++ed by:
48 non-PAUSE users
Author image Marc A. Lehmann


JSON::XS - JSON serialising/deserialising, done correctly and fast

JSON::XS - 正しくて高速な JSON シリアライザ/デシリアライザ (http://fleur.hio.jp/perldoc/mix/lib/JSON/XS.html)


 use JSON::XS;

 # exported functions, they croak on error
 # and expect/generate UTF-8

 $utf8_encoded_json_text = encode_json $perl_hash_or_arrayref;
 $perl_hash_or_arrayref  = decode_json $utf8_encoded_json_text;

 # OO-interface

 $coder = JSON::XS->new->ascii->pretty->allow_nonref;
 $pretty_printed_unencoded = $coder->encode ($perl_scalar);
 $perl_scalar = $coder->decode ($unicode_json_text);

 # Note that JSON version 2.0 and above will automatically use JSON::XS
 # if available, at virtually no speed overhead either, so you should
 # be able to just:
 use JSON;

 # and do the same things, except that you have a pure-perl fallback now.


This module converts Perl data structures to JSON and vice versa. Its primary goal is to be correct and its secondary goal is to be fast. To reach the latter goal it was written in C.

Beginning with version 2.0 of the JSON module, when both JSON and JSON::XS are installed, then JSON will fall back on JSON::XS (this can be overriden) with no overhead due to emulation (by inheritign constructor and methods). If JSON::XS is not available, it will fall back to the compatible JSON::PP module as backend, so using JSON instead of JSON::XS gives you a portable JSON API that can be fast when you need and doesn't require a C compiler when that is a problem.

As this is the n-th-something JSON module on CPAN, what was the reason to write yet another JSON module? While it seems there are many JSON modules, none of them correctly handle all corner cases, and in most cases their maintainers are unresponsive, gone missing, or not listening to bug reports for other reasons.

See COMPARISON, below, for a comparison to some other JSON modules.

See MAPPING, below, on how JSON::XS maps perl values to JSON values and vice versa.


  • correct Unicode handling

    This module knows how to handle Unicode, documents how and when it does so, and even documents what "correct" means.

  • round-trip integrity

    When you serialise a perl data structure using only datatypes supported by JSON, the deserialised data structure is identical on the Perl level. (e.g. the string "2.0" doesn't suddenly become "2" just because it looks like a number). There minor are exceptions to this, read the MAPPING section below to learn about those.

  • strict checking of JSON correctness

    There is no guessing, no generating of illegal JSON texts by default, and only JSON is accepted as input by default (the latter is a security feature).

  • fast

    Compared to other JSON modules and other serialisers such as Storable, this module usually compares favourably in terms of speed, too.

  • simple to use

    This module has both a simple functional interface as well as an objetc oriented interface interface.

  • reasonably versatile output formats

    You can choose between the most compact guaranteed-single-line format possible (nice for simple line-based protocols), a pure-ascii format (for when your transport is not 8-bit clean, still supports the whole Unicode range), or a pretty-printed format (for when you want to read that stuff). Or you can combine those features in whatever way you like.


The following convenience methods are provided by this module. They are exported by default:

$json_text = encode_json $perl_scalar

Converts the given Perl data structure to a UTF-8 encoded, binary string (that is, the string contains octets only). Croaks on error.

This function call is functionally identical to:

   $json_text = JSON::XS->new->utf8->encode ($perl_scalar)

except being faster.

$perl_scalar = decode_json $json_text

The opposite of encode_json: expects an UTF-8 (binary) string and tries to parse that as an UTF-8 encoded JSON text, returning the resulting reference. Croaks on error.

This function call is functionally identical to:

   $perl_scalar = JSON::XS->new->utf8->decode ($json_text)

except being faster.

$is_boolean = JSON::XS::is_bool $scalar

Returns true if the passed scalar represents either JSON::XS::true or JSON::XS::false, two constants that act like 1 and 0, respectively and are used to represent JSON true and false values in Perl.

See MAPPING, below, for more information on how JSON values are mapped to Perl.


Since this often leads to confusion, here are a few very clear words on how Unicode works in Perl, modulo bugs.

1. Perl strings can store characters with ordinal values > 255.

This enables you to store Unicode characters as single characters in a Perl string - very natural.

2. Perl does not associate an encoding with your strings.

... until you force it to, e.g. when matching it against a regex, or printing the scalar to a file, in which case Perl either interprets your string as locale-encoded text, octets/binary, or as Unicode, depending on various settings. In no case is an encoding stored together with your data, it is use that decides encoding, not any magical meta data.

