10.10.1 Unicode Character Sets

MySQL implements the xxx_unicode_ci collations according to the Unicode Collation Algorithm (UCA) described at http://www.unicode.org/reports/tr10/. The collation uses the version-4.0.0 UCA weight keys: http://www.unicode.org/Public/UCA/4.0.0/allkeys-4.0.0.txt. The xxx_unicode_ci collations have only partial support for the Unicode Collation Algorithm. Some characters are not supported, and combining marks are not fully supported. This affects primarily Vietnamese, Yoruba, and some smaller languages such as Navajo. A combined character is considered different from the same character written with a single unicode character in string comparisons, and the two characters are considered to have a different length (for example, as returned by the CHAR_LENGTH() function or in result set metadata).

Unicode collations based on UCA versions higher than 4.0.0 include the version in the collation name. Examples:

The LOWER() and UPPER() functions perform case folding according to the collation of their argument. A character that has uppercase and lowercase versions only in a Unicode version higher than 4.0.0 is converted by these functions only if the argument collation uses a high enough UCA version.

Collations based on UCA 9.0.0 and higher are faster than collations based on UCA versions prior to 9.0.0. They also have a pad attribute of NO PAD, in contrast to PAD SPACE as used in collations based on UCA versions prior to 9.0.0. For comparison of nonbinary strings, NO PAD collations treat spaces at the end of strings like any other character (see Trailing Space Handling in Comparisons).

To determine the pad attribute for a collation, use the INFORMATION_SCHEMA COLLATIONS table, which has a PAD_ATTRIBUTE column. For example:

Comparison of nonbinary string values (CHAR, VARCHAR, and TEXT) that have a NO PAD collation differ from other collations with respect to trailing spaces. For example, 'a' and 'a ' compare as different strings, not the same string. This can be seen using the binary collations for utf8mb4. The pad attribute for utf8mb4_bin is PAD SPACE, whereas for utf8mb4_0900_bin it is NO PAD. Consequently, operations involving utf8mb4_0900_bin do not add trailing spaces, and comparisons involving strings with trailing spaces may differ for the two collations:

MySQL implements language-specific Unicode collations if the ordering based only on the Unicode Collation Algorithm (UCA) does not work well for a language. Language-specific collations are UCA-based, with additional language tailoring rules. Examples of such rules appear later in this section. For questions about particular language orderings, unicode.org provides Common Locale Data Repository (CLDR) collation charts at http://www.unicode.org/cldr/charts/30/collation/index.html.

For example, the nonlanguage-specific utf8mb4_0900_ai_ci and language-specific utf8mb4_LOCALE_0900_ai_ci Unicode collations each have these characteristics:

A collation name that includes a locale code or language name shown in the following table is a language-specific collation. Unicode character sets may include collations for one or more of these languages.

Croatian collations are tailored for these Croatian letters: Č, Ć, Dž, Đ, Lj, Nj, Š, Ž.

Danish collations may also be used for Norwegian.

For Japanese, the utf8mb4 character set includes utf8mb4_ja_0900_as_cs and utf8mb4_ja_0900_as_cs_ks collations. Both collations are accent sensitive and case-sensitive. utf8mb4_ja_0900_as_cs_ks is also kana sensitive and distinguishes Katakana characters from Hiragana characters, whereas utf8mb4_ja_0900_as_cs treats Katakana and Hiragana characters as equal for sorting. Applications that require a Japanese collation but not kana sensitivity may use utf8mb4_ja_0900_as_cs for better sort performance. utf8mb4_ja_0900_as_cs uses three weight levels for sorting; utf8mb4_ja_0900_as_cs_ks uses four.

For Classical Latin collations that are accent insensitive, I and J compare as equal, and U and V compare as equal. I and J, and U and V compare as equal on the base letter level. In other words, J is regarded as an accented I, and U is regarded as an accented V.

Spanish collations are available for modern and traditional Spanish. For both, ñ (n-tilde) is a separate letter between n and o. In addition, for traditional Spanish, ch is a separate letter between c and d, and ll is a separate letter between l and m.

Traditional Spanish collations may also be used for Asturian and Galician.

