Steven Haryanto


Sah - Schema for data structures (specification)


version 0.9.21




In the 0.9.0 series, there will probably still be incompatible syntax changes between revision before the spec stabilizes into 1.0 series.


This document specifies Sah, a schema language for validating data structures.

Data structures are written in pseudo-JSON (JSON with comments // ..., ellipsis ..., or some JavaScript).


Although it can contain extra stuffs, a schema is essentially a type definition, stating a set of valid values for data.

Sah schemas are regular data structures, specifically arrays:


TYPE_NAME is a string, CLAUSE_SET is a hash of clauses, and EXTRAS is a hash and is optional. Some examples:

 ["int", {"min": 0, "max": 100}]

 // a definition of pos_even (positive even natural numbers). "pos" is defined
 // in the EXTRAS part.
 ["pos", {"div_by": 2}, {"def": {"pos": ["int": {"min": 0}]}}]

A shortcut string form containing only the type name is allowed when there are no clauses. It will be normalized into the array form:


The type name can have a * suffix as a shortcut for the "req": 1 clause. This shortcut exists because stating something is required is very common.


 // equivalent to
 ["int", {"req": 1}]

 ["int*", {"min": 0}]

 // equivalent to
 ["int", {"req": 1, "min": 0}]

A flattened array form is also supported when there are no EXTRAS. It will be normalized into the non-flattened form. This shortcut exists to save a couple of keystrokes :-) And also reduce the number of nested structure, which can get a bit unwieldy for complex schemas.

 ["int", "min", 1, "max", 10]

 // is equivalent to
 ["int", {"min": 1, "max": 10}]


Type classifies data and specifies the possible values of data.

Sah defines several standard types like bool, int, float, str, array, hash, and a few others. Please see Sah::Type for the complete list.

Type name must match this regular expression:


A type can have clauses. Most clauses declare constraints (thus, constraint clauses). Constraint clauses are like functions, they accept an argument, are evaluated against data and return a value. The returned value need not strictly be boolean, but for the clause to succeed, the return value must evaluate to true. The notion of true/false follows Perl's notion: undefined value, empty string (""), the string "0", and number 0 are considered false. Everything else is true.

For the schema to succeed, all constraint clauses must evaluate to true.

Aside from declaring constraints, clauses can also declare other stuffs. There is the default clause which specifies default value. There are metadata clauses which specify metadata, e.g. the summary, description, tags clauses.

Aside from clauses, type can also have type properties. Properties are different from clauses in the following ways: 1) they are used to find out something about the data, not to test/validate data; 2) they are allowed to not accept any argument. A type can have a property and a clause with the same name, for example the str type have a len clause to test its length against an integer, as well as a len property which returns its length. Properties are differentiated from clauses so that compilers to human text can generate a description like "string where its length is at least 1".

Type properties can be validated against a schema using the prop or if clause.

Base schema. You can define a schema, declare it as a new type, and then write subsequent schemas against that type, along with additional clauses. This is very much like subtyping. See "BASE SCHEMA" for more information.


As mentioned before, you can define a schema as a type and then write other schemas against that type. For example:

 // defined as pos_int type
 ["int", {"min": 0}]

and later:

 // a positive integer, divisible by 5
 ["pos_int", {"div_by": 5}]

During data validation, base schemas will be replaced by its original definition, and all the clause sets will be evaluated. Illustrated by the plus sign:

 ["int", {"min": 0} + {"div_by": 5}]

You can also declare base schemas/types locally using the def key in EXTRAS, for example:

 ["throws", {},
      "def": {
          "single_dice_throw":  ["int": {"in": [1, 2, 3, 4, 5, 6]}],
          "sdt":                "single_dice_throw", // short notation
          "dice_pair_throw":    ["array": {"len": 2, "elems": ["sdt", "sdt"]}],
          "dpt":                "dice_pair_throw",   // short notation
          "throw":              ["any": {"of": ["sdt", "dpt"]}],
          "throws":             ["array": {"of": "throw"}],

The above schema describes a list of dice throws (throws). Each throw can be a single dice throw (sdt) which is a number between 1 and 6, or a throw of two dices (dpt) which is a 2-element array (where each element is a number between 1 and 6).

Examples of valid data for this schema:

 [1, [1,3], 6, 4, 2, [3,5]]

Examples of invalid data:

 1                  // not an array
 [1, [2, 3], 0]     // the third throw is invalid
 [1, [2, 0, 4], 4]  // the second throw is invalid

All the base schemas names throw, throws, sdt, etc is only declared locally and unknown outside the schema. You can even nest this.

