sasa1977 / gpb

A Google Protobuf implementation for Erlang

The gpb is a compiler for Google protocol buffer definitions files for Erlang.

Shortcuts: API documentation ~ gpb on hex.pm

Basic example of using gpb

Let's say we have a protobuf file, x.proto

message Person {
  required string name = 1;
  required int32 id = 2;
  optional string email = 3;
}

We can generate code for this definition in a number of different ways. Here we use the command line tool. For info on integration with rebar, see further down.

# .../gpb/bin/protoc-erl -I. x.proto

Now we've got x.erl and x.hrl. First we compile it and then we can try it out in the Erlang shell:

# erlc -I.../gpb/include x.erl
# erl
Erlang/OTP 19 [erts-8.0.3] [source] [64-bit] [smp:12:12] [async-threads:10] [kernel-poll:false]

Eshell V8.0.3  (abort with ^G)
1> rr("x.hrl").
['Person']
2> x:encode_msg(#'Person'{name="abc def", id=345, email="[e-mail]"}).    
<<10,7,97,98,99,32,100,101,102,16,217,2,26,13,97,64,101,
  120,97,109,112,108,101,46,99,111,109>>
3> Bin = v(-1).
<<10,7,97,98,99,32,100,101,102,16,217,2,26,13,97,64,101,
  120,97,109,112,108,101,46,99,111,109>>
4> x:decode_msg(Bin, 'Person').
#'Person'{name = "abc def",id = 345,email = "[e-mail]"}

In the Erlang shell, the rr("x.hrl") reads record definitions, and the v(-1) references a value one step earlier in the history.

Mapping of protocol buffer datatypes to erlang

Protobuf type	Erlang type
double, float	float() | infinity | '-infinity' | nan When encoding, integers, too, are accepted
int32, int64 uint32, uint64 sint32, sint64 fixed32, fixed64 sfixed32, sfixed64	integer()
bool	true | false When encoding, the integers 1 and 0, too, are accepted
enum	atom() unknown enums decode to integer()
message	record (thus tuple()) or map() if the maps (-maps) option is specified
string	unicode string, thus list of integers or binary() if the strings_as_binaries (-strbin) option is specified When encoding, iolists, too, are accepted
bytes	binary() When encoding, iolists, too, are accepted
oneof	{ChosenFieldName, Value} or ChosenFieldName => Value if the {maps_oneof,flat} (-maps_oneof flat) option is specified
map<_,_>	An unordered list of 2-tuples, [{Key,Value}] or a map(), if the maps (-maps) option is specified

Repeated fields are represented as lists.

Optional fields are represented as either the value or undefined if not set. However, for maps, if the option maps_unset_optional is set to omitted, then unset optional values are omitted from the map, instead of being set to undefined when encoding messages. When decoding messages, even with maps_unset_optional set to omitted, the default value will be set in the decoded map.

Examples of Erlang format for protocol buffer messages

Repeated and required fields

   message m1 {
     repeated uint32 i   = 1;
     required bool   b   = 2;
     required eee    e   = 3;
     required submsg sub = 4;
   }
   message submsg {
     required string s = 1;
     required bytes  b = 2;
   }
   enum eee {
     INACTIVE = 0;
     ACTIVE   = 1;
   }

Corresponding Erlang

   #m1{i   = [17, 4711],
       b   = true,
       e   = 'ACTIVE',
       sub = #submsg{s = "abc",
                     b = <<0,1,2,3,255>>}}

   %% If compiled to with the option maps:
   #{i   => [17, 4711],
     b   => true,
     e   => 'ACTIVE',
     sub => #{s => "abc",
              b => <<0,1,2,3,255>>}}

Optional fields

   message m2 {
     optional uint32 i1 = 1;
     optional uint32 i2 = 2;
   }

Corresponding Erlang

   #m2{i1 = 17}    % i2 is implicitly set to undefined

   %% With the maps option
   #{i1 => 17}

   %% With the maps option and the maps_unset_optional set to present_undefined:
   #{i1 => 17,
     i2 => undefined}

Oneof fields

This construct first appeared in Google protobuf version 2.6.0.

   message m3 {
     oneof u {
       int32  a = 1;
       string b = 2;
     }
   }

Corresponding Erlang

A oneof field is automatically always optional.

