Collation support¶

Understanding collation¶

Text strings in Snowflake are stored using the UTF-8 character set and, by default, strings are compared according to the Unicode codes that represent the characters in the string.

However, comparing strings based on their UTF-8 character representations might not provide the desired or expected behavior. For example:

• If special characters in a given language do not sort according to that language’s ordering standards, then sorting might return unexpected results.

• You might want the strings to be ordered by other rules, such as ignoring whether the characters are uppercase or lowercase.

Collation allows you to explicitly specify the rules to use for comparing strings, based on:

• Different locales (i.e. different character sets for different languages).

• Case-sensitivity (i.e. whether to use case-sensitive or case-insensitive string comparisons without explicitly calling the UPPER or LOWER functions to convert the strings).

• Accent-sensitivity (e.g. whether Z, Ź, and Ż are considered the same letter or different letters).

• Punctuation-sensitivity (i.e. whether comparisons use only letters or include all characters). For example, if a comparison is punctuation-insensitive, then A-B-C and ABC are treated as equivalent.

• Additional options, such as preferences for sorting based on the first letter in a string and trimming of leading and/or trailing blank spaces.

Uses for collation¶

Collation can be used in a wide variety of operations, including (but not limited to):

Collation control¶

Collation control is granular. You can explicitly specify the collation to use for:

• An account, using the account-level parameter DEFAULT_DDL_COLLATION.

• All columns in all tables added to a database, using the ALTER DATABASE command.

• All columns in all tables added to a schema, using the ALTER SCHEMA command.

• All columns added to a table, using the ALTER TABLE command.

• Individual columns in a table, using the CREATE TABLE command.

• A specific comparison within a SQL statement (e.g. WHERE col1 = col2). If multiple collations are applied to a statement, Snowflake determines the collation to use based on precedence. For more details about precedence, see Collation precedence in multi-string operations.

COLLATE clause for table column definitions¶

Adding the optional COLLATE clause to the definition of a table column indicates that the specified collation is used for comparisons and other related operations performed on the data in the column:

CREATE TABLE <table_name> ( <col_name> <col_type> COLLATE '<collation_specification>'
                            [ , <col_name> <col_type> COLLATE '<collation_specification>' ... ]
                            [ , ... ]
                          )

If no COLLATE clause is specified for a column, Snowflake uses the default, which compares strings based on their UTF-8 character representations.

Also, Snowflake supports specifying an empty string for the collation specification (e.g. COLLATE ''), which is equivalent to specifying no collation for the column.

However, due to precedence, specifying COLLATE '' for a column does not have the same effect as explicitly specifying COLLATE 'utf8'. For more details, see Collation precedence in multi-string operations.

To see whether collation has been specified for the columns in a table, use DESCRIBE TABLE (or use the COLLATION function to view the collation, if any, for a specific column).

COLLATE function¶

The COLLATE function uses the specified collation on the input string expression:

COLLATE( <expression> , '<collation_specification>' )

This function can also be called using infix notation:

<expression> COLLATE '<collation_specification>'

This function is particularly useful for explicitly specifying a particular collation for a particular operation (e.g. sorting), but it can also be used to:

• Allow collation in the SELECT clause of a subquery, making all operations on the specified column in the outer query use the collation.

• Create a table using CTAS with a specified collation.

This example valuates using English case-insensitive collation:

SELECT * FROM t1 WHERE COLLATE(col1 , 'en-ci') = 'Tango';

This example sorts the results using German (Deutsch) collation:

SELECT * FROM t1 ORDER BY COLLATE(col1 , 'de');

This example creates a table with a column using French collation:

CREATE TABLE t2 AS SELECT COLLATE(col1, 'fr') AS col1 FROM t1;

This example uses infix notation to create a table with a column using French collation:

CREATE TABLE t2 AS SELECT col1 COLLATE 'fr' AS col1 FROM t1;

COLLATION function¶

The COLLATION function returns the collation specification used by an expression, including a table column:

COLLATION( <expression> )

If no collation has been specified for the expression, the function returns NULL.