3. The internal utf-8 flag has no meaning with regards to the encoding of your string.

Just ignore that flag unless you debug a Perl bug, a module written in XS or want to dive into the internals of perl. Otherwise it will only confuse you, as, despite the name, it says nothing about how your string is encoded. You can have Unicode strings with that flag set, with that flag clear, and you can have binary data with that flag set and that flag clear. Other possibilities exist, too.

If you didn't know about that flag, just the better, pretend it doesn't exist.

4. A "Unicode String" is simply a string where each character can be validly interpreted as a Unicode codepoint.

If you have UTF-8 encoded data, it is no longer a Unicode string, but a Unicode string encoded in UTF-8, giving you a binary string.

5. A string containing "high" (> 255) character values is not a UTF-8 string.

It's a fact. Learn to live with it.

I hope this helps :)


The object oriented interface lets you configure your own encoding or decoding style, within the limits of supported formats.

$json = new JSON::XS

Creates a new JSON::XS object that can be used to de/encode JSON strings. All boolean flags described below are by default disabled.

The mutators for flags all return the JSON object again and thus calls can be chained:

   my $json = JSON::XS->new->utf8->space_after->encode ({a => [1,2]})
   => {"a": [1, 2]}
$json = $json->ascii ([$enable])
$enabled = $json->get_ascii

If $enable is true (or missing), then the encode method will not generate characters outside the code range 0..127 (which is ASCII). Any Unicode characters outside that range will be escaped using either a single \uXXXX (BMP characters) or a double \uHHHH\uLLLLL escape sequence, as per RFC4627. The resulting encoded JSON text can be treated as a native Unicode string, an ascii-encoded, latin1-encoded or UTF-8 encoded string, or any other superset of ASCII.

If $enable is false, then the encode method will not escape Unicode characters unless required by the JSON syntax or other flags. This results in a faster and more compact format.

See also the section ENCODING/CODESET FLAG NOTES later in this document.

The main use for this flag is to produce JSON texts that can be transmitted over a 7-bit channel, as the encoded JSON texts will not contain any 8 bit characters.

  JSON::XS->new->ascii (1)->encode ([chr 0x10401])
  => ["\ud801\udc01"]
$json = $json->latin1 ([$enable])
$enabled = $json->get_latin1

If $enable is true (or missing), then the encode method will encode the resulting JSON text as latin1 (or iso-8859-1), escaping any characters outside the code range 0..255. The resulting string can be treated as a latin1-encoded JSON text or a native Unicode string. The decode method will not be affected in any way by this flag, as decode by default expects Unicode, which is a strict superset of latin1.

If $enable is false, then the encode method will not escape Unicode characters unless required by the JSON syntax or other flags.

See also the section ENCODING/CODESET FLAG NOTES later in this document.

The main use for this flag is efficiently encoding binary data as JSON text, as most octets will not be escaped, resulting in a smaller encoded size. The disadvantage is that the resulting JSON text is encoded in latin1 (and must correctly be treated as such when storing and transferring), a rare encoding for JSON. It is therefore most useful when you want to store data structures known to contain binary data efficiently in files or databases, not when talking to other JSON encoders/decoders.

  JSON::XS->new->latin1->encode (["\x{89}\x{abc}"]
  => ["\x{89}\\u0abc"]    # (perl syntax, U+abc escaped, U+89 not)
$json = $json->utf8 ([$enable])
$enabled = $json->get_utf8

If $enable is true (or missing), then the encode method will encode the JSON result into UTF-8, as required by many protocols, while the decode method expects to be handled an UTF-8-encoded string. Please note that UTF-8-encoded strings do not contain any characters outside the range 0..255, they are thus useful for bytewise/binary I/O. In future versions, enabling this option might enable autodetection of the UTF-16 and UTF-32 encoding families, as described in RFC4627.

If $enable is false, then the encode method will return the JSON string as a (non-encoded) Unicode string, while decode expects thus a Unicode string. Any decoding or encoding (e.g. to UTF-8 or UTF-16) needs to be done yourself, e.g. using the Encode module.

See also the section ENCODING/CODESET FLAG NOTES later in this document.

Example, output UTF-16BE-encoded JSON:

  use Encode;
  $jsontext = encode "UTF-16BE", JSON::XS->new->encode ($object);

Example, decode UTF-32LE-encoded JSON:

  use Encode;
  $object = JSON::XS->new->decode (decode "UTF-32LE", $jsontext);
$json = $json->pretty ([$enable])

This enables (or disables) all of the indent, space_before and space_after (and in the future possibly more) flags in one call to generate the most readable (or most compact) form possible.

Example, pretty-print some simple structure:

   my $json = JSON::XS->new->pretty(1)->encode ({a => [1,2]})
      "a" : [
$json = $json->indent ([$enable])
$enabled = $json->get_indent

If $enable is true (or missing), then the encode method will use a multiline format as output, putting every array member or object/hash key-value pair into its own line, indenting them properly.