   #m3{u = {a, 17}}
   #m3{u = {b, "hello"}}
   #m3{}                 % u is implicitly set to undefined

   %% With the maps option
   #{u => {a, 17}}
   #{u => {b, "hello"}}
   #{}                   % If maps_unset_optional = omitted (default)
   #{u => undefined}     % With maps_unset_optional set to present_undefined

   %% With the {maps_oneof,flat} option (requires maps_unset_optional = omitted)
   #{a => 17}
   #{b => "hello"}
   #{}

Map fields

Not to be confused with Erlang maps. This construct first appeared in Google protobuf version 3.0.0 (for both the proto2 and the proto3 syntax)

   message m4 {
     map<uint32,string> f = 1;
   }

Corresponding Erlang

For records, the order of items is undefined when decoding.

   #m4{f = []}
   #m4{f = [{1, "a"}, {2, "b"}, {13, "hello"}]}

   %% With the maps option
   #{f => #{}}
   #{f => #{1 => "a", 2 => "b", 13 => "hello"}}

Unset optionals and the default option

For proto2 syntax

This describes how decoding works for optional fields that are not present in the binary-to-decode.

The documentation for Google protobuf says these decode to the default value if specified, or else to the field's type-specific default. The code generated by Google's protobuf compiler also contains has_<field>() methods so one can examine whether a field was actually present or not.

However, in Erlang, the natural way to set and read fields is to just use the syntax for records (or maps), and this leaves no good way to at the same time both convey whether a field was present or not and to read the defaults.

So the approach in gpb is that you have to choose: either or. Normally, it is possible to see whether an optional field is present or not, eg by checking if the value is undefined. But there are options to the compiler to instead decode to defaults, in which case you lose the ability to see whether a field is present or not. The options are defaults_for_omitted_optionals and type_defaults_for_omitted_optionals, for decoding to default=<x> values, or to type-specific defaults respectively.

It works this way:

message o1 {
  optional uint32 a = 1 [default=33];
  optional uint32 b = 2; // the type-specific default is 0
}

Given binary data <<>>, that is, neither field a nor b is present, then the call decode_msg(Input, o1) results in:

#o1{a=undefined, b=undefined} % None of the options

#o1{a=33, b=undefined}        % with option defaults_for_omitted_optionals

#o1{a=33, b=0}                % with both defaults_for_omitted_optionals
                              %       and type_defaults_for_omitted_optionals

#o1{a=0, b=0}                 % with only type_defaults_for_omitted_optionals

The last of the alternatives is perhaps not very useful, but still possible, and implemented for completeness.

Google's Reference

For proto3 syntax

For proto3, there is neither required nor default=<x> for fields. Instead, unless marked with optional, all scalar fields, strings and bytes are implicitly optional. On decoding, if such a field is missing in the binary to decode, they always decode to the type-specific default value. On encoding, such fields are only included in the resulting encoded binary if they have a value different from the type-specific default value. Even though all fields are implicitly optional, one could also say that on a conceptual level, all such fields always have a value. At decoding, it is not possible to determine whether at encoding, a value was present---with a type-specific value---or not.

Fields marked as optional are essentially represented the same way as in proto2 syntax; in a record the field has the value undefined if it is not set, and in maps the field is not present if it is not set.

A recommendation I've seen for if you need detection of "missing" data, is to define has_<field> boolean fields and set them appropriately. Another alternative could be to use the well-known wrapper messages.

Fields that are sub-messages and oneof fields, do not have any type-specific default. A sub-message field that was not set encodes differently from a sub-message field set to the sub-message, and it decodes differently. This holds even when the sub-message has no fields. It works a bit similarly for oneof fields. Either none of the alternative oneof fields is set, or one of them is. The encoded format is different, and on decoding it is possible to tell a difference.