Typically, if you use this function on a column name, it is best to use DISTINCT to avoid getting one row of output for each row in the table. For example:

SELECT DISTINCT COLLATION(column1) FROM table1;

Specification examples¶

Some examples of collation specification strings include:

• 'de': German (Deutsch) locale.

• 'de-ci-pi': German locale, with case-insensitive and punctuation-insensitive comparisons.

• 'fr_CA-ai': Canadian French locale, with accent-insensitive comparisons.

• 'en_US-trim': US English locale, with leading spaces and trailing spaces trimmed before the comparison.

You can also specify an empty string for a collation specification (e.g. COLLATE '' or COLLATE(col1, '')), which indicates to use no collation.

Supported specifiers¶

Locale:

Specifies the language-specific and country-specific rules to apply.

Supports valid locale strings, consisting of a language code (required) and country code (optional) in the form of language_country. Some locale examples include:

• en - English

• en_US - American English

• fr - French

• fr_CA - Canadian French

In addition, the utf8 pseudo-locale specifies Unicode ordering, which is the default. For more details, see Differences in sorting when using UTF-8 or locale collation (in this topic).

The locale specifier is optional, but, if used, must be the first specifier in the string.

For the full list of locales supported by Snowflake, see Collation locales supported by Snowflake.

Case-sensitivity:

Determines whether case is considered when comparing values. Possible values:

• cs - Case-sensitive (default)

• ci - Case-insensitive

For example:

Collation Specification	Value	Result
'en-ci'	Abc = abc	True
'en-cs' / en	Abc = abc	False

Accent-sensitivity:

Determines whether accented characters are considered equal to, or different from, their base characters. Possible values:

• as - Accent-sensitive (default)

• ai - Accent-insensitive

For example:

Collation Specification	Value	Result	Notes
'fr-ai'	E = É	True
'fr-as' / 'fr'	E = É	False
'en-ai'	a = ą	True	In English, these letters are treated as having only accent differences, so specifying accent-insensitivity results in the values comparing as equal.
'pl-ai'	a = ą	False	In Polish, these letters are treated as separate base letters, so they always compare as unequal regardless of whether accent-insensitivity is specified.
'pl-as' / 'pl'	a = ą	False

The rules for accent-sensitivity and collation vary between languages. For example, in some languages, collation is always accent-sensitive, and you cannot turn it off even by specifying accent-insensitive collation.

Punctuation-sensitivity:

Determines whether non-letter characters matter. Possible values:

• ps - Punctuation-sensitive.

• pi - Punctuation-insensitive.

Note that the default is locale-specific (i.e. if punctuation-sensitivity is not specified, locale-specific rules are used). In most cases, the rules are equivalent to ps.

For example:

Collation Specification	Value	Result	Notes
'en-pi'	A-B-C = ABC	True
'en-ps'	A-B-C = ABC	False

First-letter Preference:

Determines whether, when sorting, uppercase or lowercase letters are first. Possible values:

• fl - Lowercase letters sorted first.

• fu - Uppercase letters sorted first.

The default is locale-specific (i.e. if no value is specified, locale-specific ordering is used). In most cases, the ordering is equivalent to fl.

Also, this specifier has no impact on equality comparisons.

Case-conversion:

Results in strings being converted to lowercase or uppercase before comparisons. In some situations, this is faster than full locale-specific collation. Possible values:

• upper - Convert the string to uppercase before comparisons.

• lower - Convert the string to lowercase before comparisons.

This specifier does not have a default (i.e. if no value is specified, neither of the conversions occurs).

Space-trimming:

Removes leading/trailing spaces from strings before comparisons. This functionality can be useful for performing comparisons equivalent (except in extremely rare corner cases) in semantics to the SQL CHAR data type.

Possible values:

• trim - Remove both leading and trailing spaces before comparisons.

• ltrim - Remove only leading spaces before comparisons.

• rtrim - Remove only trailing spaces before comparisons.

This specifier does not have a default (i.e. if no value is specified, trimming is not performed).