If $enable is false, no newlines or indenting will be produced, and the resulting JSON text is guaranteed not to contain any newlines.

This setting has no effect when decoding JSON texts.

$json = $json->space_before ([$enable])
$enabled = $json->get_space_before

If $enable is true (or missing), then the encode method will add an extra optional space before the : separating keys from values in JSON objects.

If $enable is false, then the encode method will not add any extra space at those places.

This setting has no effect when decoding JSON texts. You will also most likely combine this setting with space_after.

Example, space_before enabled, space_after and indent disabled:

   {"key" :"value"}
$json = $json->space_after ([$enable])
$enabled = $json->get_space_after

If $enable is true (or missing), then the encode method will add an extra optional space after the : separating keys from values in JSON objects and extra whitespace after the , separating key-value pairs and array members.

If $enable is false, then the encode method will not add any extra space at those places.

This setting has no effect when decoding JSON texts.

Example, space_before and indent disabled, space_after enabled:

   {"key": "value"}
$json = $json->relaxed ([$enable])
$enabled = $json->get_relaxed

If $enable is true (or missing), then decode will accept some extensions to normal JSON syntax (see below). encode will not be affected in anyway. Be aware that this option makes you accept invalid JSON texts as if they were valid!. I suggest only to use this option to parse application-specific files written by humans (configuration files, resource files etc.)

If $enable is false (the default), then decode will only accept valid JSON texts.

Currently accepted extensions are:

  • list items can have an end-comma

    JSON separates array elements and key-value pairs with commas. This can be annoying if you write JSON texts manually and want to be able to quickly append elements, so this extension accepts comma at the end of such items not just between them:

          2, <- this comma not normally allowed
          "k1": "v1",
          "k2": "v2", <- this comma not normally allowed
  • shell-style '#'-comments

    Whenever JSON allows whitespace, shell-style comments are additionally allowed. They are terminated by the first carriage-return or line-feed character, after which more white-space and comments are allowed.

         1, # this comment not allowed in JSON
            # neither this one...
$json = $json->canonical ([$enable])
$enabled = $json->get_canonical

If $enable is true (or missing), then the encode method will output JSON objects by sorting their keys. This is adding a comparatively high overhead.

If $enable is false, then the encode method will output key-value pairs in the order Perl stores them (which will likely change between runs of the same script).

This option is useful if you want the same data structure to be encoded as the same JSON text (given the same overall settings). If it is disabled, the same hash might be encoded differently even if contains the same data, as key-value pairs have no inherent ordering in Perl.

This setting has no effect when decoding JSON texts.

$json = $json->allow_nonref ([$enable])
$enabled = $json->get_allow_nonref

If $enable is true (or missing), then the encode method can convert a non-reference into its corresponding string, number or null JSON value, which is an extension to RFC4627. Likewise, decode will accept those JSON values instead of croaking.

If $enable is false, then the encode method will croak if it isn't passed an arrayref or hashref, as JSON texts must either be an object or array. Likewise, decode will croak if given something that is not a JSON object or array.

Example, encode a Perl scalar as JSON value with enabled allow_nonref, resulting in an invalid JSON text:

   JSON::XS->new->allow_nonref->encode ("Hello, World!")
   => "Hello, World!"
$json = $json->allow_blessed ([$enable])
$enabled = $json->get_allow_blessed

If $enable is true (or missing), then the encode method will not barf when it encounters a blessed reference. Instead, the value of the convert_blessed option will decide whether null (convert_blessed disabled or no TO_JSON method found) or a representation of the object (convert_blessed enabled and TO_JSON method found) is being encoded. Has no effect on decode.

If $enable is false (the default), then encode will throw an exception when it encounters a blessed object.

$json = $json->convert_blessed ([$enable])
$enabled = $json->get_convert_blessed

If $enable is true (or missing), then encode, upon encountering a blessed object, will check for the availability of the TO_JSON method on the object's class. If found, it will be called in scalar context and the resulting scalar will be encoded instead of the object. If no TO_JSON method is found, the value of allow_blessed will decide what to do.

The TO_JSON method may safely call die if it wants. If TO_JSON returns other blessed objects, those will be handled in the same way. TO_JSON must take care of not causing an endless recursion cycle (== crash) in this case. The name of TO_JSON was chosen because other methods called by the Perl core (== not by the user of the object) are usually in upper case letters and to avoid collisions with any to_json function or method.

This setting does not yet influence decode in any way, but in the future, global hooks might get installed that influence decode and are enabled by this setting.

If $enable is false, then the allow_blessed setting will decide what to do when a blessed object is found.