Optional/conditional definition

If you put a ? suffix after the definition name then it means that the definition is optional and can be skipped if the type is already defined, e.g.:

  "def": {
      "emailaddr?": ["str", {"req": 1, "match": ".+\@.+"}],
      "username":   ["str", {"req": 1, "match": "^[a-z0-9_]+$"}]

In the above example, if there is already an emailaddr type defined at that time, the definition will be skipped instead of a "cannot redefine type" error being generated.

Optional definition is useful if you want to provide some defaults (e.g. a rudimentary validation for email address) but don't mind if the validator already has something probably better (a stricter or more precise definition of email address).


A clause set is a defhash (see DefHash) containing a mapping of clause name and clause values or clause attribute names and clause attribute values. Defhash properties map to Sah clauses, while defhash property attributes map to Sah clause attributes.

     "_IGNORED": ...,

For convenience, there are also some shortcuts:

  • & suffix (multiple clause values, all must succeed)

     "CLAUSENAME&": [VAL, ...]

    is equivalent to:

     "CLAUSENAME":    [VAL, ...],
     "CLAUSENAME.op": "and"
  • | suffix (multiple clause values, only one must succeed)

     "CLAUSENAME|": [VAL, ...]

    is equivalent to:

     "CLAUSENAME":    [VAL, ...],
     "CLAUSENAME.op": "or"
  • ! prefix (negation)


    is a shortcut for this:

     "CLAUSENAME.op": "not"
  • = suffix (expression)


    are respectively equivalent to:

     "CLAUSENAME.is_expr": 1
     "CLAUSENAME.ATTRNAME1.is_expr": 1
  • (LANG) suffix (value for alternate languages)


    are respectively equivalent to:

     "CLAUSENAME.alt.lang.LANG": VAL


     "name(id_ID)": "bilangan bulat positif"
     "name(en_US)": ["positive integer", "positive integers"]

    are equivalent to:

     "name.alt.lang.id_ID": "bilangan bulat positif"
     "name.alt.lang.en_US": ["positive integer", "positive integers"]

Every clause has a priority between 0 and 100 to determine the order of evaluation (the lower the number, the higher the priority and the earlier the clause is evaluated). Most constraint clauses are at priority 50 (normal) so the order does not matter, but some clauses are early (like default and prefilters) and some are late (like postfilters). Variables mentioned in expression also determine ordering, for example:

 ["int", {"min=": "0.5*$clause:max", "max": 10}]

In the above example, although max and min are both at priority 50, min needs to be evaluated first because it refers to max (XXX syntax of variable not yet finalized).

Clause name

This specification comes from DefHash: Clause names must begin with letter/underscore and contain letters/numbers/underscores only. All clauses which begin with an _ (underscore) is ignored. You can use this to embed extra data for other purposes.

Clause attribute

This specification comes from DefHash: Attribute name must also only contain letters/numbers/underscores, but it can be a dotted-separated series of parts, e.g. alt.lang.id_ID. As with clauses, clause attributes which begin with _ (underscore) is ignored. You can use this to embed extra data.

Currently known general attributes:

  • prio : INT

    Change the clause's priority for this clause set. Note that this only works for clauses which have equal priorities. Otherwise, priority value from clause definition takes precedence.


     // both "min" and "max" clauses have priority of 50, but we want to make sure
     // that "min" is evaluated first
     ["int*", {"min=": "some expr", "min.prio": 1, "max": 10}]
  • is_multi : BOOL (default: 0)

    This attribute can be used to signify that clause contain multiple (an array of) values instead of a single value. Each value will be tested against the clause. Example:

     ["int*", {"div_by.is_multi": 1, "div_by": [2, 3, 5]}]

    The above schema requires an integer which is divisible by 2, 3, and 5.

  • op : STR

    Specify operator for (multiple) clause values. Possible values for this attribute include: and, or, none, not. Except for not, the presence of op signifies that clause contains multiple values instead of a single one.