For example:

Collation Specification	Value	Result	Notes
'en-trim'	__ABC_ = ABC	True	For the purposes of these examples, underscore characters represent blank spaces.
'en-ltrim'	__ABC_ = ABC	False
'en-rtrim'	__ABC_ = ABC	False
'en'	__ABC_ = ABC	False

Case-insensitive comparisons¶

The following sections describe case-insensitive comparisons:

• Differences when comparing uppercase strings and original strings

• Character weights

CREATE OR REPLACE TABLE test_table (col1 VARCHAR, col2 VARCHAR);
INSERT INTO test_table VALUES ('ı', 'i');

SELECT col1 = col2,
       COLLATE(col1, 'lower') = COLLATE(col2, 'lower'),
       COLLATE(col1, 'upper') = COLLATE(col2, 'upper')
  FROM test_table;

+-------------+-------------------------------------------------+-------------------------------------------------+
| COL1 = COL2 | COLLATE(COL1, 'LOWER') = COLLATE(COL2, 'LOWER') | COLLATE(COL1, 'UPPER') = COLLATE(COL2, 'UPPER') |
|-------------+-------------------------------------------------+-------------------------------------------------|
| False       | False                                           | True                                            |
+-------------+-------------------------------------------------+-------------------------------------------------+

Differences in sorting when using UTF-8 or locale collation¶

Strings are always stored internally in Snowflake in UTF-8, and can represent any character in any language supported by UTF-8. Therefore, the default collation is UTF-8 (i.e. 'utf8').

UTF-8 collation is based on the numeric representation of the character as opposed to the alphabetic order of the character.

This is analogous to sorting by the ordinal value of each ASCII character, which is important to note because uppercase letters have ordinal values lower than lowercase letters:

As a result:

• If you sort in UTF-8 order, all uppercase letters are returned before all lowercase letters:

  A , B , … , Y , Z , … , a , b , … , y , z

• In contrast, the 'en' collation specification sorts alphabetically (instead of using the UTF-8 internal representation), resulting in both A and a returned before both B and b:

  a , A , b , B , …

Additionally, the differences between the cs and ci case-sensitivity specifiers affect sorting:

• cs (case-sensitive) always returns the lowercase version of a letter before the uppercase version of the same letter. For example, using 'en-cs':

  a , A , b , B , …

  Case-sensitive is the default and, therefore, 'en-cs' and 'en' are equivalent.

• ci (case-insensitive) returns uppercase and lowercase versions of letters randomly with respect to each other, but still before both uppercase and lowercase version of later letters. For example, using 'en-ci':

  A , a , b , B , …

Some non-alphabetic characters can also be sorted differently depending upon the collation setting. The following example shows that the plus character (+) and minus character (-) are sorted differently for different collation settings:

Create the table:

CREATE OR REPLACE TABLE demo (
    no_explicit_collation VARCHAR,
    en_ci VARCHAR COLLATE 'en-ci',
    en VARCHAR COLLATE 'en',
    utf_8 VARCHAR collate 'utf8');
INSERT INTO demo (no_explicit_collation) VALUES
    ('-'),
    ('+');
UPDATE demo SET
    en_ci = no_explicit_collation,
    en = no_explicit_collation,
    utf_8 = no_explicit_collation;

Query the data:

SELECT MAX(no_explicit_collation), MAX(en_ci), MAX(en), MAX(utf_8)
  FROM demo;

+----------------------------+------------+---------+------------+
| MAX(NO_EXPLICIT_COLLATION) | MAX(EN_CI) | MAX(EN) | MAX(UTF_8) |
|----------------------------+------------+---------+------------|
| -                          | +          | +       | -          |
+----------------------------+------------+---------+------------+

Collation precedence in multi-string operations¶

When performing an operation on two (or more) strings, different collations might be specified for different strings. Determining the collation to apply depends on how collation was specified for each input and the precedence of each specifier.

There are three precedence levels (from highest to lowest):

Function:

Collation is specified using the COLLATE function in a SQL statement.

Column:

Collation was specified in the column definition.

None:

No collation is/was specified for a given expression/column, or collation with an empty specification is/was used (e.g. COLLATE(col1, '') or col1 STRING COLLATE '').

When determining the collation to use, the collation specification with the highest precedence is used. If multiple collations are specified with the same precedence level, their values are compared, and if they are not equal, an error is returned.

