Protocol Buffer Concepts
External
- https://protobuf.dev/programming-guides/proto3/
- https://learning.oreilly.com/videos/complete-introduction-to/9781789349344/
- https://protobuf.dev/programming-guides/style/
Internal
Overview
Protocol Buffers is a language-neutral, platform-neutral, extensible mechanism for serializing structured data. The schema is define once, and it used to generate code that serializes and deserializes schema-conforming data. The main use case for Protocol Buffers is sharing data across programming languages. Data can be written and serialized in one language, sent over the network and then and deserialized and interpreted in a different programming language.
Protocol Buffers offers the following advantages:
- It allows defining types, and the data is fully typed when exchanged. We know the type of data in transit.
- Data is compressed automatically.
- Serialization/deserialization is efficient.
- Comes with a schema, in form of the
.protofiles, which is used to generate code that writes and reads the data. - Schema supports embedded documentation.
- Schema can evolve over time in a safe manner. The implementations that rely on schema can stay backward and forward compatible.
One of the disadvantages is that the data is encoded in a binary format, so it can't be visualized with a text editor.
The typical workflow consist in defining the data types, called messages, in Protocol Buffer .proto files, then automatically generating the data structures to support and validate the data types, in the programming language of choice. In Go, the messages are represented as structs. With the help of a framework like gRPC, which uses Protocol Buffers as default serialization format and native mechanism to exchange data, client and server code can also be automatically generated. This renders Protocol Buffer convenient for use as serialization format for microservices.
The current version of the protocol is 3, released mid 2016.
Message
Protocol Buffers represent arbitrary data types as messages. "Message" terminology is probably used because the instances of the types defines as such are mainly intended to be sent over the wire.
A message has fields.
Message names must be conventionally rendered in CamelCase with an initial capital. Prefer to capitalize abbreviations as single words: GetDnsRequest rather than GetDNSRequest. Also see Style.
One or more messages are declared in .proto text file with the following format:
syntax = "proto3";
/* Person is used to identify
a user in the system.
*/
message Person {
// the age as of person's creation
int32 age = 1;
string first_name = 2;
string last_name = 3;
bytes small_picture = 4; // a small JPEG file
bool is_profile_verified = 5;
float height = 6;
repeated string phone_numbers = 7;
}
A message (type) can reference other type:
message Something {
...
}
message SomethingElse {
Something something = 1;
}
Nested Messages
Message definitions can be nested:
message Something {
...
message SomethingElse {
...
}
}
Fields
Each field has a field type, a field name and a field tag
<field_type> <field_name> = <field_tag>;
Unless explicitly set up by the program, every field is initialized with the default type value in serialized messages.
Field Names
Field names must be conventionally rendered in lower_snake_case, including oneof field and extension names. Also see Style.
Fields names are not important when the message is serialized/deserialized. Only the tags matter. Field names are important for your program, though, in the generated code.
Field Types
Field Tags
A tag is an integral value between 1 and 229-1 (536,879,911). The numbers between 19,000 and 19,999 cannot be used. The tag of an enum's first element must be 0.
Tags from 1 to 15 use one byte of space, so use them for frequently populated fields. The tags from 16 to 2047 use 2 bytes.
Default Value
Every field has a default value, which is defined by the field's type. The default value is always used unless the field is explicitly set up by the program. There's no such a concept as "required" field or "optional" field. If the field is not explicitly set in the program, it takes the default value.
This could potentially be a problem, because we cannot tell by just looking at a message that has a default value for a field if the sender explicitly set the field to the default value or they did not provide any value. For example, if we look at an account balance field and we see zero, we can't say if this is the actual account balance or is a default that kicked in.
If it is possible, choose default values that do not have any meaning for the business. 0, "", etc. should be very special values, meaning "missing".
Comments
// this is a single-line comment
/* This is
a multi-line
comment.
*/
Services
Protocol Buffers can define services that use messages to exchange data.
A service is a set of endpoints with different semantics that can be used to call into the service, by sending a request, and then receiving a response.
service SomeService {
rpc SomeEndpoint(SomeRequest) returns (SomeResponse);
rpc SomeOtherEndpoint(SomeOtherRequest) returns (SomeOtherResponse);
}
message SomeRequest {
...
}
message SomeResponse {
...
}
message SomeOtherRequest {
...
}
message SomeOtherResponse {
...
}
.proto Files
Multiple messages can be defined in the same .proto file.
Simple Project Layout
This is an example of a simple project that contains just one Protocol Buffer package, which translates into one Go package. Much more complex combinations are possible, and some are explained below in Imports and Packages. Project layout:
.
├── go.mod
├── pkg
│ ├── main
│ │ └── main.go
│ └── somepkgpb # Generated
│ └── file1.pb.go # Generated
└── protobuf
└── somepkgpb
└── file1.proto
For more details on Go project layouts, see
The simplest .proto file:
syntax = "proto3";
option go_package = "./somepkgpb"; // The Go package and the Protocol Buffer Package have the same name.
// This is not a requirement, but makes the code base easier to understand.
package somepkgpb; // The Protocol Buffer package and the Go have the same name. This is
// not a requirement, but makes the code base easier to understand.
message SomeMessage {
}
To generate the Go code:
protoc --proto_path=./protobuf --go_out=./pkg somepkgpb/file1.proto
A slightly more complex example is available in
For more details on code generation, see:
Import
Different messages can live in different .proto files and can be imported. This feature encourages modularization and sharing.
The types declared in file B.proto can be imported in the file A.proto by using an import statement in file A.proto. Note that unlike for Go, Protocol Buffer import requires paths to .proto files, not packages.
.