$json = $json->filter_json_object ([$coderef->($hashref)])

When $coderef is specified, it will be called from decode each time it decodes a JSON object. The only argument is a reference to the newly-created hash. If the code references returns a single scalar (which need not be a reference), this value (i.e. a copy of that scalar to avoid aliasing) is inserted into the deserialised data structure. If it returns an empty list (NOTE: not undef, which is a valid scalar), the original deserialised hash will be inserted. This setting can slow down decoding considerably.

When $coderef is omitted or undefined, any existing callback will be removed and decode will not change the deserialised hash in any way.

Example, convert all JSON objects into the integer 5:

   my $js = JSON::XS->new->filter_json_object (sub { 5 });
   # returns [5]
   $js->decode ('[{}]')
   # throw an exception because allow_nonref is not enabled
   # so a lone 5 is not allowed.
   $js->decode ('{"a":1, "b":2}');
$json = $json->filter_json_single_key_object ($key [=> $coderef->($value)])

Works remotely similar to filter_json_object, but is only called for JSON objects having a single key named $key.

This $coderef is called before the one specified via filter_json_object, if any. It gets passed the single value in the JSON object. If it returns a single value, it will be inserted into the data structure. If it returns nothing (not even undef but the empty list), the callback from filter_json_object will be called next, as if no single-key callback were specified.

If $coderef is omitted or undefined, the corresponding callback will be disabled. There can only ever be one callback for a given key.

As this callback gets called less often then the filter_json_object one, decoding speed will not usually suffer as much. Therefore, single-key objects make excellent targets to serialise Perl objects into, especially as single-key JSON objects are as close to the type-tagged value concept as JSON gets (it's basically an ID/VALUE tuple). Of course, JSON does not support this in any way, so you need to make sure your data never looks like a serialised Perl hash.

Typical names for the single object key are __class_whatever__, or $__dollars_are_rarely_used__$ or }ugly_brace_placement, or even things like __class_md5sum(classname)__, to reduce the risk of clashing with real hashes.

Example, decode JSON objects of the form { "__widget__" => <id> } into the corresponding $WIDGET{<id>} object:

   # return whatever is in $WIDGET{5}:
      ->filter_json_single_key_object (__widget__ => sub {
            $WIDGET{ $_[0] }
      ->decode ('{"__widget__": 5')

   # this can be used with a TO_JSON method in some "widget" class
   # for serialisation to json:
   sub WidgetBase::TO_JSON {
      my ($self) = @_;

      unless ($self->{id}) {
         $self->{id} = ..get..some..id..;
         $WIDGET{$self->{id}} = $self;

      { __widget__ => $self->{id} }
$json = $json->shrink ([$enable])
$enabled = $json->get_shrink

Perl usually over-allocates memory a bit when allocating space for strings. This flag optionally resizes strings generated by either encode or decode to their minimum size possible. This can save memory when your JSON texts are either very very long or you have many short strings. It will also try to downgrade any strings to octet-form if possible: perl stores strings internally either in an encoding called UTF-X or in octet-form. The latter cannot store everything but uses less space in general (and some buggy Perl or C code might even rely on that internal representation being used).

The actual definition of what shrink does might change in future versions, but it will always try to save space at the expense of time.

If $enable is true (or missing), the string returned by encode will be shrunk-to-fit, while all strings generated by decode will also be shrunk-to-fit.

If $enable is false, then the normal perl allocation algorithms are used. If you work with your data, then this is likely to be faster.

In the future, this setting might control other things, such as converting strings that look like integers or floats into integers or floats internally (there is no difference on the Perl level), saving space.

$json = $json->max_depth ([$maximum_nesting_depth])
$max_depth = $json->get_max_depth

Sets the maximum nesting level (default 512) accepted while encoding or decoding. If the JSON text or Perl data structure has an equal or higher nesting level then this limit, then the encoder and decoder will stop and croak at that point.

Nesting level is defined by number of hash- or arrayrefs that the encoder needs to traverse to reach a given point or the number of { or [ characters without their matching closing parenthesis crossed to reach a given character in a string.

Setting the maximum depth to one disallows any nesting, so that ensures that the object is only a single hash/object or array.

The argument to max_depth will be rounded up to the next highest power of two. If no argument is given, the highest possible setting will be used, which is rarely useful.

See SECURITY CONSIDERATIONS, below, for more info on why this is useful.

$json = $json->max_size ([$maximum_string_size])
$max_size = $json->get_max_size

Set the maximum length a JSON text may have (in bytes) where decoding is being attempted. The default is 0, meaning no limit. When decode is called on a string longer then this number of characters it will not attempt to decode the string but throw an exception. This setting has no effect on encode (yet).

The argument to max_size will be rounded up to the next highest power of two (so may be more than requested). If no argument is given, the limit check will be deactivated (same as when 0 is specified).

