Skip to content
Rapid YAML library
C++ Python CMake Shell Other
Branch: master
Clone or download
Latest commit c97a964 Aug 19, 2019

Rapid YAML

MIT Licensed Build Status Build status Coverage: coveralls Coverage: codecov Total alerts Language grade: C/C++

Or ryml, for short. ryml is a library to parse and emit YAML, and do it fast.

ryml parses both read-only and in-situ source buffers; the resulting data nodes hold only views to sub-ranges of the source buffer. No string copies or duplications are done, and no virtual functions are used. The data tree is a flat index-based structure stored in a single array. Serialization happens only at your request, after parsing / before emitting. Internally the data tree representation has no knowledge of types (but of course, every node can have a YAML type tag). It is easy and fast to read, write and iterate through the data tree.

ryml can use custom per-tree memory allocators, and is exception-agnostic. Errors are reported via a custom error handler callback (and a default implementation using std::abort is provided, but you can opt out).

ryml has respect for your compilation times and therefore it is NOT header-only. (Work is still needed on installing it, but you can just add it as a subdirectory of your CMake project).

ryml has no dependencies, not even on the STL (although it does use the libc). But it provides optional headers that let you serialize/deserialize STL strings and containers (or show you how to do it).

ryml is written in C++11, and is known to compile with:

  • Visual Studio 2015 and later
  • clang++ 3.9 and later
  • g++ 5 and later

ryml is extensively unit-tested in Linux and Windows. The tests include analysing ryml with:

  • valgrind
  • clang-tidy
  • clang sanitizers:
    • memory
    • address
    • undefined behavior
    • thread

ryml is also partially available in Python, with more languages to follow (see below).

Is it rapid?

You bet!

The first benchmarks results are extremely satisfying. On a i7-6800K CPU @ 3.40GHz:

  • ryml parses YAML at about ~150MB/s on Linux and ~100MB/s on Windows (vs2017).
  • ryml parses JSON at about ~450MB/s on Linux, faster than sajson (didn't try yet on Windows).
  • compared against the other existing YAML libraries for C/C++:
    • ryml is in general between 2 and 3 times faster than libyaml
    • ryml is in general between 20 and 70 times faster than yaml-cpp

(The results presented below are a bit scattered; and they need to be sistematized.) Here's the benchmark. Using different approaches within ryml (in-situ/read-only vs. with/without reuse), a YAML / JSON buffer is repeatedly parsed, and compared against other libraries.

Comparison with yaml-cpp

The first result set is for Windows, and is using a appveyor.yml config file. A comparison of these results is summarized on the table below:

Read rates (MB/s) ryml yamlcpp compared
appveyor / vs2017 / Release 101.5 5.3 20x / 5.2%
appveyor / vs2017 / Debug 6.4 0.0844 76x / 1.3%

The next set of results is taken in Linux, comparing g++ 8.2 and clang++ 7.0.1 in parsing a YAML buffer from a travis.yml config file or a JSON buffer from a compile_commands.json file. You can see the full results here. Summarizing:

Read rates (MB/s) ryml yamlcpp compared
json / clang++ / Release 453.5 15.1 30x / 3%
json / g++ / Release 430.5 16.3 26x / 4%
json / clang++ / Debug 61.9 1.63 38x / 3%
json / g++ / Debug 72.6 1.53 47x / 2%
travis / clang++ / Release 131.6 8.08 16x / 6%
travis / g++ / Release 176.4 8.23 21x / 5%
travis / clang++ / Debug 10.2 1.08 9x / 1%
travis / g++ / Debug 12.5 1.01 12x / 8%

The 450MB/s read rate for JSON puts ryml squarely in the same ballpark as RapidJSON and other fast json readers (data from here). Even parsing full YAML is at ~150MB/s, which is still in that performance ballpark, allbeit at its lower end. This is something to be proud of, as the YAML specification is much more complex than JSON.

Performance reading JSON

So how does ryml compare against other JSON readers? Well, it's one of the fastest!

The benchmark is the same as above, and it is reading the compile_commands.json, The _ro suffix notes parsing a read-only buffer (so buffer copies are performed), while the _rw suffix means that the source buffer can be parsed in situ. The _reuse means the data tree and/or parser are reused on each benchmark repeat.