Features of gpb

• Parses protocol buffer definition files and can generate:

  • record definitions, one record for each message
  • erlang code for encoding/decoding the messages to/from binaries

• Features of the protocol buffer definition files: gpb supports:

  • message definitions (also messages in messages)
  • scalar types
  • importing other proto files
  • nested types
  • message extensions
  • the packed and default options for fields
  • the allow_alias enum option (treated as if it is always set true)
  • generating metadata information
  • package namespacing (optional)
  • oneof (introduced in protobuf 2.6.0)
  • map<_,_> (introduced in protobuf 3.0.0)
  • proto3 support:
    • syntax and general semantics
    • import of well-known types
    • Callback functions can be specified for automatically translating google.protobuf.Any messages
  • groups
  • JSON mapping is supported, see the json (-json) option(s)

  gpb reads but ignores:

  • options other than packed or default
  • custom options

  gpb does not support:

  • aggregate custom options introduced in protobuf 2.4.0
  • rpc
  • JSON limitations:
    • does not handle the special JSON mapping of the google.protobuf.Any wellknown

• Characteristics of gpb:

  • Skipping over unknown message fields or groups, when decoding, is supported
  • Merging of messages, also recursive merging, is supported
  • Gpb can optionally generate code for verification of values during encoding this makes it easy to catch e.g integers out of range, or values of the wrong type.
  • Gpb can optionally or conditionally copy the contents of bytes fields, in order to let the runtime system free the larger message binary.
  • Gpb can optionally make use of the package attribute by prepending the name of the package to every contained message type (if defined), which is useful to avoid name clashes of message types across packages. See the use_packages option or the -pkgs command line option.
  • The generated encode/decoder has no run-time dependency to gpb, but there is normally a compile-time dependency for the generated code: to the #field{} record in gpb.hrl for the get_msg_defs function, but it is possible to avoid this dependency by using the also the defs_as_proplists or -pldefs option.
  • Gpb can generate code both to files and to binaries.
  • Proto input files are expected to be UTF-8, but the file reader will fall back to decode the files as latin1 in UTF-8 decode errors, for backwards compatibility and behaviour that most closely emulates what Google protobuf does.

• Introspection

  gpb generates some functions for examining messages, enums and services:

  • get_msg_defs() (or get_proto_defs() if introspect_get_proto_defs is set), get_msg_names(), get_enum_names()
  • find_msg_def(MsgName) and fetch_msg_def(MsgName)
  • find_enum_def(MsgName) and fetch_enum_def(MsgName)
  • enum_symbol_by_value(EnumName, Value),
  • enum_symbol_by_value_<EnumName>(Value), enum_value_by_symbol(EnumName, Enum) and enum_value_by_symbol_<EnumName>(Enum)
  • get_service_names(), get_service_def(ServiceName), get_rpc_names(ServiceName)
  • find_rpc_def(ServiceName, RpcName), fetch_rpc_def(ServiceName, RpcName)

  There are also some functions for translating between fully qualified names and internal names. These take any renaming options into consideration. They may be useful for instance with grpc reflection.

  • fqbin_to_service_name(<<"Package.ServiceName">>) and service_name_to_fqbin('ServiceName')
  • fqbins_to_service_and_rpc_name(<<"Package.ServiceName">>, <<"RpcName">>) and service_and_rpc_name_to_fqbins('ServiceName', 'RpcName')
  • fqbin_to_msg_name(<<"Package.MsgName">>) and msg_name_to_fqbin('MsgName')
  • fqbin_to_enum_name(<<"Package.EnumName">>) and enum_name_to_fqbin('EnumName')

  There are also some functions for querying what proto a type belongs to. Each type belongs to some "name" which is a string, usually the file name, sans extension, for example "name" if the proto file was "name.proto".

  • get_all_proto_names() -> ["name1", ...]
  • get_msg_containment("name") -> ['MsgName1', ...]
  • get_pkg_containment("name") -> 'Package'
  • get_service_containment("name") -> ['Service1', ...]
  • get_rpc_containment("name") -> [{'Service1', 'RpcName1}, ...]
  • get_proto_by_msg_name_as_fqbin(<<"Package.MsgName">>) -> "name"
  • get_proto_by_enum_name_as_fqbin(<<"Package.EnumName">>) -> "name"
  • get_protos_by_pkg_name_as_fqbin(<<"Package">>) -> ["name1", ...]

  There are also some version information functions:

  • gpb:version_as_string(), gpb:version_as_list() and gpb:version_source()
  • GeneratedCode:version_as_string(), GeneratedCode:version_as_list() and
  • GeneratedCode:version_source()
  • ?gpb_version (in gpb_version.hrl)
  • ?'GeneratedCode_gpb_version' (in GeneratedCode.hrl)

  The gpb can also generate a self-description of the proto file. The self-description is a description of the proto file, encoded to a binary using the descriptor.proto that comes with the Google protocol buffers library. Note that such an encoded self-descriptions won't be byte-by-byte identical to what the Google protocol buffers compiler will generate for the same proto, but should be roughly equivalent.