For example, consider a table with the following column-level collation specifications:

CREATE OR REPLACE TABLE collation_precedence_example(
  col1    VARCHAR,               -- equivalent to COLLATE ''
  col2_fr VARCHAR COLLATE 'fr',  -- French locale
  col3_de VARCHAR COLLATE 'de'   -- German locale
);

If the table is used in a statement comparing two strings, collation is applied as follows:

• This comparison uses the 'fr' collation because the precedence for col2_fr is higher than the precedence for col1:

  ... WHERE col1 = col2_fr ...
  

  Copy

• This comparison uses the 'en' collation, because it is explicitly specified in the statement, which takes precedence over the collation for col2_fr:

  ... WHERE col1 COLLATE 'en' = col2_fr ...
  

  Copy

• This comparison returns an error because the expressions have different collations at the same precedence level:

  ... WHERE col2_fr = col3_de ...
  

  Copy

• This comparison uses the 'de' collation because collation for col2_fr has been removed:

  ... WHERE col2_fr COLLATE '' = col3_de ...
  

  Copy

• This comparison returns an error because the expressions have different collations at the same precedence level:

  ... WHERE col2_fr COLLATE 'en' = col3_de COLLATE 'de' ...
  

  Copy

Even though Snowflake’s default collation is 'utf8', specifying an empty string (or specifying no collation) is different from explicitly specifying 'utf8' because explicit collation has higher precedence than no collation. The last two statements in the code examples below show the difference:

For example, consider a table with the following column-level collation specifications:

CREATE OR REPLACE TABLE collation_precedence_example2(
  s1 STRING COLLATE '',
  s2 STRING COLLATE 'utf8',
  s3 STRING COLLATE 'fr'
);

If the table is used in a statement comparing two strings, collation is applied as follows:

• This comparison uses 'utf8' because s1 has no collation and 'utf8' is the default:

  ... WHERE s1 = 'a' ...
  

  Copy

• This comparison uses 'utf8' because s1 has no collation and s2 has explicit 'utf8' collation

  ... WHERE s1 = s2 ...
  

  Copy

• This comparison executes without error because s1 has no collation and s3 has explicit fr collation, so the explicit collation takes precedence:

  ... WHERE s1 = s3 ...
  

  Copy

• This comparison causes an error because s2 and s3 have different collations specified at the same precedence level:

  ... WHERE s2 = s3 ...
  

  Copy

  002322 (42846): SQL compilation error: Incompatible collations: 'fr' and 'utf8'
  

Limited support for collation in built-in functions¶

Collation is supported in only a subset of string functions. Functions that could reasonably be expected to implement collation, but do not yet support collation, return an error when used with collation. These error messages are displayed not only when calling the COLLATE function, but also when calling a string function on a column that was defined as collated in the CREATE TABLE or ALTER TABLE statement that created that column.

Currently, collation influences only simple comparison operations. For example, POSITION('abc' in COLLATE('ABC', 'en-ci')) does not find abc in ABC, even though case-insensitive collation is specified.

Performance implications of using collation¶

Using collation can affect the performance of various database operations:

• Operations involving comparisons might be slower.

  This can impact simple WHERE clauses, as well as joins, sorts, GROUP BY operations, etc.

• When used with some functions in WHERE predicates, micro-partition pruning might be less efficient.

• Using collation in a WHERE predicate that is different from the collation specified for the column might result in reduced pruning efficiency or the complete elimination of pruning.

Additional considerations for using collation¶

• Remember that, despite the similarity in their names, the following collation functions return different results:

  • COLLATE explicitly specifies which collation to use.

  • COLLATION shows which collation is used if none is specified explicitly.

• A column with a collation specification can use characters that are not from the locale for the collation, which might impact sorting.

  For example, if a column is created with a COLLATE 'en' clause, the data in the column can contain the non-English character É. In this situation, the character É is sorted close to E.

• You can specify collation operations that are not necessarily meaningful.

  For example, you could specify that Polish data is compared to French data using German collation:

  SELECT ... WHERE COLLATE(French_column, 'de') = Polish_column;
  

  Copy

  However, Snowflake does not recommend using the feature this way because it might return unexpected or unintended results.