Here's what we get with g++ 8.2:

|                  |           Release             |           Debug               
| Benchmark        |  Iterations    Bytes/sec      |  Iterations    Bytes/sec      
| rapidjson_ro     |        7941    509.855M/s     |         633    43.3632M/s     
| rapidjson_rw     |       21400    1.32937G/s     |        1067    68.171M/s     
| sajson_rw        |        6808    434.245M/s     |        2770    176.478M/s     
| sajson_ro        |        6726    430.723M/s     |        2748    175.613M/s     
| jsoncpp_ro       |        2871    183.616M/s     |        2941    187.937M/s     
| nlohmann_json_ro |        1807    115.801M/s     |         337    21.5237M/s     
| yamlcpp_ro       |         261    16.6322M/s     |          25    1.58178M/s     
| libyaml_ro       |        1786    113.909M/s     |         560    35.6599M/s     
| libyaml_ro_reuse |        1797    114.594M/s     |         561    35.8531M/s     
| ryml_ro          |        6088    388.585M/s     |         576    36.8634M/s     
| ryml_rw          |        6179    393.658M/s     |         577    36.8474M/s     
| ryml_ro_reuse    |        6986    446.248M/s     |        1164    74.636M/s      
| ryml_rw_reuse    |        7157    457.076M/s     |        1175    74.8721M/s     

You can verify that (at least for this test) ryml beats most json parsers at their own game, with the notable exception of rapidjson --- but this occurs only in Release mode. When in Debug mode, rapidjson is actually slower than ryml, and only sajson manages to be faster.

More json comparison benchmarks will be added, but seem unlikely to significantly alter these results.

Quick start

If you're wondering whether ryml's speed comes at a usage cost, you need not. With ryml, you can have your cake and eat it too: being rapid is definitely NOT the same as being unpractical! ryml was written with easy AND efficient usage in mind, and comes with a two level API for accessing and traversing the data tree.

The low-level interface is an index-based API available through the ryml::Tree class (see examples below). This class is essentially a contiguous array of NodeData elements; these are linked to parent, children and siblings via indices.

On top of this index-based API, there is a thin abstraction ryml::NodeRef which is essentially a non-owning pointer to a NodeData element. It provides convenient methods for accessing the NodeData properties wrapping it via a class allowing for a more object-oriented use.


A parser takes a source buffer and fills a ryml::Tree object:

#include <ryml.hpp>

// not needed by ryml, just for these examples (and below)
#include <iostream>
// convenience functions to print a node
void show_keyval(ryml::NodeRef n)
    std::cout << n.key() << ": " << n.val() << "\n";
void show_val(ryml::NodeRef n)
    std::cout << n.val() << "\n";
int main()
    // ryml can parse in situ (and read-only buffers too):
    char src[] = "{foo: 1, bar: [a: 2, b: 3]}";
    // there are also overloads for reusing the tree and parser
    ryml::Tree tree = ryml::parse(src);

    // get a reference to the "foo" node
    ryml::NodeRef node = tree["foo"];

    show_keyval(node);  // "foo: 1"
    show_val(node["bar"][0]);  // "2"
    show_val(node["bar"][1]);  // "3"

    // deserializing:
    int foo;
    node >> foo; // now foo == 1

It is also possible to parse read-only buffers, but note these will be copied over to an arena buffer in the tree object, and that buffer copy will be the one to be parsed:

// "{foo: 1}" is a read-only buffer; it will be
// copied to the tree's arena before parsing
ryml::Tree tree = ryml::parse("{foo: 1}");

When parsing, you can reuse the existing trees and parsers. You can also parse into particular tree nodes, so that you can parse an entire file into a node which is deep in the hierarchy of an existing tree. To see the various parse overloads, consult the c4/yml/parse.hpp header. The free-standing parse() functions (towards the end of the file) are just convenience wrappers for calling the several Parser::parse() overloads.

