PostgreSQL

The first form of CREATE TYPE creates a composite type. The composite type is specified by a list of attribute names and data types. An attribute’s collation can be specified too, if its data type is collatable. A composite type is essentially the same as the row type of a table, but using CREATE TYPE avoids the need to create an actual table when all that is wanted is to define a type. A stand-alone composite type is useful, for example, as the argument or return type of a function.

To be able to create a composite type, you must have USAGE privilege on all attribute types.

The second form of CREATE TYPE creates an enumerated (enum) type, as described in Section 8.7. Enum types take a list of quoted labels, each of which must be less than NAMEDATALEN bytes long (64 bytes in a standard PostgreSQL build). (It is possible to create an enumerated type with zero labels, but such a type cannot be used to hold values before at least one label is added using ALTER TYPE.)

The third form of CREATE TYPE creates a new range type, as described in Section 8.17.

The range type’s `subtype can be any type with an associated b-tree operator class (to determine the ordering of values for the range type). Normally the subtype’s default b-tree operator class is used to determine ordering; to use a non-default operator class, specify its name with subtype_opclass. If the subtype is collatable, and you want to use a non-default collation in the range’s ordering, specify the desired collation with the collation` option.

The optional `canonical function must take one argument of the range type being defined, and return a value of the same type. This is used to convert range values to a canonical form, when applicable. See Section 8.17.8 for more information. Creating a canonical` function is a bit tricky, since it must be defined before the range type can be declared. To do this, you must first create a shell type, which is a placeholder type that has no properties except a name and an owner. This is done by issuing the command `CREATE TYPE +`_`+name`_, with no additional parameters. Then the function can be declared using the shell type as argument and result, and finally the range type can be declared using the same name. This automatically replaces the shell type entry with a valid range type.

The optional `subtype_diff function must take two values of the subtype type as argument, and return a `double precision value representing the difference between the two given values. While this is optional, providing it allows much greater efficiency of GiST indexes on columns of the range type. See Section 8.17.8 for more information.

The fourth form of CREATE TYPE creates a new base type (scalar type). To create a new base type, you must be a superuser. (This restriction is made because an erroneous type definition could confuse or even crash the server.)

The parameters can appear in any order, not only that illustrated above, and most are optional. You must register two or more functions (using CREATE FUNCTION) before defining the type. The support functions `input_function and output_function are required, while the functions receive_function, send_function, type_modifier_input_function, type_modifier_output_function and analyze_function` are optional. Generally these functions have to be coded in C or another low-level language.

The `input_function converts the type’s external textual representation to the internal representation used by the operators and functions defined for the type. output_function performs the reverse transformation. The input function can be declared as taking one argument of type `cstring, or as taking three arguments of types cstring, oid, integer. The first argument is the input text as a C string, the second argument is the type’s own OID (except for array types, which instead receive their element type’s OID), and the third is the typmod of the destination column, if known (-1 will be passed if not). The input function must return a value of the data type itself. Usually, an input function should be declared STRICT; if it is not, it will be called with a NULL first parameter when reading a NULL input value. The function must still return NULL in this case, unless it raises an error. (This case is mainly meant to support domain input functions, which might need to reject NULL inputs.) The output function must be declared as taking one argument of the new data type. The output function must return type cstring. Output functions are not invoked for NULL values.

The optional `receive_function converts the type’s external binary representation to the internal representation. If this function is not supplied, the type cannot participate in binary input. The binary representation should be chosen to be cheap to convert to internal form, while being reasonably portable. (For example, the standard integer data types use network byte order as the external binary representation, while the internal representation is in the machine’s native byte order.) The receive function should perform adequate checking to ensure that the value is valid. The receive function can be declared as taking one argument of type `internal, or as taking three arguments of types internal, oid, integer. The first argument is a pointer to a StringInfo buffer holding the received byte string; the optional arguments are the same as for the text input function. The receive function must return a value of the data type itself. Usually, a receive function should be declared STRICT; if it is not, it will be called with a NULL first parameter when reading a NULL input value. The function must still return NULL in this case, unless it raises an error. (This case is mainly meant to support domain receive functions, which might need to reject NULL inputs.) Similarly, the optional `send_function converts from the internal representation to the external binary representation. If this function is not supplied, the type cannot participate in binary output. The send function must be declared as taking one argument of the new data type. The send function must return type `bytea. Send functions are not invoked for NULL values.

You should at this point be wondering how the input and output functions can be declared to have results or arguments of the new type, when they have to be created before the new type can be created. The answer is that the type should first be defined as a shell type, which is a placeholder type that has no properties except a name and an owner. This is done by issuing the command CREATE TYPE +`_`+name`_, with no additional parameters. Then the C I/O functions can be defined referencing the shell type. Finally, `CREATE TYPE with a full definition replaces the shell entry with a complete, valid type definition, after which the new type can be used normally.

The optional `type_modifier_input_function and type_modifier_output_function are needed if the type supports modifiers, that is optional constraints attached to a type declaration, such as `char(5) or numeric(30,2). PostgreSQL allows user-defined types to take one or more simple constants or identifiers as modifiers. However, this information must be capable of being packed into a single non-negative integer value for storage in the system catalogs. The `type_modifier_input_function is passed the declared modifier(s) in the form of a `cstring array. It must check the values for validity (throwing an error if they are wrong), and if they are correct, return a single non-negative integer value that will be stored as the column “[.quote]#typmod”#. Type modifiers will be rejected if the type does not have a `type_modifier_input_function. The type_modifier_output_function converts the internal integer typmod value back to the correct form for user display. It must return a `cstring value that is the exact string to append to the type name; for example numeric’s function might return `(30,2). It is allowed to omit the `type_modifier_output_function`, in which case the default display format is just the stored typmod integer value enclosed in parentheses.