• After a table column is defined, you cannot change the collation for the column. In other words, after a column has been created with a particular collation using a CREATE TABLE statement, you cannot use ALTER TABLE to change the collation.

  However, you can specify a different collation in a DML statement, such as a SELECT statement, that references the column.

Differences in comparisons of widths, spaces, and scripts¶

During string comparisons, the ci collation specification recognizes that different visual representations of a character might still refer to the same character, and treats them accordingly. To allow for more performant comparisons, the upper and lower collation specifications do not recognize these different visual representations of a character as the same character.

Specifically, the ci collation specification ignores some differences in the following categories, while the upper and lower collation specifications do not ignore them:

• Character widths

• Types of spaces

• Character scripts

The following sections include examples that illustrate these differences.

CREATE OR REPLACE TABLE different_widths(codepoint STRING, description STRING);

INSERT INTO different_widths VALUES
  ('a', 'ASCII a'),
  ('A', 'ASCII A'),
  ('ａ', 'Full-width a'),
  ('Ａ', 'Full-width A');

SELECT codepoint VISUAL_CHAR,
       'U+'  || TO_CHAR(UNICODE(codepoint), '0XXX') codepoint_representation,
       description
  FROM different_widths;

+-------------+--------------------------+--------------+
| VISUAL_CHAR | CODEPOINT_REPRESENTATION | DESCRIPTION  |
|-------------+--------------------------+--------------|
| a           | U+0061                   | ASCII a      |
| A           | U+0041                   | ASCII A      |
| ａ          | U+FF41                   | Full-width a |
| Ａ          | U+FF21                   | Full-width A |
+-------------+--------------------------+--------------+

SELECT COUNT(*) NumRows,
       COUNT(DISTINCT UNICODE(codepoint)) DistinctCodepoints,
       COUNT(DISTINCT codepoint COLLATE 'en-ci') DistinctCodepoints_EnCi,
       COUNT(DISTINCT codepoint COLLATE 'upper') DistinctCodepoints_Upper,
       COUNT(DISTINCT codepoint COLLATE 'lower') DistinctCodepoints_Lower
  FROM different_widths;

+---------+--------------------+-------------------------+--------------------------+--------------------------+
| NUMROWS | DISTINCTCODEPOINTS | DISTINCTCODEPOINTS_ENCI | DISTINCTCODEPOINTS_UPPER | DISTINCTCODEPOINTS_LOWER |
|---------+--------------------+-------------------------+--------------------------+--------------------------|
|       4 |                  4 |                       1 |                        2 |                        2 |
+---------+--------------------+-------------------------+--------------------------+--------------------------+

CREATE OR REPLACE TABLE different_whitespaces(codepoint STRING, description STRING);

INSERT INTO different_whitespaces VALUES
  (' ', 'ASCII space'),
  ('\u00A0', 'Non-breaking space'),
  (' ', 'Ogham space mark'),
  (' ', 'en space'),
  (' ', 'em space');

SELECT codepoint visual_char,
       'U+'  || TO_CHAR(unicode(codepoint), '0XXX')
       codepoint_representation, description
  FROM different_whitespaces;

+-------------+--------------------------+--------------------+
| VISUAL_CHAR | CODEPOINT_REPRESENTATION | DESCRIPTION        |
|-------------+--------------------------+--------------------|
|             | U+0020                   | ASCII space        |
|             | U+00A0                   | Non-breaking space |
|             | U+1680                   | Ogham space mark   |
|             | U+2002                   | en space           |
|             | U+2003                   | em space           |
+-------------+--------------------------+--------------------+

SELECT COUNT(*) NumRows,
       COUNT(DISTINCT UNICODE(codepoint)) NumDistinctCodepoints,
       COUNT(DISTINCT codepoint COLLATE 'en-ci') DistinctCodepoints_EnCi,
       COUNT(DISTINCT codepoint COLLATE 'upper') DistinctCodepoints_Upper,
       COUNT(DISTINCT codepoint COLLATE 'lower') DistinctCodepoints_Lower
  FROM different_whitespaces;