See SECURITY CONSIDERATIONS, below, for more info on why this is useful.

$json_text = $json->encode ($perl_scalar)

Converts the given Perl data structure (a simple scalar or a reference to a hash or array) to its JSON representation. Simple scalars will be converted into JSON string or number sequences, while references to arrays become JSON arrays and references to hashes become JSON objects. Undefined Perl values (e.g. undef) become JSON null values. Neither true nor false values will be generated.

$perl_scalar = $json->decode ($json_text)

The opposite of encode: expects a JSON text and tries to parse it, returning the resulting simple scalar or reference. Croaks on error.

JSON numbers and strings become simple Perl scalars. JSON arrays become Perl arrayrefs and JSON objects become Perl hashrefs. true becomes 1, false becomes 0 and null becomes undef.

($perl_scalar, $characters) = $json->decode_prefix ($json_text)

This works like the decode method, but instead of raising an exception when there is trailing garbage after the first JSON object, it will silently stop parsing there and return the number of characters consumed so far.

This is useful if your JSON texts are not delimited by an outer protocol (which is not the brightest thing to do in the first place) and you need to know where the JSON text ends.

   JSON::XS->new->decode_prefix ("[1] the tail")
   => ([], 3)


This section describes how JSON::XS maps Perl values to JSON values and vice versa. These mappings are designed to "do the right thing" in most circumstances automatically, preserving round-tripping characteristics (what you put in comes out as something equivalent).

For the more enlightened: note that in the following descriptions, lowercase perl refers to the Perl interpreter, while uppercase Perl refers to the abstract Perl language itself.



A JSON object becomes a reference to a hash in Perl. No ordering of object keys is preserved (JSON does not preserve object key ordering itself).


A JSON array becomes a reference to an array in Perl.


A JSON string becomes a string scalar in Perl - Unicode codepoints in JSON are represented by the same codepoints in the Perl string, so no manual decoding is necessary.


A JSON number becomes either an integer, numeric (floating point) or string scalar in perl, depending on its range and any fractional parts. On the Perl level, there is no difference between those as Perl handles all the conversion details, but an integer may take slightly less memory and might represent more values exactly than floating point numbers.

If the number consists of digits only, JSON::XS will try to represent it as an integer value. If that fails, it will try to represent it as a numeric (floating point) value if that is possible without loss of precision. Otherwise it will preserve the number as a string value (in which case you lose roundtripping ability, as the JSON number will be re-encoded toa JSON string).

Numbers containing a fractional or exponential part will always be represented as numeric (floating point) values, possibly at a loss of precision (in which case you might lose perfect roundtripping ability, but the JSON number will still be re-encoded as a JSON number).

true, false

These JSON atoms become JSON::XS::true and JSON::XS::false, respectively. They are overloaded to act almost exactly like the numbers 1 and 0. You can check whether a scalar is a JSON boolean by using the JSON::XS::is_bool function.


A JSON null atom becomes undef in Perl.


The mapping from Perl to JSON is slightly more difficult, as Perl is a truly typeless language, so we can only guess which JSON type is meant by a Perl value.

hash references

Perl hash references become JSON objects. As there is no inherent ordering in hash keys (or JSON objects), they will usually be encoded in a pseudo-random order that can change between runs of the same program but stays generally the same within a single run of a program. JSON::XS can optionally sort the hash keys (determined by the canonical flag), so the same datastructure will serialise to the same JSON text (given same settings and version of JSON::XS), but this incurs a runtime overhead and is only rarely useful, e.g. when you want to compare some JSON text against another for equality.

array references

Perl array references become JSON arrays.

other references

Other unblessed references are generally not allowed and will cause an exception to be thrown, except for references to the integers 0 and 1, which get turned into false and true atoms in JSON. You can also use JSON::XS::false and JSON::XS::true to improve readability.

   encode_json [\0,JSON::XS::true]      # yields [false,true]
JSON::XS::true, JSON::XS::false

These special values become JSON true and JSON false values, respectively. You can also use \1 and \0 directly if you want.

blessed objects

Blessed objects are not directly representable in JSON. See the allow_blessed and convert_blessed methods on various options on how to deal with this: basically, you can choose between throwing an exception, encoding the reference as if it weren't blessed, or provide your own serialiser method.

simple scalars

Simple Perl scalars (any scalar that is not a reference) are the most difficult objects to encode: JSON::XS will encode undefined scalars as JSON null values, scalars that have last been used in a string context before encoding as JSON strings, and anything else as number value:

   # dump as number
   encode_json [2]                      # yields [2]
   encode_json [-3.0e17]                # yields [-3e+17]
   my $value = 5; encode_json [$value]  # yields [5]

   # used as string, so dump as string
   print $value;
   encode_json [$value]                 # yields ["5"]

   # undef becomes null
   encode_json [undef]                  # yields [null]

You can force the type to be a JSON string by stringifying it:

   my $x = 3.1; # some variable containing a number
   "$x";        # stringified
   $x .= "";    # another, more awkward way to stringify
   print $x;    # perl does it for you, too, quite often

You can force the type to be a JSON number by numifying it:

   my $x = "3"; # some variable containing a string
   $x += 0;     # numify it, ensuring it will be dumped as a number
   $x *= 1;     # same thing, the choice is yours.