    There are shortcuts for and, or, and not; see "CLAUSE AND CLAUSE SET".

    and specifies that all clause values must succeed for the clause to succeed. Example:

     ["str", {"clause": [["min_len", 8], ["match", "\\W"]], "clause.op": "and"}]

    The above schema requires a string to be at least 8 characters long, or contains a non-word character. Strings that would validate include: abcdefgh or $ or $abcdefg. Strings that would not validate include: abcd (fails both min_len and match clauses) or abcdefgh (fails the match clause).

    or specifies that any one of clause values must succeed for the clause to succeed. Example:

     ["str", {"match": [RE1, RE2, RE3], "match.op": "or"}]

    The above schema specifies that string can match any of the regexes RE1/RE2/RE3.

    none specifies that all clause values must fail for the clause to succeed. For example:

     ["str", {"match": [RE1, RE2, RE3], "match.op": "none"}]

    The above schema specifies that string must not match any of the regexes RE1/RE2/RE3.

    not reverts the success status of clause (in other words, clause must fail for validation to succeed). Example:

     ["str", {"match": RE, "match.op", "not"}]

    The above schema specifies not string must not match regex RE.

  • is_expr : BOOL

    Signify that clause contains expression (see "EXPRESSION") instead of literal value. Example:

     // a string, minimum 4 characters
     ["str", {"min_len": 4}]
     // same thing, albeit a bit fancier
     ["str", {"min_len.is_expr": 1, "min_len": "2*2"}]
     // same thing, shortcut notation
     ["str", {"min_len=": "2*2"}]
     // for default, we pick a random number between 1 and 10
     ["int", {"default=": "int(10*rand())+1"}]

    Expression is useful for more complex schema, when a clause/attribute value needs to be calculated in terms of other values, and/or using functions.

    Note that an implementation might not support expression in some clauses or attributes, especially clauses that accept argument containing schemas as dynamically generated schemas needs the compiler to embed an interpreter or compiler in the generated code.

    When is_expr attribute is true, and is_multi is also true, then the expression is expected to return an array of values. Otherwise, the clause will fail. Example:

     // number which must be divisible by 2, 3, 5
     ["int", {"div_by.is_expr": 1, "div_by.is_multi": 1, "div_by": "[2, 3, 5]"}]
     // string must not match any of the blacklist
     ["str", {
         "contains.is_expr": 1,
         "contains.op": "none",
         "contains": "get_blacklist()"
  • err_level : STR (default: error)

    Valid value: fatal, error, warn. Normally, when clause checking fails, an error is generated and it causes validation of the whole schema to fail. If err_level is set to warn, however, this only generates a warning and does not cause the validation to fail.

     // password
     ["str*", {"cset&": [
       {"min_len": 4},
       {"min_len": 8,
        "min_len.err_level": "warn",
        "min_len.err_msg":   "Although a password less than 8 letters are " +
                             "valid it's highly recommended that a password is " +
                             "at least 8 letters long, for security reasons"}

    In the above example, the err_level and err_msg are attributes for the min_len clause. The second clause set basically adds an optional restriction for the password: when the min_len clause is not satisfied, instead of making the data fails the validation, only a warning is issued.

    fatal is the same as error but will make validation exit early, without collecting further errors. This only takes effect when validation collects full errors instead of just stopping after the first error is found.

  • err_msg[.LANGCODE]

    This tells the compiler that instead of the default error message from the type handler, a custom error message is supplied. You can add translations by adding more attributes with language code suffixes. For example:

     ["str", {"match":                "[^A-Za-z0-9_-]",
              "match.err_msg":        "Must not contain naughty characters",
              "match.err_msg.id_ID": "Tidak boleh mengandung karakter aneh-aneh"

    Another example:

     ["str", {"!in": ["root", "admin"],
              "in.err_msg":       "Sorry, username is reserved",
              "in.err_msg.id_ID": "Maaf, nama user dilarang digunakan"
  • human[.LANGCODE]

    This is also ignored when validating data, but will be used by the human compiler to supply description. You can add translations by adding more attributes.

     ["str", {"match":             "[^A-Za-z0-9_-]",
              "match.human"        "Must not contain naughty characters",
              "match.human.id_ID": "Tidak boleh mengandung karakter aneh-aneh"
  • alt

    This comes from DefHash, mainly used to store translations for name, summary, description.

  • result_var : VARNAME (EXPERIMENTAL)

    Specify variable name to store results in.

    Aside from pass/failure, a clause or clause set can also produce some value. This attribute specifies where to put the results in. The value can then be used by referring to the variable in expression. Example:

     ["any", {
         "of": [
             ["str*":    {"min_len": 1, "max_len": 10}], // 1
             ["str*":    {"min_len": 11}],               // 2
             ["array*":  {}],                            // 3
             ["hash*":   {}]                             // 4
        "of.result_var": "a"

    Aside from passing/failing the validation, the of clause above also produces an index to the schema in the list which matches. So if you validate an array, $a in the schema will be set to 3. If you validate a string with length 12, $a will be set to 2. If you pass an empty string (which does not pass the of clause, $a will not be set.