+---------+-----------------------+-------------------------+--------------------------+--------------------------+
| NUMROWS | NUMDISTINCTCODEPOINTS | DISTINCTCODEPOINTS_ENCI | DISTINCTCODEPOINTS_UPPER | DISTINCTCODEPOINTS_LOWER |
|---------+-----------------------+-------------------------+--------------------------+--------------------------|
|       5 |                     5 |                       1 |                        5 |                        5 |
+---------+-----------------------+-------------------------+--------------------------+--------------------------+

CREATE OR REPLACE TABLE different_scripts(codepoint STRING, description STRING);

INSERT INTO different_scripts VALUES
  ('1', 'ASCII digit 1'),
  ('¹', 'Superscript 1'),
  ('₁', 'Subscript 1'),
  ('①', 'Circled digit 1'),
  ('੧', 'Gurmukhi digit 1'),
  ('௧', 'Tamil digit 1');

SELECT codepoint VISUAL_CHAR,
       'U+'  || TO_CHAR(UNICODE(codepoint), '0XXX') codepoint_representation,
       description
  FROM different_scripts;

+-------------+--------------------------+------------------+
| VISUAL_CHAR | CODEPOINT_REPRESENTATION | DESCRIPTION      |
|-------------+--------------------------+------------------|
| 1           | U+0031                   | ASCII digit 1    |
| ¹           | U+00B9                   | Superscript 1    |
| ₁           | U+2081                   | Subscript 1      |
| ①           | U+2460                   | Circled digit 1  |
| ੧           | U+0A67                   | Gurmukhi digit 1 |
| ௧           | U+0BE7                   | Tamil digit 1    |
+-------------+--------------------------+------------------+

SELECT COUNT(*) NumRows,
       COUNT(DISTINCT UNICODE(codepoint)) DistinctCodepoints,
       COUNT(DISTINCT codepoint COLLATE 'en-ci') DistinctCodepoints_EnCi,
       COUNT(DISTINCT codepoint COLLATE 'upper') DistinctCodepoints_Upper,
       COUNT(DISTINCT codepoint COLLATE 'lower') DistinctCodepoints_Lower
  FROM different_scripts;

+---------+--------------------+-------------------------+--------------------------+--------------------------+
| NUMROWS | DISTINCTCODEPOINTS | DISTINCTCODEPOINTS_ENCI | DISTINCTCODEPOINTS_UPPER | DISTINCTCODEPOINTS_LOWER |
|---------+--------------------+-------------------------+--------------------------+--------------------------|
|       6 |                  6 |                       1 |                        6 |                        6 |
+---------+--------------------+-------------------------+--------------------------+--------------------------+

Differences in handling ignorable code points¶

The Unicode Collation Algorithm specifies that collation elements (code points) can be ignorable, which means that a code point is not considered during string comparison and sorting.

• With the ci collation specification, these code points are ignored. This can make it difficult to search for or replace ignorable code points.

• With the upper and lower collation specifications, these code points are not ignored.

For example, the code point U+0001 is ignorable. If you compare this code point to an empty string with the en-ci collation specification, the result is TRUE because U+0001 is ignored:

SELECT '\u0001' = '' COLLATE 'en-ci';

+-------------------------------+
| '\U0001' = '' COLLATE 'EN-CI' |
|-------------------------------|
| True                          |
+-------------------------------+

On the other hand, if you use the upper or lower collation specification, the result is FALSE because U+0001 is not ignored:

SELECT '\u0001' = '' COLLATE 'upper';

+-------------------------------+
| '\U0001' = '' COLLATE 'UPPER' |
|-------------------------------|
| False                         |
+-------------------------------+

Similarly, suppose that you call the REPLACE function to remove this code point from a string. If you use the en-ci collation specification, the function does not remove the code point because U+0001 is ignored.

As shown in the example below, the string returned by the REPLACE function has the same length as the string passed into the function because the function does not remove the U+0001 character.

SELECT
  LEN('abc\u0001') AS original_length,
  LEN(REPLACE('abc\u0001' COLLATE 'en-ci', '\u0001')) AS length_after_replacement;

+-----------------+--------------------------+
| ORIGINAL_LENGTH | LENGTH_AFTER_REPLACEMENT |
|-----------------+--------------------------|
|               4 |                        4 |
+-----------------+--------------------------+

On the other hand, if you use the upper or lower collation specification, the function removes the code point from the string, returning a shorter string.