• Erroneously encoded protobuf messages and fields will generally cause the decoder to crash. Examples of such erroneous encodings are:

  • varints with too many bits
  • strings, bytes, sub messages or packed repeated fields, where the encoded length is longer than the remaining binary

• Maps

  Gpb can generate encoders/decoders for maps.

  The option maps_unset_optional can be used to specify behavior for non-present optional fields: whether they are omitted from maps, or whether they are present, but have the value undefined like for records.

• Reporting of errors in .proto files

  Gpb is not very good at error reporting, especially referencing errors, such as references to messages that are not defined. You might want to first verify with protoc that the .proto files are valid before feeding them to gpb.

Interaction with rebar

For info on how to use gpb with rebar3, see https://rebar3.org/docs/configuration/plugins/#protocol-buffers

Compatibility with rebar2

In rebar there is support for gpb since version 2.6.0. See the proto compiler section of rebar.sample.config file at https://github.com/rebar/rebar/blob/master/rebar.config.sample

For older versions of rebar---prior to 2.6.0---the text below outlines how to proceed:

Place the .proto files for instance in a proto/ subdirectory. Any subdirectory, other than src/, is fine, since rebar will try to use another protobuf compiler for any .proto it finds in the src/ subdirectory. Here are some some lines for the rebar.config file:

%% -*- erlang -*-
{pre_hooks,
 [{compile, "mkdir -p include"}, %% ensure the include dir exists
  {compile,
   "/path/to/gpb/bin/protoc-erl -I`pwd`/proto"
   "-o-erl src -o-hrl include `pwd`/proto/*.proto"
  }]}.

{post_hooks,
 [{clean,
   "bash -c 'for f in proto/*.proto; "
   "do "
   "  rm -f src/$(basename $f .proto).erl; "
   "  rm -f include/$(basename $f .proto).hrl; "
   "done'"}
 ]}.

{erl_opts, [{i, "/path/to/gpb/include"}]}.

Performance

Here is a comparison between gpb (interpreted by the erlang vm) and the C++, Python and Java serializers/deserializers of protobuf-2.6.1rc1

[MB/s]        | gpb   |pb/c++ |pb/c++ | pb/c++ | pb/py |pb/java| pb/java|
              |       |(speed)|(size) | (lite) |       |(size) | (speed)|
--------------+-------+-------+-------+--------+-------+-------+--------+
small msgs    |       |       |       |        |       |       |        |
  serialize   |   52  | 1240  |   85  |   750  |  6.5  |   68  |  1290  |
  deserialize |   63  |  880  |   85  |   950  |  5.5  |   90  |   450  |
--------------+-------+-------+-------+--------+-------+-------+--------+
large msgs    |       |       |       |        |       |       |        |
  serialize   |   36  |  950  |   72  |   670  |  4.5  |   55  |   670  |
  deserialize |   54  |  620  |   71  |   480  |  4.0  |   60  |   360  |
--------------+-------+-------+-------+--------+-------+-------+--------+

The performances are measured as number of processed MB/s, serialized form. Higher values means better performance.

The benchmarks are run with small and large messages (228 and 84584 bytes, respectively, in serialized form)

The Java benchmark is run with optimization both for code size and for speed. The Python implementation cannot optimize for speed.

SW: Python 2.7.11, Java 1.8.0_77 (Oracle JDK), Erlang/OTP 18.3, g++ 5.3.1
    Linux kernel 4.4, Debian (in 64 bit mode), protobuf-2.6.1rc1
HW: Intel Core i7 5820k, 3.3GHz, 6x256 kB L2 cache, 15MB L3 cache
    (CPU frequency pinned to 3.3 GHz)

NB: Optimizations landed in Erlang 22, but the performance figures have not yet been updated. The way the benchmarks are implemented is causing more GC than normal, and just running the benchmarks will produce pessimistic values. I have yet to figure out a better way of benchmarking, before I can update the figures.. See https://github.com/erlang/otp/commit/7d941c529d#commitcomment-31091771 for more info.

