An open source project by FPL.
Use in C++

Assuming you have written a schema using the above language in say mygame.fbs (FlatBuffer Schema, though the extension doesn't matter), you've generated a C++ header called mygame_generated.h using the compiler (e.g. flatc -c mygame.fbs), you can now start using this in your program by including the header. As noted, this header relies on flatbuffers/flatbuffers.h, which should be in your include path.

Writing in C++

To start creating a buffer, create an instance of FlatBufferBuilder which will contain the buffer as it grows:

FlatBufferBuilder fbb;

Before we serialize a Monster, we need to first serialize any objects that are contained there-in, i.e. we serialize the data tree using depth first, pre-order traversal. This is generally easy to do on any tree structures. For example:

auto name = fbb.CreateString("MyMonster");
unsigned char inv[] = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 };
auto inventory = fbb.CreateVector(inv, 10);

CreateString and CreateVector serialize these two built-in datatypes, and return offsets into the serialized data indicating where they are stored, such that Monster below can refer to them.

CreateString can also take an std::string, or a const char * with an explicit length, and is suitable for holding UTF-8 and binary data if needed.

CreateVector can also take an std::vector. The offset it returns is typed, i.e. can only be used to set fields of the correct type below. To create a vector of struct objects (which will be stored as contiguous memory in the buffer, use CreateVectorOfStructs instead.

To create a vector of nested objects (e.g. tables, strings or other vectors) collect their offsets in a temporary array/vector, then call CreateVector on that (see e.g. the array of strings example in test.cpp CreateFlatBufferTest).

Vec3 vec(1, 2, 3);

Vec3 is the first example of code from our generated header. Structs (unlike tables) translate to simple structs in C++, so we can construct them in a familiar way.

We have now serialized the non-scalar components of of the monster example, so we could create the monster something like this:

auto mloc = CreateMonster(fbb, &vec, 150, 80, name, inventory, Color_Red, 0, Any_NONE);

Note that we're passing 150 for the mana field, which happens to be the default value: this means the field will not actually be written to the buffer, since we'll get that value anyway when we query it. This is a nice space savings, since it is very common for fields to be at their default. It means we also don't need to be scared to add fields only used in a minority of cases, since they won't bloat up the buffer sizes if they're not actually used.

We do something similarly for the union field test by specifying a 0 offset and the NONE enum value (part of every union) to indicate we don't actually want to write this field. You can use 0 also as a default for other non-scalar types, such as strings, vectors and tables. To pass an actual table, pass a preconstructed table as mytable.Union() that corresponds to union enum you're passing.

Tables (like Monster) give you full flexibility on what fields you write (unlike Vec3, which always has all fields set because it is a struct). If you want even more control over this (i.e. skip fields even when they are not default), instead of the convenient CreateMonster call we can also build the object field-by-field manually:

MonsterBuilder mb(fbb);
auto mloc = mb.Finish();

We start with a temporary helper class MonsterBuilder (which is defined in our generated code also), then call the various add_ methods to set fields, and Finish to complete the object. This is pretty much the same code as you find inside CreateMonster, except we're leaving out a few fields. Fields may also be added in any order, though orderings with fields of the same size adjacent to each other most efficient in size, due to alignment. You should not nest these Builder classes (serialize your data in pre-order).

Regardless of whether you used CreateMonster or MonsterBuilder, you now have an offset to the root of your data, and you can finish the buffer using:

FinishMonsterBuffer(fbb, mloc);

The buffer is now ready to be stored somewhere, sent over the network, be compressed, or whatever you'd like to do with it. You can access the start of the buffer with fbb.GetBufferPointer(), and it's size from fbb.GetSize().

Calling code may take ownership of the buffer with fbb.ReleaseBufferPointer(). Should you do it, the FlatBufferBuilder will be in an invalid state, and must be cleared before it can be used again. However, it also means you are able to destroy the builder while keeping the buffer in your application.

samples/sample_binary.cpp is a complete code sample similar to the code above, that also includes the reading code below.

Reading in C++

If you've received a buffer from somewhere (disk, network, etc.) you can directly start traversing it using:

auto monster = GetMonster(buffer_pointer);

monster is of type Monster *, and points to somewhere inside your buffer (root object pointers are not the same as buffer_pointer !). If you look in your generated header, you'll see it has convenient accessors for all fields, e.g.

assert(monster->hp() == 80);
assert(monster->mana() == 150); // default
assert(strcmp(monster->name()->c_str(), "MyMonster") == 0);

These should all be true. Note that we never stored a mana value, so it will return the default.

