Protobuf resources

Style guide

This is our style guide for Protobuf. It's purposefully concise, and is meant as a short reference for developers to refer to when writing Protobuf schemas. These recommendations are not enforced by the Buf Schema Registry, but are here for reference.

The requirements follow the DEFAULT lint category in the Buf CLI. For details on each rule and its rationale, see that documentation. Within this style guide, we provide (Why?) links where relevant for each check.

Our style guide is designed to provide consistency and maintainability across a Protobuf schema of any size and any purpose, but without being so opinionated as to restrict organizations from making the design decisions they need to make for their individual APIs.

Requirements

Files and packages

All files should have a package defined. (Why?)

All files of the same package should be in the same directory. All files should be in a directory that matches their package name. (Why?)

For example, if we have a module defined in the proto directory, we expect these package values:

.
β”œβ”€β”€ buf.yaml
└── proto
    └── foo
        └── bar
            β”œβ”€β”€ bat
            β”‚Β Β  └── v1
            β”‚Β Β      └── bat.proto // package foo.bar.bat.v1
            └── baz
                └── v1
                    β”œβ”€β”€ baz.proto         // package foo.bar.baz.v1
                    └── baz_service.proto // package foo.bar.baz.v1

Packages should be lower_snake_case. (Why?)

The last component of a package should be a version. (Why?)

Files should be named lower_snake_case.proto (Why?)

All of the file options below should have the same value, or all be unset, for all files that have the same package: (Why?)

  • csharp_namespace
  • go_package
  • java_multiple_files
  • java_package
  • php_namespace
  • ruby_package
  • swift_prefix

For example, if we have file foo_one.proto:

foo_one.proto
syntax = "proto3";

package foo.v1;

option go_package = "foov1";
option java_multiple_files = true;
option java_package = "com.foo.v1";

Another file foo_two.proto with package foo.v1 must have these three options set to the same value, and the other options unset:

foo_two.proto
syntax = "proto3";

package foo.v1;

option go_package = "foov1";
option java_multiple_files = true;
option java_package = "com.foo.v1";

Imports

No imports should be declared as public or weak. (Why?)

Enums

Enums should not have the allow_alias option set. (Why?)

Enum names should be PascalCase. (Why?)

Enum value names should be UPPER_SNAKE_CASE. (Why?)

Enum value names should be prefixed with the UPPER_SNAKE_CASE of the enum name. (Why?). For example, given the enum FooBar, all enum value names should be prefixed with FOO_BAR_.

The zero value for all enums should be suffixed with _UNSPECIFIED. (Why?) The suffix is configurable. For example, given the enum FooBar, the zero value should be FOO_BAR_UNSPECIFIED = 0;.

Messages

Message names should be PascalCase. (Why?)

Field names should be lower_snake_case. (Why?)

Oneof names should be lower_snake_case. (Why?)

Services

Service names should be PascalCase. (Why?)

Service names should be suffixed with Service. (Why?) The suffix is configurable.

RPC names should be PascalCase. (Why?)

All RPC request and responses messages should be unique across your Protobuf schema. (Why?)

All RPC request and response messages should be named after the RPC, either by naming them MethodNameRequest, MethodNameResponse or ServiceNameMethodNameRequest, ServiceNameMethodNameResponse. (Why?)

Use of the WKT Empty as a request and response type

We recommend that you use a separate request and response message for each RPC, because it gives you maximum flexibility to evolve your RPCs without breaking backward compatibility. This applies to Protobuf in generalβ€”both gRPC and ConnectRPC.

As a consequence, you shouldn't import and use Empty if your RPC doesn't happen to have any request or response data yet. You should instead define a custom empty request and/or response message per RPC. That way, when your request or response message does eventually contain fields, you can add them without fear of breaking changes.

If you're really trying to model an empty message (as opposed to a message whose shape you're not yet sure about) and you're sure the message will always remain empty, then using Empty is fine. However, using custom empty request and response types is much more future-proof.

If you want to use Empty and you're using the Buf CLI to lint, you should set the corresponding flags in your buf.yaml file to avoid lint errors:

buf.yaml with flags set to allow Empty
version: v1
lint:
  use:
    - DEFAULT
  rpc_allow_google_protobuf_empty_requests: true
  rpc_allow_google_protobuf_empty_responses: true

Recommendations

While not strictly related to style, you should always set up breaking change detection for your Protobuf schema. See the breaking change detection documentation for more details on how Buf helps enforce this.

Use // instead of /* */ for comments.

Over-document, and use complete sentences for comments. Put documentation above the type, instead of inline.

Avoid widely used keywords for all types, especially packages. For example, if your package name is foo.internal.bar, the internal component blocks importing the generated stubs in other packages for Go.

Files should be laid out in this order (this matches Google's current recommendations), and the buf format command checks for all but the first two:

  • License header (if applicable)
  • File overview
  • Syntax
  • Package
  • Imports (sorted)
  • File options
  • Everything else

Use pluralized names for repeated fields.

Name fields after their type as much as possible. For example, for a field of message type FooBar, name the field foo_bar unless there is a specific reason to do otherwise.

Avoid using nested enums and nested messages. You may end up wanting to use them outside of their context message in the future, even if you don't think so at the moment.

Avoid streaming RPCs. They're difficult to implement and call, and they often require special configuration in proxies, firewalls, and other network infrastructure. In most circumstances, polling and pagination are much simpler and nearly as efficient. For the handful of cases where streaming RPCs are worth the complexity, add exceptions to your lint configuration.