Browsing the tree

The data tree is an index-linked array of NodeData elements. These are defined roughly as (browse the c4/yml/tree.hpp header):

// (inside namespace c4::yml)

typedef enum : int // bitflags for marking node features
   ANCHOR=..., // etc 
} NodeType_e;
struct NodeType
   NodeType_e m_flags;
   // ... predicate methods such as
   // has_key(), is_map(), is_seq(), etc
struct NodeScalar // this is both for keys and vals
    csubstr tag;
    csubstr scalar;
    csubstr anchor;
    // csubstr is a constant substring:
    // a non-owning read-only string view
    // consisting of a pointer and a length
constexpr const size_t NONE = (size_t)-1;
struct NodeData
    NodeType   type;
    NodeScalar key; // data for the key (if applicable)
    NodeScalar val; // data for the value
    size_t     parent;      // NONE when this is the root node
    size_t     first_child; // NONE if this is a leaf node
    size_t     last_child;  // etc
    size_t     next_sibling;
    size_t     prev_sibling;

Please note that you should not rely on this particular structure; the above definitions are given only to provide an idea on how the tree is structured. To access and modify node properties, please use the APIs provided through the Tree (low-level) or the NodeRef (high-level) classes.

You may have noticed above the use of a csubstr class. This class is defined in another library, c4core, which is imported by ryml (so technically it's not a dependency, is it?). This is a library I use with my projects consisting of multiplatform low-level utilities. One of these is c4::csubstr (the name comes from "constant substring") which is a non-owning read-only string view, with many methods that make it practical to use (I would certainly argue more practical than std::string). (In fact, c4::csubstr and its writeable counterpart c4::substr are the workhorses of the ryml parsing and serialization code; you can browse these classes here: c4/substr.hpp.)

Now, let's parse and go through a tree. To obtain a NodeRef from the tree, you only need to invoke operator[]. This operator can take indices (when invoked on sequence and map nodes) and also strings (only when invoked on map nodes):

ryml::Tree tree = ryml::parse("[a, b, {c: 0, d: 1}]");

// note: show_val() was defined above

show_val(tree[0]); // "a"
show_val(tree[1]); // "b"
show_val(tree[2][ 0 ]); // "0" // index-based
show_val(tree[2][ 1 ]); // "1" // index-based
show_val(tree[2]["c"]); // "0" // string-based
show_val(tree[2]["d"]); // "1" // string-based

// note that trying to obtain the value on a non-value
// node such as a container will fail an assert:
// ERROR, assertion triggered: a container has no value
// ERROR: the same

// the same for keys:
show_keyval(tree[0]); // ERROR: sequence element has no key
show_keyval(tree[2][0]); // ok

Please note that since a ryml tree uses linear storage, the complexity of operator[] is linear on the number of children of the node on which it is invoked. If you use it with a large tree with many siblings at the root level, you may get a performance hit. To avoid this, you can create your own accelerator structure (eg, do a single loop at the root level to fill an std::map<csubstr,size_t> mapping key names to node indices; with a node index, a lookup is O(1), so this way you can get O(log n) lookup from a key.)

What about NodeRef? Let's consider when a non-existing key or index is requested via operator[]. Unlike with std::map, this operator does not modify the tree. Instead you get a seed-state NodeRef, and the tree will be modified only when this seed-state reference is written to. Thus NodeRef can either point to a valid tree node, or if no such node exists it will be in seed-state by holding the index or name passed to operator[]. To allow for this, NodeRef is a simple structure with a declaration like:

class NodeRef
    // a pointer to the tree
    Tree * m_tree; 
    // either the (tree-scoped) index of an existing node or the (node-scoped) index of a seed state
    size_t m_node_or_seed_id;
    // the key name of a seed state. null when valid
    const char* m_seed_name;


    // this can be used to query whether a node is in seed state
    bool valid()
        return m_node_or_seed_id != NONE
               m_seed_name == nullptr;

    // forward all calls to m_tree. For example:
    csubstr val() const { assert(valid()); return m_tree->val(m_node_or_seed_id); }
    void set_val(csubstr v) { if(!valid()) {/*create node in tree*/;} m_tree->set_val(m_node_or_seed_id, v); }

    // etc...

To iterate over children:

for(NodeRef c : node.children())
    std::cout << c.key() << "---" << c.val() << "\n";

To iterate over siblings:

for(NodeRef c : node.siblings())
    std::cout << c.key() << "---" << c.val() << "\n";

Creating a tree

To create a tree programatically:

ryml::Tree tree;
NodeRef r = tree.rootref();

// Each container node must be explicitly set (either MAP or SEQ):
r |= ryml::MAP;

r["foo"] = "1"; // ryml works only with strings.
// Note that the tree will be __pointing__ at the
// strings "foo" and "1" used here. You need
// to make sure they have at least the same
// lifetime as the tree.