To access sub-objects, in this case the Vec3:

auto pos = monster->pos();
assert(pos->z() == 3);

If we had not set the pos field during serialization, it would be NULL.

Similarly, we can access elements of the inventory array:

auto inv = monster->inventory();
assert(inv->Get(9) == 9);

Mutating FlatBuffers

As you saw above, typically once you have created a FlatBuffer, it is read-only from that moment on. There are however cases where you have just received a FlatBuffer, and you'd like to modify something about it before sending it on to another recipient. With the above functionality, you'd have to generate an entirely new FlatBuffer, while tracking what you modify in your own data structures. This is inconvenient.

For this reason FlatBuffers can also be mutated in-place. While this is great for making small fixes to an existing buffer, you generally want to create buffers from scratch whenever possible, since it is much more efficient and the API is much more general purpose.

To get non-const accessors, invoke flatc with --gen-mutable.

Similar to the reading API above, you now can:

auto monster = GetMutableMonster(buffer_pointer); // non-const
monster->mutate_hp(10); // Set table field.
monster->mutable_pos()->mutate_z(4); // Set struct field.
monster->mutable_inventory()->Mutate(0, 1); // Set vector element.

We use the somewhat verbose term mutate instead of set to indicate that this is a special use case, not to be confused with the default way of constructing FlatBuffer data.

After the above mutations, you can send on the FlatBuffer to a new recipient without any further work!

Note that any mutate_ functions on tables return a bool, which is false if the field we're trying to set isn't present in the buffer. Fields are not present if they weren't set, or even if they happen to be equal to the default value. For example, in the creation code above we set the mana field to 150, which is the default value, so it was never stored in the buffer. Trying to call mutate_mana() on such data will return false, and the value won't actually be modified!

One way to solve this is to call ForceDefaults() on a FlatBufferBuilder to force all fields you set to actually be written. This of course increases the size of the buffer somewhat, but this may be acceptable for a mutable buffer.

Alternatively, you can use the more powerful reflection functionality:

Reflection (& Resizing)

If the above ways of accessing a buffer are still too static for you, there is experimental support for reflection in FlatBuffers, allowing you to read and write data even if you don't know the exact format of a buffer, and even allows you to change sizes of strings and vectors in-place.

The way this works is very elegant, there is actually a FlatBuffer schema that describes schemas (!) which you can find in reflection/reflection.fbs. The compiler flatc can write out any schemas it has just parsed as a binary FlatBuffer, corresponding to this meta-schema.

Loading in one of these binary schemas at runtime allows you traverse any FlatBuffer data that corresponds to it without knowing the exact format. You can query what fields are present, and then read/write them after.

For convenient field manipulation, you can include the header flatbuffers/reflection.h which includes both the generated code from the meta schema, as well as a lot of helper functions.

And example of usage for the moment you can find in test.cpp/ReflectionTest().

Storing maps / dictionaries in a FlatBuffer

FlatBuffers doesn't support maps natively, but there is support to emulate their behavior with vectors and binary search, which means you can have fast lookups directly from a FlatBuffer without having to unpack your data into a std::map or similar.

To use it:

  • Designate one of the fields in a table as they "key" field. You do this by setting the key attribute on this field, e.g. name:string (key). You may only have one key field, and it must be of string or scalar type.
  • Write out tables of this type as usual, collect their offsets in an array or vector.
  • Instead of CreateVector, call CreateVectorOfSortedTables, which will first sort all offsets such that the tables they refer to are sorted by the key field, then serialize it.
  • Now when you're accessing the FlatBuffer, you can use Vector::LookupByKey instead of just Vector::Get to access elements of the vector, e.g.: myvector->LookupByKey("Fred"), which returns a pointer to the corresponding table type, or nullptr if not found. LookupByKey performs a binary search, so should have a similar speed to std::map, though may be faster because of better caching. LookupByKey only works if the vector has been sorted, it will likely not find elements if it hasn't been sorted.

Direct memory access

As you can see from the above examples, all elements in a buffer are accessed through generated accessors. This is because everything is stored in little endian format on all platforms (the accessor performs a swap operation on big endian machines), and also because the layout of things is generally not known to the user.

For structs, layout is deterministic and guaranteed to be the same accross platforms (scalars are aligned to their own size, and structs themselves to their largest member), and you are allowed to access this memory directly by using sizeof() and memcpy on the pointer to a struct, or even an array of structs.