You can not currently force the type in other, less obscure, ways. Tell me if you need this capability (but don't forget to explain why its needed :).


The interested reader might have seen a number of flags that signify encodings or codesets - utf8, latin1 and ascii. There seems to be some confusion on what these do, so here is a short comparison:

utf8 controls wether the JSON text created by encode (and expected by decode) is UTF-8 encoded or not, while latin1 and ascii only control wether encode escapes character values outside their respective codeset range. Neither of these flags conflict with each other, although some combinations make less sense than others.

Care has been taken to make all flags symmetrical with respect to encode and decode, that is, texts encoded with any combination of these flag values will be correctly decoded when the same flags are used - in general, if you use different flag settings while encoding vs. when decoding you likely have a bug somewhere.

Below comes a verbose discussion of these flags. Note that a "codeset" is simply an abstract set of character-codepoint pairs, while an encoding takes those codepoint numbers and encodes them, in our case into octets. Unicode is (among other things) a codeset, UTF-8 is an encoding, and ISO-8859-1 (= latin 1) and ASCII are both codesets and encodings at the same time, which can be confusing.

utf8 flag disabled

When utf8 is disabled (the default), then encode/decode generate and expect Unicode strings, that is, characters with high ordinal Unicode values (> 255) will be encoded as such characters, and likewise such characters are decoded as-is, no canges to them will be done, except "(re-)interpreting" them as Unicode codepoints or Unicode characters, respectively (to Perl, these are the same thing in strings unless you do funny/weird/dumb stuff).

This is useful when you want to do the encoding yourself (e.g. when you want to have UTF-16 encoded JSON texts) or when some other layer does the encoding for you (for example, when printing to a terminal using a filehandle that transparently encodes to UTF-8 you certainly do NOT want to UTF-8 encode your data first and have Perl encode it another time).

utf8 flag enabled

If the utf8-flag is enabled, encode/decode will encode all characters using the corresponding UTF-8 multi-byte sequence, and will expect your input strings to be encoded as UTF-8, that is, no "character" of the input string must have any value > 255, as UTF-8 does not allow that.

The utf8 flag therefore switches between two modes: disabled means you will get a Unicode string in Perl, enabled means you get an UTF-8 encoded octet/binary string in Perl.

latin1 or ascii flags enabled

With latin1 (or ascii) enabled, encode will escape characters with ordinal values > 255 (> 127 with ascii) and encode the remaining characters as specified by the utf8 flag.

If utf8 is disabled, then the result is also correctly encoded in those character sets (as both are proper subsets of Unicode, meaning that a Unicode string with all character values < 256 is the same thing as a ISO-8859-1 string, and a Unicode string with all character values < 128 is the same thing as an ASCII string in Perl).

If utf8 is enabled, you still get a correct UTF-8-encoded string, regardless of these flags, just some more characters will be escaped using \uXXXX then before.

Note that ISO-8859-1-encoded strings are not compatible with UTF-8 encoding, while ASCII-encoded strings are. That is because the ISO-8859-1 encoding is NOT a subset of UTF-8 (despite the ISO-8859-1 codeset being a subset of Unicode), while ASCII is.

Surprisingly, decode will ignore these flags and so treat all input values as governed by the utf8 flag. If it is disabled, this allows you to decode ISO-8859-1- and ASCII-encoded strings, as both strict subsets of Unicode. If it is enabled, you can correctly decode UTF-8 encoded strings.

So neither latin1 nor ascii are incompatible with the utf8 flag - they only govern when the JSON output engine escapes a character or not.

The main use for latin1 is to relatively efficiently store binary data as JSON, at the expense of breaking compatibility with most JSON decoders.

The main use for ascii is to force the output to not contain characters with values > 127, which means you can interpret the resulting string as UTF-8, ISO-8859-1, ASCII, KOI8-R or most about any character set and 8-bit-encoding, and still get the same data structure back. This is useful when your channel for JSON transfer is not 8-bit clean or the encoding might be mangled in between (e.g. in mail), and works because ASCII is a proper subset of most 8-bit and multibyte encodings in use in the world.