The benchmarks are all done with the exact same messages files and proto files. The source of the benchmarks was found in the Google protobuf's svn repository. The gpb originally did not support groups, and the benchmarks in the protobuf used groups, so I converted the google_message*.dat to use sub message structures instead. For protobuf, that change was only barely noticeable.

For performance, the generated Erlang code avoids creating sub binaries as far as possible. It has to for sub messages, strings and bytes, but for the rest of the types, it avoids creating sub binaries, both during encoding and decoding (for info, compile with the bin_opt_info option)

The Erlang code ran in the smp emulator, though only one CPU core was utilized.

The generated C++ core was compiled with -O3.

Version numbering

The gpb version number is fetched from the git latest git tag matching N.M where N and M are integers. This version is inserted into the gpb.app file as well as into the include/gpb_version.hrl. The version is the result of the command

git describe --always --tags --match '[0-9]*.[0-9]*'

Thus, to create a new version of gpb, the single source from where this version is fetched, is the git tag. (If you are importing gpb into another version control system than git, or using another build tool than rebar, you might have to adapt rebar.config and src/gpb.app.src accordingly. See also the section below about building outside of a git work tree for info on exporting gpb from git.)

The version number from the git describe command above will look like

• <x>.<y>.<z> (on master on Github)
• <x>.<y>.<z>-<n>-g<sha> (on branches or between commits)

The version number on the master branch of the gpb on Github is intended to always be only integers with dots, in order to be compatible with reltool. In other words, each push to Github's master branch is considered a release, and the version number is bumped. To ensure this, there is a pre-push git hook and two scripts, install-git-hooks and tag-next-minor-vsn, in the helpers subdirectory. The ChangeLog file will not necessarily reflect all minor version bumps, only important updates.

Places to update when making a new version:

• Write about the changes in the ChangeLog file, if it is a non-minor version bump.
• tag it in git

Building outside of a git work tree

The gpb build process expects a (non-shallow) git work tree, with tags, to get the version numbering right, as described in the Version numbering section, but it is also possible to build outside of git. To do that, you have two options:

• set the version manually by creating a file, gpb.vsn, with the version on the first line
• or create a versioned archive, using the helpers/mk-versioned-archive script, then unpack the archive and build inside it.

If you create the versioned archive in a git worktree, the version will be set automatically, otherwise you will need to specify it manually. Run mk-versioned-archive --help for info on what options to use.

When downloading from Github, the gpb-<x.y.z>.tar.gz archives have been created using the mk-versioned-archive script, so it is possible to just unpack and build directly.

If you use Github's automatic Source code zip or tar.gz archives, you will need to either create a gpb.vsn file as described above, or re-create a versioned archive using the mk-versioned-archive script and the --override-version=<x> option (or possibly the or the --override-version-from-cwd-path option if the directory name contains a proper version.)

Related projects

• rebar3_gpb_plugin for using gpb with rebar3
• exprotobuf for using gpb from Elixir
• enif_protobuf for a NIF encoder/decoder
• grpcbox, for creating grpc services (client and server)
• grpc and grpc_client for a grpc server and client

Contributing

Contributions are welcome, preferably as pull requests or git patches or git fetch requests. Here are some guide lines:

• Use only spaces for indentation, no tabs. Indentation is 4 spaces.
• The code must fit 80 columns
• Verify that the code and documentation compiles and that tests are ok:
  rebar clean; rebar eunit && rebar doc
  (if you are still on rebar2, you will need to run rebar compile before eunit)
• If you add a feature, test cases are most welcome, so that the feature won't get lost in any future refactorization
• Use a git branch for your feature. This way, the git history will look better in case there is need to refetch.
• By submitting patches to this project, you agree to allow them to be redistributed under the project's license according to the normal forms and usages of the open-source community.

Version history

See the ChangeLog for defails.

Major change in version 4.0.0:

The default value for the maps_unset_optional option has changed to omitted, from present_undefined This concerns only code generated with the maps (-maps) options. Projects that already set this option explicitly are not impacted. Projects that relied on the default to be present_undefined will need to set the option explicitly in order to upgrade to 4.0.0.

For type specs, the default has changed to generate them when possible. The option {type_specs,false} (-no_type) can be used to avoid generating type specs.