SELECT
  LEN('abc\u0001') AS original_length,
  LEN(REPLACE('abc\u0001' COLLATE 'upper', '\u0001')) AS length_after_replacement;

+-----------------+--------------------------+
| ORIGINAL_LENGTH | LENGTH_AFTER_REPLACEMENT |
|-----------------+--------------------------|
|               4 |                        3 |
+-----------------+--------------------------+

Differences when characters are represented by different code points¶

In Unicode, different sequences of code points can represent the same character. For example, the Greek Small Letter Iota with Dialytika and Tonos can be represented by the precomposed character with the code point U+0390 or by the sequence of code points U+03b9 U+0308 U+0301 for the decomposed characters.

If you use the ci collation specification, the different sequences of code points for a character are treated as the same character. For example, the code point U+0390 and the sequence of code points U+03b9 U+0308 U+0301 are treated as equivalent:

SELECT '\u03b9\u0308\u0301' = '\u0390' COLLATE 'en-ci';

+-------------------------------------------------+
| '\U03B9\U0308\U0301' = '\U0390' COLLATE 'EN-CI' |
|-------------------------------------------------|
| True                                            |
+-------------------------------------------------+

In order to improve performance for the upper and lower collation specifications, the sequences are not handled in the same way. Two sequences of code points are considered to be equivalent only if they result in the same binary representation after they are converted to uppercase or lowercase.

For example, using the upper specification with the code point U+0390 and the sequence of code points U+03b9 U+0308 U+0301 results in characters that are treated as equal:

SELECT '\u03b9\u0308\u0301' = '\u0390' COLLATE 'upper';

+-------------------------------------------------+
| '\U03B9\U0308\U0301' = '\U0390' COLLATE 'UPPER' |
|-------------------------------------------------|
| True                                            |
+-------------------------------------------------+

Using the lower specification results in characters that are not equal:

SELECT '\u03b9\u0308\u0301' = '\u0390' COLLATE 'lower';

+-------------------------------------------------+
| '\U03B9\U0308\U0301' = '\U0390' COLLATE 'LOWER' |
|-------------------------------------------------|
| False                                           |
+-------------------------------------------------+

These differences are less likely to occur when using upper (rather than lower) because there is only one composite uppercase code point (U+0130), compared to over 100 composite lowercase code points.

Differences with sequences of code points representing a single character¶

In cases where a sequence of code points represents a single character, the ci collation specification recognizes that the sequence represents a single character and does not match individual code points in the sequence.

For example, the sequence of code points U+03b9 U+0308 U+0301 represents a single character (the Greek Small Letter Iota with Dialytika and Tonos). U+0308 and U+0301 represent accents applied to U+03b9.

For the ci collation specification, if you use the CONTAINS function to determine if the sequence U+03b9 U+0308 contains U+03b9 or U+0308, the function returns FALSE because the sequence U+03b9 U+0308 is treated as a single character:

SELECT CONTAINS('\u03b9\u0308', '\u03b9' COLLATE 'en-ci');

+----------------------------------------------------+
| CONTAINS('\U03B9\U0308', '\U03B9' COLLATE 'EN-CI') |
|----------------------------------------------------|
| False                                              |
+----------------------------------------------------+

SELECT CONTAINS('\u03b9\u0308', '\u0308' COLLATE 'en-ci');

+----------------------------------------------------+
| CONTAINS('\U03B9\U0308', '\U0308' COLLATE 'EN-CI') |
|----------------------------------------------------|
| False                                              |
+----------------------------------------------------+

To improve performance, the upper and lower specifications do not treat these sequences as a single character. In the example above, the CONTAINS function returns TRUE because these specifications treat the sequence of code points as separate characters:

SELECT CONTAINS('\u03b9\u0308', '\u03b9' COLLATE 'upper');

+----------------------------------------------------+
| CONTAINS('\U03B9\U0308', '\U03B9' COLLATE 'UPPER') |
|----------------------------------------------------|
| True                                               |
+----------------------------------------------------+

SELECT CONTAINS('\u03b9\u0308', '\u0308' COLLATE 'upper');

+----------------------------------------------------+
| CONTAINS('\U03B9\U0308', '\U0308' COLLATE 'UPPER') |
|----------------------------------------------------|
| True                                               |
+----------------------------------------------------+

Differences when changes to case result in multiple code points¶

For some composite characters, the uppercase or lowercase version of the character is represented by a sequence of code points. For example, the uppercase character for the German character ß is a sequence of two S characters (SS).