    Refer to each clause's documentation to find out what value the clause returns.

Aside from the above general attributes, each clause might recognize its own specific attributes. See documentation of respective clauses.


    This is a namespace for specifying compiler options. Each compiler will have its specific options; see documentation on respective compiler to see available options. For example:

     // skip clauses which are not implemented in JavaScript. we'll check on the
     // server-side anyway.
     ["str", {
       "soundex": "E460",
       "c.js.ignore_missing_clause_handler": true

Clause set merging

Clause set merging happens when a schema is based on another schema and the child schema's clause set contains merge prefixes (explained later) in its keys. For example:

 // schema1

 // schema2, based on schema1
 [schema1, CSET2]

 // schema3, based on schema2
 [schema2, CSET3]

When compiling/evaluating schema2, Sah will check against TYPE1 and CSET1 and then CSET2. However, when CSET2 contains a merge prefix (marked with an asterisk here for illustration), then Sah will check against TYPE1 and merge(CSET1, *CSET2).

When compiling/evaluating schema3, Sah will check against TYPE1 and CSET1 and then CSET2 and then CSET3. However, when CSET2 contains a merge prefix, then Sah will check against TYPE1, merge(CSET1, *CSET2), and then CSET3. When CSET2 and CSET3 contains merge prefixes, Sah will check against TYPE1 and merge(CSET1, *CSET2, *CSET3). So merging will be done from left to right.

The base schema's clause set must not contain any merge prefixes.

Merging is done using Data::ModeMerge, with merge prefixes changed to 'merge.add.', 'merge.delete.' and so on. In merging, Data::ModeMerge allows keys on the right side hash not only to replace but also add, subtract, remove keys from the left side. This is powerful because it allows schema definition to not only add clauses (restrict types even more), but also replace clauses (change type restriction) as well as delete clauses (relax type restriction). For more information, refer to the Data::ModeMerge documentation.


 int + {"div_by": 2} + {"div_by": 3}               // must be divisible by 2 & 3

 int + {"div_by": 2} + {"merge.normal.div_by": 3} // will be merged and become:
 int + {"div_by": 3}                              // must be divisible by 3 ONLY

 int + {"div_by": 2} + {"merge.delete.div_by": 0}  // will be merged and become:
 int + {}                                          // need not be divisible

 int + {"in": [1,2,3,4,5]} + {"in": [6]}           // impossible to satisfy

 int + {"in": [1,2,3,4,5]} + {"": [6]} // will be merged and become:
 int + {"in": [1,2,3,4,5,6]}

 int + {"in": [1,2,3,4,5]}, {"": [4]}  // will become:
 int + {"in": [1,2,3,  5]}

Merging is performed before schema is normalized.

Merging is not recursive.


XXX: Syntax of variables not yet finalized.

Sah supports expressions, using Language::Expr minilanguage. See Language::Expr::Manual::Syntax for details on the syntax. You can specify expression in the check clause, e.g.:

 ["int", {"check": "$_ >= 4"}]

Alternatively, expression can also be specified in any clause's attribute:

 ["int", {"min=": "floor(4.9)"}]

The above three schemas are equivalent to:

 ["int", {"min", 4}]

Expression can refer to elements of data and (normalized) schema, and can call functions, enabling more complex schema to be defined, for example:

 ["array*", {"len": 2, "elems": [
   ["str*", {"match": "^\w+$"}],
   ["str*", {"match=": "${../../0/clause_sets/0/match}",
             "min_len=": "2*length(${data:../0})"}]

The above schema requires data to be a two-element array containing strings, where the length of the second string has to be at least twice the length of the first. Both strings have to comply to the same regex, ^\w+$ (which is declared on the first string's clause and referred to in the second string's clause).


Functions can be used in expressions. The syntax of calling function is:

 func(ARG, ...)

Functions in Sah can sometimes accept several types of arguments, e.g. len(ARRAY) will return the number of elements in the ARRAY, while len(STR) will return the number of characters in the string. However, when an inappropriate argument is given, an exception will be thrown.


2012-07-21 split specification to Sah

2011-11-23 Data::Sah

2009-03-30 Data::Schema (first CPAN release)

Previous incarnation as Schema-Nested (internal)



Sah::Type, Sah::FAQ


Steven Haryanto <>


This software is copyright (c) 2012 by Steven Haryanto.

This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.

2 POD Errors

The following errors were encountered while parsing the POD:

Around line 510:

'=item' outside of any '=over'

Around line 523:

You forgot a '=back' before '=head2'