Swedish collations include Swedish rules. For example, in Swedish, the following relationship holds, which is not something expected by a German or French speaker:

Ü = Y < Ö

For any Unicode character set, operations performed using the xxx_general_ci collation are faster than those for the xxx_unicode_ci collation. For example, comparisons for the utf8_general_ci collation are faster, but slightly less correct, than comparisons for utf8_unicode_ci. The reason is that utf8_unicode_ci supports mappings such as expansions; that is, when one character compares as equal to combinations of other characters. For example, ß is equal to ss in German and some other languages. utf8_unicode_ci also supports contractions and ignorable characters. utf8_general_ci is a legacy collation that does not support expansions, contractions, or ignorable characters. It can make only one-to-one comparisons between characters.

To further illustrate, the following equalities hold in both utf8_general_ci and utf8_unicode_ci (for the effect of this in comparisons or searches, see Section 10.8.6, “Examples of the Effect of Collation”):

Ä = A
Ö = O
Ü = U

A difference between the collations is that this is true for utf8_general_ci:

ß = s

Whereas this is true for utf8_unicode_ci, which supports the German DIN-1 ordering (also known as dictionary order):

ß = ss

MySQL implements utf8 language-specific collations if the ordering with utf8_unicode_ci does not work well for a language. For example, utf8_unicode_ci works fine for German dictionary order and French, so there is no need to create special utf8 collations.

utf8_general_ci also is satisfactory for both German and French, except that ß is equal to s, and not to ss. If this is acceptable for your application, you should use utf8_general_ci because it is faster. If this is not acceptable (for example, if you require German dictionary order), use utf8_unicode_ci because it is more accurate.

If you require German DIN-2 (phone book) ordering, use the utf8_german2_ci collation, which compares the following sets of characters equal:

Ä = Æ = AE
Ö = Œ = OE
Ü = UE
ß = ss

utf8_german2_ci is similar to latin1_german2_ci, but the latter does not compare Æ equal to AE or Œ equal to OE. There is no utf8_german_ci corresponding to latin1_german_ci for German dictionary order because utf8_general_ci suffices.

A character's collating weight is determined as follows:

Collating weights can be displayed using the WEIGHT_STRING() function. (See Section 12.5, “String Functions and Operators”.) If a collation uses a weight lookup table, but a character is not in the table (for example, because it is a “new” character), collating weight determination becomes more complex:

There is a difference between “ordering by the character's code value” and “ordering by the character's binary representation,” a difference that appears only with utf16_bin, because of surrogates.

Suppose that utf16_bin (the binary collation for utf16) was a binary comparison “byte by byte” rather than “character by character.” If that were so, the order of characters in utf16_bin would differ from the order in utf8_bin. For example, the following chart shows two rare characters. The first character is in the range E000-FFFF, so it is greater than a surrogate but less than a supplementary. The second character is a supplementary.

Code point  Character                    utf8         utf16
----------  ---------                    ----         -----
0FF9D       HALFWIDTH KATAKANA LETTER N  EF BE 9D     FF 9D
10384       UGARITIC LETTER DELTA        F0 90 8E 84  D8 00 DF 84

The two characters in the chart are in order by code point value because 0xff9d < 0x10384. And they are in order by utf8 value because 0xef < 0xf0. But they are not in order by utf16 value, if we use byte-by-byte comparison, because 0xff > 0xd8.

So MySQL's utf16_bin collation is not “byte by byte.” It is “by code point.” When MySQL sees a supplementary-character encoding in utf16, it converts to the character's code-point value, and then compares. Therefore, utf8_bin and utf16_bin are the same ordering. This is consistent with the SQL:2008 standard requirement for a UCS_BASIC collation: “UCS_BASIC is a collation in which the ordering is determined entirely by the Unicode scalar values of the characters in the strings being sorted. It is applicable to the UCS character repertoire. Since every character repertoire is a subset of the UCS repertoire, the UCS_BASIC collation is potentially applicable to every character set. NOTE 11: The Unicode scalar value of a character is its code point treated as an unsigned integer.”

If the character set is ucs2, comparison is byte-by-byte, but ucs2 strings should not contain surrogates, anyway.

The xxx_general_mysql500_ci collations preserve the pre-5.1.24 ordering of the original xxx_general_ci collations and permit upgrades for tables created before MySQL 5.1.24 (Bug #27877).