├── pkg
│ └── somethingpb
│ ├── A.pb.go
│ └── B.pb.go
│
└── protobuf
└── somethingpb
├── A.proto
└── B.proto
A.proto:
syntax = "proto3";
option go_package = "./somethingpb";
import "./somethingpb/B.proto"; // The .proto file path is relative to
// the --proto_path used by the compiler.
message A {
B b = 1;
}
B.proto:
syntax = "proto3";
option go_package = "./somethingpb";
message B {
}
Note that the path used in the import statement is relative to the --proto_path used by the compiler:
protoc --proto_path=./protobuf --go_out=./pkg somethingpb/A.proto somethingpb/B.proto
There is no correlation between the Go import path and the .proto import path.
For more details on code generation, see:
Packages
Protocol Buffers have the concept of package, whose semantics is similar to that of a Go package: a namespace for names. In Protocol Buffer's case, these are message names. When a message is declared in a package, with the package specifier, the package must be imported, and the name of a message must be prefixed with the fully qualified package name if we want to use that message in other packages.
Package names should be in lowercase. Package names should have unique names based on the project name, and possibly based on the path of the file containing the protocol buffer type definitions. If the Protocol Buffer package name contains dots ("."), they will be translated to underscores "_" upon Go code generation.
There is no correlation between the Go import path and package name, and the package specifier in the .proto file. The latter is only relevant to the protobuf namespace, while the formers are only relevant to the Go namespace. Conventionally, Go package names end in "pb". Protocol Buffer package names do not follow such convention.
However, if we want to generate Go code from Protocol Buffer files that declare multiple packages, it is a good idea to design Protocol Buffer packages to map onto Go packages - each Protocol Buffer package should have its corresponding Go package. If we maintain this convention, it'll help with code comprehension. To exemplify this, we will declare two different Protocol Buffer packages that are going to be translate into their corresponding, homonymous Go packages.
.
├── go.mod
├── pkg
│ ├── blue
│ │ └── blue.pb.go
│ └── red
│ └── red.pb.go
└── protobuf
└── protobuf-internal-dir
├── blue.proto
└── red.proto
blue.proto
syntax = "proto3";
option go_package = "./blue"; // The Go package and the Protocol Buffer Package have the same name.
// This is not a requirement, but makes the code base easier to understand.
package blue; // The Protocol Buffer package and the Go have the same name. This is
// not a requirement, but makes the code base easier to understand.
// The "A" name is used both in the "blue" and "red" package, so when used in the "red" package
// will have to be qualified as blue.A.
message A {
}
red.proto
syntax = "proto3";
option go_package = "./red"; // The Go package and the Protocol Buffer Package have the same name.
// This is not a requirement, but makes the code base easier to understand.
package red; // The Protocol Buffer package and the Go have the same name. This is
// not a requirement, but makes the code base easier to understand.
import "protobuf-internal-dir/blue.proto";
// The "A" name is used both in the "blue" and "red" package. They can be used even in the
// same file (this one), if they are appropriately prefixed with the package name.
message A {
blue.A something = 2;
}
The files are compiled with:
protoc \ --proto_path=./protobuf \ --go_out=./pkg \ protobuf-internal-dir/blue.proto \ protobuf-internal-dir/red.proto
For more details on code generation, see:
Services
TODO: https://protobuf.dev/programming-guides/style/#services
Go Code Generation
Style
Protocol Buffer Go Code Examples
- Marshal/Unmarshal Protocol Buffers in Go
- Converting Protocol Buffers Messages to and from JSON in Go
Data Evolution with Protocol Buffers
Schemas evolve over time to keep up with business requirements. Protocol Buffers has been designed to allow data evolution. Data formats defined with Protocol Buffers can evolve in a safe way, so code that uses the old and new versions of the schema continue to interoperate.
There are essential two scenarios we need to think about.
One involves old clients, which rely on old versions of the schema, continue to send data into a newer service version. They must be supported, which means that the current system has to be compatible with old version of itself, hence backward compatible.
The second scenario involve new clients, which use a newer version of the schema, sending data into older service instances that use an older version of the schema. The old system must be compatible with newer versions of itself, hence forward compatible.
Protocol buffer ensures full compatibility.
Full compatibility with Protocol Buffer is ensured if the following rules are followed:
- Don't change the numeric tags for any existing fields.
- New fields can be added. The old code will ignore them, but will correctly process the old fields. This is forward compatibility.
- If old fields are deprecated and not sent anymore (they are marked OBSOLETE_, or literally not embedded in the message anymore, by making the tag
reserved), they don't disappear from schema, but they will always carry the default value upon deserialization. - There are rules for type changes.
Adding Fields
To add a field, use the next, unallocated tag number.
Code reading data using old schema will not know what the new tag number corresponds to, and it will just ignore and drop it.
If old data that was generated with old schema is read by code that uses the new schema and expects the field, it will obviously not find it, and use the default value for that type.
This implies that default values must always interpreted with care.
Renaming Fields
Field names can be changes, nothing really happens, as long as the tag is not changed.
Removing Fields
If a field is removed, the code using the old schema won't find it, and will use the default value.
However, care must be taken to prevent reusing the name and the tag corresponding to the removed field with a new fields, with a different semantics. For that, both the field name and the tag of the field that is being retired must be declared as reserved. The tags are reserved so they won't be reused in communication. The names are reserved to prevent code bugs.
Never remove reserved tags or field names.
// version 1
message SomeMessage {
int32 id = 1;
string name = 2;
}
// version 2 with the "name" field removed
message SomeMessage {
reserved 2;
reserved "name";
int32 id = 1;
}
Declaring Fields OBSOLETE_
The disadvantage to declaring fields OBSOLETE_ instead of removing them is that you will have to continue to provide values.
TODO.
Evolving Enums
Enums can evolve like any other field: we can add, remove and reserve values.