// does not change the tree until s is written to.
ryml::NodeRef s = r["seq"]; // here, s is not valid()
s |= ryml::SEQ; // now s is valid()

s.append_child() = "bar0"; // this child is now __pointing__ at "bar0"
s.append_child() = "bar1";
s.append_child() = "bar2";

// emit to stdout (can also emit to FILE* or ryml::span)
emit(tree); // prints the following:
            // foo: 1
            // seq:
            //  - bar0
            //  - bar1
            //  - bar2

// serializing: using operator<< instead of operator=
// will make the tree serialize the value into a char
// arena inside the tree. This arena can be reserved at will.
int ch3 = 33, ch4 = 44;
s.append_child() << ch3;
s.append_child() << ch4;

    std::string tmp = "child5";
    s.append_child() << tmp;
    // now tmp can go safely out of scope, as it was
    // serialized to the tree's internal string arena

emit(tree); // now prints the following:
            // foo: 1
            // seq:
            //  - bar0
            //  - bar1
            //  - bar2
            //  - 33
            //  - 44
            //  - child5

// to serialize keys:
r.append_child() << ryml::key(66) << 7;

emit(tree); // now prints the following:
            // foo: 1
            // seq:
            //  - bar0
            //  - bar1
            //  - bar2
            //  - 33
            //  - 44
            //  - child5
            // 66: 7

Low-level API

The low-level api is an index-based API accessible from the ryml::Tree object. Here are some examples:

void print_keyval(Tree const& t, size_t elm_id)
    std::cout << t.get_key(elm_id)
              << ": "
              << t.get_val(elm_id) << "\n";

ryml::Tree t = parse("{foo: 1, bar: 2, baz: 3}")

size_t root_id = t.root_id();
size_t foo_id  = t.first_child(root_id);
size_t bar_id  = t.next_sibling(foo_id);
size_t baz_id  = t.last_child(root_id);

assert(baz == t.next_sibling(bar_id));
assert(bar == t.prev_sibling(baz_id));

print_keyval(t, foo_id); // "foo: 1"
print_keyval(t, bar_id); // "bar: 2"
print_keyval(t, baz_id); // "baz: 3"

// to iterate over the children of a node:
for(size_t i  = t.first_child(root_id);
           i != ryml::NONE;
           i  = t.next_sibling(i))
    // ...

// to iterate over the siblings of a node:
for(size_t i  = t.first_sibling(foo_id);
           i != ryml::NONE;
           i  = t.next_sibling(i))
    // ...

Custom types

ryml provides code to serialize the basic intrinsic types (integers, floating points and strings): you can see it in the the c4/to_chars.hpp header. For types other than these, you need to instruct ryml how to serialize your type.

There are two distinct type categories when serializing to a YAML tree:

  • Container types requiring child nodes (ie, either sequences or maps). For these, overload the write()/read() functions. For example,

    namespace foo {
    struct MyStruct; // a container-type struct 
        int subject;
        std::map<std::string, int> counts;
    // ... will need these functions to convert to YAML:
    void write(c4::yml::NodeRef *n, MyStruct const& v);
    void  read(c4::yml::NodeRef const& n, MyStruct *v);
    } // namespace foo

    which could be implemented as:

    #include <c4/yml/std/map.hpp>
    #include <c4/yml/std/string.hpp>
    void foo::read(c4::yml::NodeRef const& n, MyStruct *v)
        n["subject"] >> v->subject;
        n["counts"] >> v->counts;
    void foo::write(c4::yml::NodeRef *n, MyStruct const& v)
        *n |= c4::yml::MAP;
        NodeRef ch = n->append_child();
        ch = n->append_child();
        write(&ch, v.counts);

    To harness C++'s ADL rules, it is important to overload these functions in the namespace where the type you're serializing was defined (or in the c4::yml namespace). Generic examples can be seen in the (optional) implementations of std::vector or std::map, at their respective headers c4/yml/std/vector.hpp and c4/yml/std/map.hpp.

