Conceptually, every expression of a collatable data type has a collation. (The built-in collatable data types are text, varchar, and char. User-defined base types can also be marked collatable, and of course a domain over a collatable data type is collatable.) If the expression is a column reference, the collation of the expression is the defined collation of the column. If the expression is a constant, the collation is the default collation of the data type of the constant. The collation of a more complex expression is derived from the collations of its inputs, as described below.

The collation of an expression can be the “default” collation, which means the locale settings defined for the database. It is also possible for an expression's collation to be indeterminate. In such cases, ordering operations and other operations that need to know the collation will fail.

When the database system has to perform an ordering or a character classification, it uses the collation of the input expression. This happens, for example, with ORDER BY clauses and function or operator calls such as <. The collation to apply for an ORDER BY clause is simply the collation of the sort key. The collation to apply for a function or operator call is derived from the arguments, as described below. In addition to comparison operators, collations are taken into account by functions that convert between lower and upper case letters, such as lower, upper, and initcap; by pattern matching operators; and by to_char and related functions.

For a function or operator call, the collation that is derived by examining the argument collations is used at run time for performing the specified operation. If the result of the function or operator call is of a collatable data type, the collation is also used at parse time as the defined collation of the function or operator expression, in case there is a surrounding expression that requires knowledge of its collation.

The collation derivation of an expression can be implicit or explicit. This distinction affects how collations are combined when multiple different collations appear in an expression. An explicit collation derivation occurs when a COLLATE clause is used; all other collation derivations are implicit. When multiple collations need to be combined, for example in a function call, the following rules are used:

For example, consider this table definition:

CREATE TABLE test1 (
    a text COLLATE "ru_RU",
    b text COLLATE "es_ES",
    ...
);

Then in

SELECT a < 'foo' FROM test1;

the < comparison is performed according to ru_RU rules, because the expression combines an implicitly derived collation with the default collation. But in

SELECT a < ('foo' COLLATE "fr_FR") FROM test1;

the comparison is performed using fr_FR rules, because the explicit collation derivation overrides the implicit one. Furthermore, given

SELECT a < b FROM test1;

the parser cannot determine which collation to apply, since the a and b columns have conflicting implicit collations. Since the < operator does need to know which collation to use, this will result in an error. The error can be resolved by attaching an explicit collation specifier to either input expression, thus:

SELECT a < b COLLATE "ru_RU" FROM test1;

or equivalently

SELECT a COLLATE "ru_RU" < b FROM test1;

On the other hand, the structurally similar case

SELECT a || b FROM test1;

does not result in an error, because the || operator does not care about collations: its result is the same regardless of the collation.

The collation assigned to a function or operator's combined input expressions is also considered to apply to the function or operator's result, if the function or operator delivers a result of a collatable data type. So, in

SELECT * FROM test1 ORDER BY a || 'foo';

the ordering will be done according to ru_RU rules. But this query:

SELECT * FROM test1 ORDER BY a || b;

results in an error, because even though the || operator doesn't need to know a collation, the ORDER BY clause does. As before, the conflict can be resolved with an explicit collation specifier:

SELECT * FROM test1 ORDER BY a || b COLLATE "fr_FR";

A collation is an SQL schema object that maps an SQL name to locales provided by libraries installed in the operating system. A collation definition has a provider that specifies which library supplies the locale data. One standard provider name is libc, which uses the locales provided by the operating system C library. These are the locales that most tools provided by the operating system use. Another provider is icu, which uses the external ICU library. ICU locales can only be used if support for ICU was configured when Postgres Pro was built.

A collation object provided by libc maps to a combination of LC_COLLATE and LC_CTYPE settings, as accepted by the setlocale() system library call. (As the name would suggest, the main purpose of a collation is to set LC_COLLATE, which controls the sort order. But it is rarely necessary in practice to have an LC_CTYPE setting that is different from LC_COLLATE, so it is more convenient to collect these under one concept than to create another infrastructure for setting LC_CTYPE per expression.) Also, a libc collation is tied to a character set encoding (see Section 23.3). The same collation name may exist for different encodings.

A collation object provided by icu maps to a named collator provided by the ICU library. ICU does not support separate “collate” and “ctype” settings, so they are always the same. Also, ICU collations are independent of the encoding, so there is always only one ICU collation of a given name in a database.

You can specify the provider of the default collation with the --locale and --lc-collate options of the initdb or createdb commands, as follows:

--locale=locale[@provider]
--lc-collate=locale[@provider]

where provider can take the icu or libc value, and locale is specified in the libc format. You can only specify a single locale provider after the @ symbol. The --lc-collate option overrides the --locale setting, regardless of whether the collation provider is specified.

If you do not specify the provider, libc is used for C and POSIX locales. For other locales, the default provider is:

Before starting the server, Postgres Pro checks that the cluster was initialized by a server compiled with the same version of ICU, even if icu is not used as the provider for the default collation. If the ICU library versions do not match, Postgres Pro emits the appropriate informational message.

Before connecting to a database with an icu default collation, Postgres Pro compares this collation version to the one provided by the ICU library. If the collation version changes, you may need to rebuild the objects that depend on a changed collation if you think this change may affect the sort order of your data. Postgres Pro displays a warning if collation versions do not match, or it is impossible to compare them because the cluster does not provide any information about the collation version for the default collation. To suppress these warnings, you can use the ALTER COLLATION "default" REFRESH VERSION command, as explained in ALTER COLLATION.

You can find the default collation and its provider in pg_database.datcollate:

locale@provider

SELECT a COLLATE "C" < b COLLATE "POSIX" FROM test1;

CREATE COLLATION russian (provider = libc, locale = 'ru_RU');

CREATE COLLATION russian FROM "ru_RU";
CREATE COLLATION french FROM "fr-x-icu";

CREATE COLLATION ndcoll (provider = icu, locale = 'und', deterministic = false);

CREATE COLLATION case_insensitive (provider = icu, locale = 'und-u-ks-level2', deterministic = false);
CREATE COLLATION ignore_accents (provider = icu, locale = 'und-u-ks-level1-kc-true', deterministic = false);