Even though ß and SS are equivalent, when you use the upper collation specification, searches of ß and SS return different results. Sequences produced by case conversion either match in their entirety or not at all.

SELECT CONTAINS('ß' , 's' COLLATE 'upper');

+--------------------------------------+
| CONTAINS('SS' , 'S' COLLATE 'UPPER') |
|--------------------------------------|
| False                                |
+--------------------------------------+

SELECT CONTAINS('ss', 's' COLLATE 'upper');

+-------------------------------------+
| CONTAINS('SS', 'S' COLLATE 'UPPER') |
|-------------------------------------|
| True                                |
+-------------------------------------+

Differences in sort order¶

Sorting for the upper and lower collation specifications works differently from sorting for the ci specification:

• With the ci specification, strings are sorted by collation key. In general, the collation key can account for case sensitivity, accent sensitivity, locale, etc.

• With the upper and lower specifications, strings are sorted by code point to improve performance.

For example, some characters within the ASCII range (such as + and -) sort differently:

SELECT '+' < '-' COLLATE 'en-ci';

+---------------------------+
| '+' < '-' COLLATE 'EN-CI '|
|---------------------------|
| False                     |
+---------------------------+

SELECT '+' < '-' COLLATE 'upper';

+---------------------------+
| '+' < '-' COLLATE 'UPPER' |
|---------------------------|
| True                      |
+---------------------------+

As another example, strings with ignored code points sort in a different order:

SELECT 'a\u0001b' < 'ab' COLLATE 'en-ci';

+-----------------------------------+
| 'A\U0001B' < 'AB' COLLATE 'EN-CI' |
|-----------------------------------|
| False                             |
+-----------------------------------+

SELECT 'a\u0001b' < 'ab' COLLATE 'upper';

+-----------------------------------+
| 'A\U0001B' < 'AB' COLLATE 'UPPER' |
|-----------------------------------|
| True                              |
+-----------------------------------+

In addition, emojis sort differently:

SELECT 'abc' < '❄' COLLATE 'en-ci';

+-----------------------------+
| 'ABC' < '❄' COLLATE 'EN-CI' |
|-----------------------------|
| False                       |
+-----------------------------+

SELECT 'abc' < '❄' COLLATE 'upper';

+-----------------------------+
| 'ABC' < '❄' COLLATE 'UPPER' |
|-----------------------------|
| True                        |
+-----------------------------+

Collation is supported only for strings up to 8 MB¶

Although the Snowflake VARCHAR data type supports strings up to 16 MB, Snowflake supports collation only when the resulting string is 8 MB or less. (Some collation operations can lengthen a string.)

Collation not supported with UDFs¶

Snowflake does not support collation with UDFs (user-defined functions):

• You cannot return a collated string value from a UDF; the server reports that the actual return type is incompatible with the declared return type.

• If you pass a collated string value to a UDF, the collation information is not passed; the UDF sees the string as an uncollated string.

Collation not supported for strings in VARIANT, ARRAY, or OBJECT values¶

Strings stored inside a VARIANT, OBJECT, or ARRAY value do not include a collation specification. Therefore:

• Comparison of these values always uses the 'utf8' collation.

• When a VARCHAR value with a collation specification is used to construct an ARRAY, OBJECT, or VARIANT value, the collation specification is not preserved.

• You can still compare a value stored inside an ARRAY, OBJECT, or VARIANT by extracting the value, casting to VARCHAR, and adding a collation specification. For example:

  COLLATE(VARIANT_COL:fld1::VARCHAR, 'en-ci') = VARIANT_COL:fld2::VARCHAR
  

  Copy