not quite the same, in the Maintenance Solution folder.

I have taken the best ideas from their Database Server Options script and merged them into my Max Server Memory Calculator script.

New Features

The SQL Server thread stack is now taken into account. This value depends on the CPU architecture (32-bit, or x64 / IA64) and the maximum worker threads configured for the SQL Server instance.

On my 64-bit laptop with 16GB RAM, the new recommended amount for Max Server Memory has dropped from 11,264 MB to 10,112 MB (1,125 MB of RAM is now reserved for the thread stack).

Improvements

By default, the generated script will enable show advanced options before trying to set the max server memory (MB) value.

The @ProductVersion parameter uses a new method to calculate the major SQL Server version.  Previously it was a hack based on the string returned by the @@VERSION function, but now it uses the @@MICROSOFTVERSION function.

This code is also from the Tiger team’s repository, and I’m sharing it here because I think it’s pretty clever how it works.

-- Get SQL Server Major Version
SELECT CONVERT(INT, (@@MICROSOFTVERSION / 0x1000000) & 0xFF);

I have also added a note on the Max Server Memory Matrix page to note that the script now accounts for the thread stack.

I hope you enjoy this new version of the script. If you have any comments or suggestions, please contact me on Twitter at @bornsql .

Look, Ma, No Surprises

Last week I demonstrated at least 30% performance improvement by switching to memory optimised table-valued parameters on SQL Server 2016.

This week I will demonstrate the same test using Azure SQL Database, on the Premium tier, where In-Memory OLTP is supported.

My test harness is the same. I will create a temp table and a table variable:

-- Temp table creation
CREATE TABLE #temptable
(ID UNIQUEIDENTIFIER NOT NULL PRIMARY KEY);

-- Table variable creation
DECLARE @tablevariable AS TABLE
(ID UNIQUEIDENTIFIER NOT NULL PRIMARY KEY);

Now I will create a standard table-valued parameter:

CREATE TYPE dbo.TVPStandard AS TABLE
(ID UNIQUEIDENTIFIER NOT NULL PRIMARY KEY);

DECLARE @TempTVPStandard AS dbo.TVPStandard;

Finally, I will create a memory-optimized table-valued parameter (there is no requirement to separately enable In-Memory OLTP in Azure SQL Database):

CREATE TYPE dbo.TVPInMemory AS TABLE
(ID UNIQUEIDENTIFIER NOT NULL PRIMARY KEY NONCLUSTERED)
WITH (MEMORY_OPTIMIZED = ON);

DECLARE @TempTVPMemory AS dbo.TVPInMemory;

So far, nothing is different from last week’s scenario. We now have the following structures at our disposal:

  • #temptable (a temp table)
  • @tablevariable (a table variable)
  • @TempTVPStandard (a standard TVP)
  • @TempTVPMemory (a memory-optimized TVP)

I’m going to use the same WHILE loop again, but instead of a million runs, I’ll do 1000, 10,000 and 100,000, because I’m paying for this instance of Azure SQL Database (I picked a Premium P1, with 125 DTUs) and I’m a cheapskate. I doubt the 125 DTUs is even enough to run a million times for the fourth option.

PRINT SYSUTCDATETIME();

SET NOCOUNT ON;

DECLARE @i INT = 0;
WHILE @i < <number of executions>
BEGIN
INSERT INTO <object name>
SELECT NEWID();
SELECT @i = @i + 1;
END;

SET NOCOUNT OFF;

PRINT SYSUTCDATETIME();

Results!

Like last week, at low row counts all four data structures performed around the same speed. Tempdb is shared with other Azure SQL Database customers, so I expected to see that slower.

I ran each series three times and took the lowest value from each run.

Data Structure 1,000 10,000 100,000 1,000,000
Temp Table (Clustered) 94 ms 453 ms 4,266 ms 44,955 ms
Table Variable (Clustered) 93 ms 344 ms 3,484 ms 34,673 ms
Standard TVP (Clustered) 94 ms 343 ms 3,500 ms 34,610 ms
Memory-Optimized TVP
(Non-Clustered)
78 ms 203 ms 1,797 ms No Time

Unsurprisingly, because Azure SQL Databases share tempdb with other customers, the IO-bound figures are higher than those on my dedicated laptop, no matter how old it is.

The big winner here, again, is the memory-optimized table-valued parameter. More than twice as fast as a temp table, and almost twice as fast as a table variable or standard TVP.

Note, however, that because my Azure SQL Database was resource-constrained, I was unable to run the one-million step WHILE loop.

This is an excellent example of how In-Memory OLTP is a trade-off, no matter whether you’re running on premises or in the cloud. While you do consistently get much better performance, it is not reliable for large data structures.

If you have any In-Memory OLTP tweaks you’d like to share, find me on Twitter at @bornsql.