As already mentioned, this module was created because none of the existing JSON modules could be made to work correctly. First I will describe the problems (or pleasures) I encountered with various existing JSON modules, followed by some benchmark values. JSON::XS was designed not to suffer from any of these problems or limitations.

JSON 2.xx

A marvellous piece of engineering, this module either uses JSON::XS directly when available (so will be 100% compatible with it, including speed), or it uses JSON::PP, which is basically JSON::XS translated to Pure Perl, which should be 100% compatible with JSON::XS, just a bit slower.

You cannot really lose by using this module, especially as it tries very hard to work even with ancient Perl versions, while JSON::XS does not.

JSON 1.07

Slow (but very portable, as it is written in pure Perl).

Undocumented/buggy Unicode handling (how JSON handles Unicode values is undocumented. One can get far by feeding it Unicode strings and doing en-/decoding oneself, but Unicode escapes are not working properly).

No round-tripping (strings get clobbered if they look like numbers, e.g. the string 2.0 will encode to 2.0 instead of "2.0", and that will decode into the number 2.

JSON::PC 0.01

Very fast.

Undocumented/buggy Unicode handling.

No round-tripping.

Has problems handling many Perl values (e.g. regex results and other magic values will make it croak).

Does not even generate valid JSON ({1,2} gets converted to {1:2} which is not a valid JSON text.

Unmaintained (maintainer unresponsive for many months, bugs are not getting fixed).

JSON::Syck 0.21

Very buggy (often crashes).

Very inflexible (no human-readable format supported, format pretty much undocumented. I need at least a format for easy reading by humans and a single-line compact format for use in a protocol, and preferably a way to generate ASCII-only JSON texts).

Completely broken (and confusingly documented) Unicode handling (Unicode escapes are not working properly, you need to set ImplicitUnicode to different values on en- and decoding to get symmetric behaviour).

No round-tripping (simple cases work, but this depends on whether the scalar value was used in a numeric context or not).

Dumping hashes may skip hash values depending on iterator state.

Unmaintained (maintainer unresponsive for many months, bugs are not getting fixed).

Does not check input for validity (i.e. will accept non-JSON input and return "something" instead of raising an exception. This is a security issue: imagine two banks transferring money between each other using JSON. One bank might parse a given non-JSON request and deduct money, while the other might reject the transaction with a syntax error. While a good protocol will at least recover, that is extra unnecessary work and the transaction will still not succeed).


Very fast. Very natural. Very nice.

Undocumented Unicode handling (but the best of the pack. Unicode escapes still don't get parsed properly).

Very inflexible.

No round-tripping.

Does not generate valid JSON texts (key strings are often unquoted, empty keys result in nothing being output)

Does not check input for validity.


You often hear that JSON is a subset of YAML. This is, however, a mass hysteria(*) and very far from the truth (as of the time of this writing), so let me state it clearly: in general, there is no way to configure JSON::XS to output a data structure as valid YAML that works in all cases.

If you really must use JSON::XS to generate YAML, you should use this algorithm (subject to change in future versions):

   my $to_yaml = JSON::XS->new->utf8->space_after (1);
   my $yaml = $to_yaml->encode ($ref) . "\n";

This will usually generate JSON texts that also parse as valid YAML. Please note that YAML has hardcoded limits on (simple) object key lengths that JSON doesn't have and also has different and incompatible unicode handling, so you should make sure that your hash keys are noticeably shorter than the 1024 "stream characters" YAML allows and that you do not have characters with codepoint values outside the Unicode BMP (basic multilingual page). YAML also does not allow \/ sequences in strings (which JSON::XS does not currently generate, but other JSON generators might).

There might be other incompatibilities that I am not aware of (or the YAML specification has been changed yet again - it does so quite often). In general you should not try to generate YAML with a JSON generator or vice versa, or try to parse JSON with a YAML parser or vice versa: chances are high that you will run into severe interoperability problems when you least expect it.


I have been pressured multiple times by Brian Ingerson (one of the authors of the YAML specification) to remove this paragraph, despite him acknowledging that the actual incompatibilities exist. As I was personally bitten by this "JSON is YAML" lie, I refused and said I will continue to educate people about these issues, so others do not run into the same problem again and again. After this, Brian called me a (quote)complete and worthless idiot(unquote).

In my opinion, instead of pressuring and insulting people who actually clarify issues with YAML and the wrong statements of some of its proponents, I would kindly suggest reading the JSON spec (which is not that difficult or long) and finally make YAML compatible to it, and educating users about the changes, instead of spreading lies about the real compatibility for many years and trying to silence people who point out that it isn't true.


It seems that JSON::XS is surprisingly fast, as shown in the following tables. They have been generated with the help of the eg/bench program in the JSON::XS distribution, to make it easy to compare on your own system.