  • The second category is for types which should serialize to a string, resulting in leaf node in the YAML tree. For these, overload the to_chars(c4::substr, T)/from_chars(c4::csubstr, *T) functions. Here's an example for a 3D vector type:

    struct vec3 { float x, y, z; };
    // format v to the given string view + return the number of
    // characters written into it. The view size (buf.len) must
    // be strictly respected. Return the number of characters
    // that need to be written. So if the return value
    // is larger than buf.len, ryml will resize the buffer and
    // call this again with a larger buffer.
    size_t to_chars(c4::substr buf, vec3 v)
        // this call to c4::format() is a type-safe version
        // of snprintf(buf.str, buf.len, "(%f,%f,%f)", v.x, v.y, v.z)
        return c4::format(buf, "({},{},{})", v.x, v.y, v.z);
    bool from_chars(c4::csubstr buf, vec3 *v)
        // equivalent to sscanf(buf.str, "(%f,%f,%f)", &v.x, &v.y, &v.z)
        // --- actually snscanf(buf.str, buf.len, ...) but there's
        // no such function in the standard.
        size_t ret = c4::unformat(buf, "({},{},{})", v.x, v.y, v.z);
        return ret != c4::csubstr::npos;

    You can also look at the std::string serialization code.

STL interoperation

ryml does not use the STL internally, but you can use ryml to serialize and deserialize STL containers. That is, the use of STL is opt-in: you need to #include the proper ryml header for the container you want to serialize, or provide an implementation of your own, as above. Having done that, you can serialize / deserialize your containers with a single step. For example:

#include <ryml_std.hpp>
int main()
    std::map<std::string, int> m({{"foo", 1}, {"bar", 2}});
    ryml::Tree t;
    t.rootref() << m; // serialization of the map happens here
    // foo: 1
    // bar: 2
    t["foo"] << 1111;  // serialize an integer into
                       // the tree's arena, and make
                       // foo's value point at it
    t["bar"] << 2222;  // the same, but for bar
    // foo: 1111
    // bar: 2222
    t.rootref() >> m; // deserialization of the map happens here

    assert(m["foo"] == 1111); // ok
    assert(m["bar"] == 2222); // ok

The <ryml_std.hpp> header includes every std type implementation available in ryml. But you can include just a specific header if you are interested only in a particular container; these headers are located under a specific directory in the ryml source folder: c4/yml/std. You can browse them to learn how to implement your custom type: for containers, see for example the std::vector implementation, or the std::map implementation; for an example of value nodes, see the std::string implementation. If you'd like to see a particular STL container implemented, feel free to submit a pull request or open an issue.

Custom allocators and error handlers

ryml accepts your own allocators and error handlers. Read through this header file to set it up.

Please note the following about the use of custom allocators with ryml. If you use static ryml trees or parsers, you need to make sure that their allocator has the same lifetime. So you can't use ryml's default allocator, as it is declared in a ryml file, and the standard provides no guarantee on the relative initialization order, such that the allocator is constructed before and destroyed after your variables (in fact you are pretty much guaranteed to see this fail). So please carefully consider your choices, and ponder whether you really need to use ryml static trees and parsers. If you do need this, then you will need to declare and use an allocator from a ryml memory resource that outlives the tree and/or parser.

Other languages

One of the aims of ryml is to provide an efficient YAML API for other languages. There's already a cursory implementation for Python (using only the low-level API). After ironing out the general approach, other languages are likely to follow: probably (in order) JavaScript, C#, Java, Ruby, PHP, Octave and R (all of this is possible because we're using SWIG, which makes it easy to do so).


(Note that this is a work in progress. Additions will be made and things will be changed.) With that said, here's an example of the Python API:

import ryml

# because ryml does not take ownership of the source buffer
# ryml cannot accept strings; only bytes or bytearrays
src = b"{HELLO: a, foo: b, bar: c, baz: d, seq: [0, 1, 2, 3]}"