To compute offsets to sub-elements of a struct, make sure they are a structs themselves, as then you can use the pointers to figure out the offset without having to hardcode it. This is handy for use of arrays of structs with calls like glVertexAttribPointer in OpenGL or similar APIs.

It is important to note is that structs are still little endian on all machines, so only use tricks like this if you can guarantee you're not shipping on a big endian machine (an assert(FLATBUFFERS_LITTLEENDIAN) would be wise).

Access of untrusted buffers

The generated accessor functions access fields over offsets, which is very quick. These offsets are not verified at run-time, so a malformed buffer could cause a program to crash by accessing random memory.

When you're processing large amounts of data from a source you know (e.g. your own generated data on disk), this is acceptable, but when reading data from the network that can potentially have been modified by an attacker, this is undesirable.

For this reason, you can optionally use a buffer verifier before you access the data. This verifier will check all offsets, all sizes of fields, and null termination of strings to ensure that when a buffer is accessed, all reads will end up inside the buffer.

Each root type will have a verification function generated for it, e.g. for Monster, you can call:

bool ok = VerifyMonsterBuffer(Verifier(buf, len));

if ok is true, the buffer is safe to read.

Besides untrusted data, this function may be useful to call in debug mode, as extra insurance against data being corrupted somewhere along the way.

While verifying a buffer isn't "free", it is typically faster than a full traversal (since any scalar data is not actually touched), and since it may cause the buffer to be brought into cache before reading, the actual overhead may be even lower than expected.

In specialized cases where a denial of service attack is possible, the verifier has two additional constructor arguments that allow you to limit the nesting depth and total amount of tables the verifier may encounter before declaring the buffer malformed. The default is Verifier(buf, len, 64 /* max depth */, 1000000, /* max tables */) which should be sufficient for most uses.

Text & schema parsing

Using binary buffers with the generated header provides a super low overhead use of FlatBuffer data. There are, however, times when you want to use text formats, for example because it interacts better with source control, or you want to give your users easy access to data.

Another reason might be that you already have a lot of data in JSON format, or a tool that generates JSON, and if you can write a schema for it, this will provide you an easy way to use that data directly.

(see the schema documentation for some specifics on the JSON format accepted).

There are two ways to use text formats:

Using the compiler as a conversion tool

This is the preferred path, as it doesn't require you to add any new code to your program, and is maximally efficient since you can ship with binary data. The disadvantage is that it is an extra step for your users/developers to perform, though you might be able to automate it.

flatc -b myschema.fbs mydata.json

This will generate the binary file mydata_wire.bin which can be loaded as before.

Making your program capable of loading text directly

This gives you maximum flexibility. You could even opt to support both, i.e. check for both files, and regenerate the binary from text when required, otherwise just load the binary.

This option is currently only available for C++, or Java through JNI.

As mentioned in the section "Building" above, this technique requires you to link a few more files into your program, and you'll want to include flatbuffers/idl.h.

Load text (either a schema or json) into an in-memory buffer (there is a convenient LoadFile() utility function in flatbuffers/util.h if you wish). Construct a parser:

flatbuffers::Parser parser;

Now you can parse any number of text files in sequence:


This works similarly to how the command-line compiler works: a sequence of files parsed by the same Parser object allow later files to reference definitions in earlier files. Typically this means you first load a schema file (which populates Parser with definitions), followed by one or more JSON files.

As optional argument to Parse, you may specify a null-terminated list of include paths. If not specified, any include statements try to resolve from the current directory.

If there were any parsing errors, Parse will return false, and Parser::err contains a human readable error string with a line number etc, which you should present to the creator of that file.

After each JSON file, the Parser::fbb member variable is the FlatBufferBuilder that contains the binary buffer version of that file, that you can access as described above.

samples/sample_text.cpp is a code sample showing the above operations.


Reading a FlatBuffer does not touch any memory outside the original buffer, and is entirely read-only (all const), so is safe to access from multiple threads even without synchronisation primitives.

Creating a FlatBuffer is not thread safe. All state related to building a FlatBuffer is contained in a FlatBufferBuilder instance, and no memory outside of it is touched. To make this thread safe, either do not share instances of FlatBufferBuilder between threads (recommended), or manually wrap it in synchronisation primites. There's no automatic way to accomplish this, by design, as we feel multithreaded construction of a single buffer will be rare, and synchronisation overhead would be costly.