First comes a comparison between various modules using a very short single-line JSON string (also available at http://dist.schmorp.de/misc/json/short.json).

   {"method": "handleMessage", "params": ["user1", "we were just talking"], \
   "id": null, "array":[1,11,234,-5,1e5,1e7, true,  false]}

It shows the number of encodes/decodes per second (JSON::XS uses the functional interface, while JSON::XS/2 uses the OO interface with pretty-printing and hashkey sorting enabled, JSON::XS/3 enables shrink). Higher is better:

   module     |     encode |     decode |
   JSON 1.x   |   4990.842 |   4088.813 |
   JSON::DWIW |  51653.990 |  71575.154 |
   JSON::PC   |  65948.176 |  74631.744 |
   JSON::PP   |   8931.652 |   3817.168 |
   JSON::Syck |  24877.248 |  27776.848 |
   JSON::XS   | 388361.481 | 227951.304 |
   JSON::XS/2 | 227951.304 | 218453.333 |
   JSON::XS/3 | 338250.323 | 218453.333 |
   Storable   |  16500.016 | 135300.129 |

That is, JSON::XS is about five times faster than JSON::DWIW on encoding, about three times faster on decoding, and over forty times faster than JSON, even with pretty-printing and key sorting. It also compares favourably to Storable for small amounts of data.

Using a longer test string (roughly 18KB, generated from Yahoo! Locals search API (http://dist.schmorp.de/misc/json/long.json).

   module     |     encode |     decode |
   JSON 1.x   |     55.260 |     34.971 |
   JSON::DWIW |    825.228 |   1082.513 |
   JSON::PC   |   3571.444 |   2394.829 |
   JSON::PP   |    210.987 |     32.574 |
   JSON::Syck |    552.551 |    787.544 |
   JSON::XS   |   5780.463 |   4854.519 |
   JSON::XS/2 |   3869.998 |   4798.975 |
   JSON::XS/3 |   5862.880 |   4798.975 |
   Storable   |   4445.002 |   5235.027 |

Again, JSON::XS leads by far (except for Storable which non-surprisingly decodes faster).

On large strings containing lots of high Unicode characters, some modules (such as JSON::PC) seem to decode faster than JSON::XS, but the result will be broken due to missing (or wrong) Unicode handling. Others refuse to decode or encode properly, so it was impossible to prepare a fair comparison table for that case.


When you are using JSON in a protocol, talking to untrusted potentially hostile creatures requires relatively few measures.

First of all, your JSON decoder should be secure, that is, should not have any buffer overflows. Obviously, this module should ensure that and I am trying hard on making that true, but you never know.

Second, you need to avoid resource-starving attacks. That means you should limit the size of JSON texts you accept, or make sure then when your resources run out, that's just fine (e.g. by using a separate process that can crash safely). The size of a JSON text in octets or characters is usually a good indication of the size of the resources required to decode it into a Perl structure. While JSON::XS can check the size of the JSON text, it might be too late when you already have it in memory, so you might want to check the size before you accept the string.

Third, JSON::XS recurses using the C stack when decoding objects and arrays. The C stack is a limited resource: for instance, on my amd64 machine with 8MB of stack size I can decode around 180k nested arrays but only 14k nested JSON objects (due to perl itself recursing deeply on croak to free the temporary). If that is exceeded, the program crashes. To be conservative, the default nesting limit is set to 512. If your process has a smaller stack, you should adjust this setting accordingly with the max_depth method.

Something else could bomb you, too, that I forgot to think of. In that case, you get to keep the pieces. I am always open for hints, though...

Also keep in mind that JSON::XS might leak contents of your Perl data structures in its error messages, so when you serialise sensitive information you might want to make sure that exceptions thrown by JSON::XS will not end up in front of untrusted eyes.

If you are using JSON::XS to return packets to consumption by JavaScript scripts in a browser you should have a look at http://jpsykes.com/47/practical-csrf-and-json-security to see whether you are vulnerable to some common attack vectors (which really are browser design bugs, but it is still you who will have to deal with it, as major browser developers care only for features, not about getting security right).


This module is not guaranteed to be thread safe and there are no plans to change this until Perl gets thread support (as opposed to the horribly slow so-called "threads" which are simply slow and bloated process simulations - use fork, its much faster, cheaper, better).

(It might actually work, but you have been warned).


While the goal of this module is to be correct, that unfortunately does not mean its bug-free, only that I think its design is bug-free. It is still relatively early in its development. If you keep reporting bugs they will be fixed swiftly, though.

Please refrain from using rt.cpan.org or any other bug reporting service. I put the contact address into my modules for a reason.


 Marc Lehmann <schmorp@schmorp.de>