def check(tree):
    # for now, only the index-based low-level API is implemented
    assert tree.size() == 10
    assert tree.root_id() == 0
    assert tree.first_child(0) == 1
    assert tree.next_sibling(1) == 2
    assert tree.first_sibling(5) == 2
    assert tree.last_sibling(1) == 5
    # use bytes objects for queries
    assert tree.find_child(0, b"foo") == 1
    assert tree.key(1) == b"foo")
    assert tree.val(1) == b"b")
    assert tree.find_child(0, b"seq") == 5
    assert tree.is_seq(5)
    # to loop over children:
    for i, ch in enumerate(ryml.children(tree, 5)):
        assert tree.val(ch) == [b"0", b"1", b"2", b"3"][i]
    # to loop over siblings:
    for i, sib in enumerate(ryml.siblings(tree, 5)):
        assert tree.key(sib) == [b"HELLO", b"foo", b"bar", b"baz", b"seq"][i]
    # to walk over all elements
    visited = [False] * tree.size()
    for n, indentation_level in ryml.walk(tree):
        # just a dumb emitter
        left = "  " * indentation_level
        if tree.is_keyval(n):
           print("{}{}: {}".format(left, tree.key(n), tree.val(n))
        elif tree.is_val(n):
           print("- {}".format(left, tree.val(n))
        elif tree.is_keyseq(n):
           print("{}{}:".format(left, tree.key(n))
        visited[inode] = True
    assert False not in visited
    # NOTE about encoding!
    k = tree.get_key(5)
    print(k)  # '<memory at 0x7f80d5b93f48>'
    assert k == b"seq"               # ok, as expected
    assert k != "seq"                # not ok - NOTE THIS! 
    assert str(k) != "seq"           # not ok
    assert str(k, "utf8") == "seq"   # ok again

# parse immutable buffer
tree = ryml.parse(src)
check(tree) # OK

# also works, but requires bytearrays or
# objects offering writeable memory
mutable = bytearray(src)
tree = ryml.parse_in_situ(mutable)
check(tree) # OK

As expected, the performance results so far are encouraging. In a timeit benchmark compared against PyYaml and ruamel.yaml, ryml parses quicker by a factor of 30x-50x:

| case                  | iters | time(ms) | avg(ms) | avg_read(MB/s) |
| parse:RuamelYaml      |    88 | 800.483  |  9.096  |      0.234     |
| parse:PyYaml          |    88 | 541.370  |  6.152  |      0.346     |
| parse:RymlRo          |  3888 | 776.020  |  0.200  |     10.667     |
| parse:RymlRoReuse     |  1888 | 381.558  |  0.202  |     10.535     |
| parse:RymlRw          |  3888 | 775.121  |  0.199  |     10.679     |
| parse:RymlRwReuse     |  3888 | 774.534  |  0.199  |     10.687     |

(Note that the results above are somewhat biased towards ryml, because it does not perform any type conversions: return types are merely memoryviews to the source buffer.)

YAML standard conformance

ryml is under active development, but is close to feature complete. (With the notable exception of UTF8 support, which we haven't had the chance to verify.)

The following core features are tested:

  • mappings
  • sequences
  • complex keys
  • literal blocks
  • quoted scalars
  • tags
  • anchors and references

Of course, there are many dark corners in YAML, and there certainly can appear some cases which ryml fails to parse. So we welcome your bug reports or pull requests!.

Integration of the ~300 cases in the YAML test suite is ongoing work.

Alternative libraries

Why this library? Because none of the existing libraries was quite what I wanted. There are two C/C++ libraries that I know of:

The standard libyaml is a bare C library. It does not create a representation of the data tree, so it can't qualify as practical. My initial idea was to wrap parsing and emitting around libyaml, but to my surprise I found out it makes heavy use of allocations and string duplications when parsing. I briefly pondered on sending PRs to reduce these allocation needs, but not having a permanent tree to store the parsed data was too much of a downside.

yaml-cpp is full of functionality, but is heavy on the use of node-pointer-based structures like std::map, allocations, string copies and slow C++ stream serializations. This is generally a sure way of making your code slower, and strong evidence of this can be seen in the benchmark results above.

When performance and low latency are important, using contiguous structures for better cache behavior and to prevent the library from trampling over the client's caches, parsing in place and using non-owning strings is of central importance. Hence this Rapid YAML library which, with minimal compromise, bridges the gap from efficiency to usability. This library takes inspiration from RapidJSON and RapidXML.


ryml is permissively licensed under the MIT license.

You can’t perform that action at this time.