The optional `analyze_function performs type-specific statistics collection for columns of the data type. By default, `ANALYZE will attempt to gather statistics using the type’s “[.quote]#equals”# and “[.quote]#less-than”# operators, if there is a default b-tree operator class for the type. For non-scalar types this behavior is likely to be unsuitable, so it can be overridden by specifying a custom analysis function. The analysis function must be declared to take a single argument of type internal, and return a boolean result. The detailed API for analysis functions appears in src/include/commands/vacuum.h.

While the details of the new type’s internal representation are only known to the I/O functions and other functions you create to work with the type, there are several properties of the internal representation that must be declared to PostgreSQL. Foremost of these is `internallength. Base data types can be fixed-length, in which case internallength is a positive integer, or variable-length, indicated by setting internallength to `VARIABLE. (Internally, this is represented by setting typlen to -1.) The internal representation of all variable-length types must start with a 4-byte integer giving the total length of this value of the type. (Note that the length field is often encoded, as described in Section 69.2; it’s unwise to access it directly.)

The optional flag PASSEDBYVALUE indicates that values of this data type are passed by value, rather than by reference. Types passed by value must be fixed-length, and their internal representation cannot be larger than the size of the Datum type (4 bytes on some machines, 8 bytes on others).

The `alignment parameter specifies the storage alignment required for the data type. The allowed values equate to alignment on 1, 2, 4, or 8 byte boundaries. Note that variable-length types must have an alignment of at least 4, since they necessarily contain an `int4 as their first component.

The `storage parameter allows selection of storage strategies for variable-length data types. (Only `plain is allowed for fixed-length types.) plain specifies that data of the type will always be stored in-line and not compressed. extended specifies that the system will first try to compress a long data value, and will move the value out of the main table row if it’s still too long. external allows the value to be moved out of the main table, but the system will not try to compress it. main allows compression, but discourages moving the value out of the main table. (Data items with this storage strategy might still be moved out of the main table if there is no other way to make a row fit, but they will be kept in the main table preferentially over extended and external items.)

All `storage values other than `plain imply that the functions of the data type can handle values that have been toasted, as described in Section 69.2 and Section 38.12.1. The specific other value given merely determines the default TOAST storage strategy for columns of a toastable data type; users can pick other strategies for individual columns using ALTER TABLE SET STORAGE.

The `like_type parameter provides an alternative method for specifying the basic representation properties of a data type: copy them from some existing type. The values of internallength, passedbyvalue, alignment, and storage are copied from the named type. (It is possible, though usually undesirable, to override some of these values by specifying them along with the `LIKE clause.) Specifying representation this way is especially useful when the low-level implementation of the new type “[.quote]#piggybacks”# on an existing type in some fashion.

The `category and preferred parameters can be used to help control which implicit cast will be applied in ambiguous situations. Each data type belongs to a category named by a single ASCII character, and each type is either “[.quote]#preferred”# or not within its category. The parser will prefer casting to preferred types (but only from other types within the same category) when this rule is helpful in resolving overloaded functions or operators. For more details see Chapter 10. For types that have no implicit casts to or from any other types, it is sufficient to leave these settings at the defaults. However, for a group of related types that have implicit casts, it is often helpful to mark them all as belonging to a category and select one or two of the “[.quote]#most general”# types as being preferred within the category. The category` parameter is especially useful when adding a user-defined type to an existing built-in category, such as the numeric or string types. However, it is also possible to create new entirely-user-defined type categories. Select any ASCII character other than an upper-case letter to name such a category.

A default value can be specified, in case a user wants columns of the data type to default to something other than the null value. Specify the default with the DEFAULT key word. (Such a default can be overridden by an explicit DEFAULT clause attached to a particular column.)

To indicate that a type is an array, specify the type of the array elements using the ELEMENT key word. For example, to define an array of 4-byte integers (int4), specify ELEMENT = int4. More details about array types appear below.

To indicate the delimiter to be used between values in the external representation of arrays of this type, `delimiter can be set to a specific character. The default delimiter is the comma (,`). Note that the delimiter is associated with the array element type, not the array type itself.

If the optional Boolean parameter `collatable is true, column definitions and expressions of the type may carry collation information through use of the `COLLATE clause. It is up to the implementations of the functions operating on the type to actually make use of the collation information; this does not happen automatically merely by marking the type collatable.

Whenever a user-defined type is created, PostgreSQL automatically creates an associated array type, whose name consists of the element type’s name prepended with an underscore, and truncated if necessary to keep it less than NAMEDATALEN bytes long. (If the name so generated collides with an existing type name, the process is repeated until a non-colliding name is found.) This implicitly-created array type is variable length and uses the built-in input and output functions array_in and array_out. The array type tracks any changes in its element type’s owner or schema, and is dropped if the element type is.

You might reasonably ask why there is an ELEMENT option, if the system makes the correct array type automatically. The only case where it’s useful to use ELEMENT is when you are making a fixed-length type that happens to be internally an array of a number of identical things, and you want to allow these things to be accessed directly by subscripting, in addition to whatever operations you plan to provide for the type as a whole. For example, type point is represented as just two floating-point numbers, which can be accessed using point[0] and point[1]. Note that this facility only works for fixed-length types whose internal form is exactly a sequence of identical fixed-length fields. A subscriptable variable-length type must have the generalized internal representation used by array_in and array_out. For historical reasons (i.e., this is clearly wrong but it’s far too late to change it), subscripting of fixed-length array types starts from zero, rather than from one as for variable-length arrays.