evanmcc/gpb

A Google Protobuf implementation for Erlang

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

See https://developers.google.com/protocol-buffers/ for further information on the Google protocol buffers.

Build Status

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="a@example.com"}).    
<<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 = "a@example.com"}

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}
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.

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,
     i2 => undefined}

   %% With the maps option and the maps_unset_optional set to omitted:
   #{i1 => 17}

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"}}
   #{u => undefined}     % If maps_unset_optional = present_undefined (default)
   #{}                   % With the maps_unset_optional set to omitted

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 optional nor default=<x> for fields. Instead all fields are implicitly optional, and if missing in the binary to decode, they always decode to the type-specific default value. Also, it is not possible to determine whether a value was present---with a type-specific value---or not; no has_<field>() methods are generated (at least for scalars). If you need detection of "missing" data, you must define has_<field> boolean fields and set them appropriately.

This maps directly and naturally to Erlang.

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

  gpb reads but ignores or throws away:

  • options other than 'packed' or 'default'
  • custom options

  gpb does not support:

  • groups
  • aggregate custom options introduced in protobuf 2.4.0
  • rpc
  • proto3 JSON mapping

• Characteristics of gpb:

  • Skipping over unknown message fields, 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 copying 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.
  • 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 the 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(), 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 version information functions:

  • gpb:version_as_string() and gpb:version_as_list()
  • GeneratedCode:version_as_string() and GeneratedCode:version_as_list()
  • ?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.

• Caveats

  The gpb does accept reserved words as names for fields (just like protoc does), but not as names for messages. To correct this, one would have to either rewrite the grammar, or stop using yecc. (maybe rewrite it all as a protoc plugin?)

Interaction with rebar

For info on how to use gpb with rebar3, see https://www.rebar3.org/docs/using-available-plugins#section-using-gpb

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)

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 does not support groups, but 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.)

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 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

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 compile eunit doc xref
• 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.