- 2020 April 27: Initial Draft
- 2020 August 5: Update guidelines
Accepted
Protocol Buffers provide a basic style guide and Buf builds upon that. To the extent possible, we want to follow industry accepted guidelines and wisdom for the effective usage of protobuf, deviating from those only when there is clear rationale for our use case.
The adoption of google.protobuf.Any as the recommended approach for encoding
interface types (as opposed to oneof) makes package naming a central part
of the encoding as fully-qualified message names now appear in encoded
messages.
Thus far we have mostly followed Buf's DEFAULT
recommendations, with the minor deviation of disabling PACKAGE_DIRECTORY_MATCH
which although being convenient for developing code comes with the warning
from Buf that:
you will have a very bad time with many Protobuf plugins across various languages if you do not do this
In ADR 021, gRPC was adopted for Protobuf native queries. The full gRPC service path thus becomes a key part of ABCI query path. In the future, gRPC queries may be allowed from within persistent scripts by technologies such as CosmWasm and these query routes would be stored within script binaries.
The goal of this ADR is to provide thoughtful naming conventions that:
- encourage a good user experience for when users interact directly with .proto files and fully-qualified protobuf names
- balance conciseness against the possibility of either over-optimizing (making names too short and cryptic) or under-optimizing (just accepting bloated names with lots of redundant information)
These guidelines are meant to act as a style guide for both the SDK and third-party modules.
As a starting point, we should adopt all of the DEFAULT
checkers in Buf's including PACKAGE_DIRECTORY_MATCH,
except:
Further guidelines to be described below.
Names should be descriptive enough to convey their meaning and distinguish them from other names.
Given that we are using fully-qualifed names within
google.protobuf.Any as well as within gRPC query routes, we should aim to
keep names concise, without going overboard. The general rule of thumb should
be if a shorter name would convey more or else the same thing, pick the shorter
name.
For instance, cosmos.bank.MsgSend (19 bytes) conveys roughly the same information
as cosmos_sdk.x.bank.v1.MsgSend (28 bytes) but is more concise.
Such conciseness makes names both more pleasant to work with and take up less space within transactions and on the wire.
We should also resist the temptation to over-optimize, by making names
cryptically short with abbreviations. For instance, we shouldn't try to
reduce cosmos.bank.MsgSend to csm.bk.MSnd just to save a few bytes.
The goal is to make names concise but not cryptic.
Package and type names should be chosen for the benefit of users, not necessarily because of legacy concerns related to the go code-base.
In the interests of long-term support, we should plan on the names we do choose to be in usage for a long time, so now is the opportunity to make the best choices for the future.
In general, schema evolution is the way to update protobuf schemas. That means that new fields, messages, and RPC methods are added to existing schemas and old fields, messages and RPC methods are maintained as long as possible.
Breaking things is often unacceptable in a blockchain scenario. For instance, immutable smart contracts may depend on certain data schemas on the host chain. If the host chain breaks those schemas, the smart contract may be irreparably broken. Even when things can be fixed (for instance in client software), this often comes at a high cost.
Instead of breaking things, we should make every effort to evolve schemas rather than just breaking them. Buf breaking change detection should be used on all stable (non-alpha or beta) packages to prevent such breakage.
With that in mind, different stable versions (i.e. v1 or v2) of a package should more or less be considered
different packages and this should be last resort approach for upgrading protobuf schemas. Scenarios where creating
a v2 may make sense are:
- we want to create a new module with similar functionality to an existing module and adding
v2is the most natural way to do this. In that case, there are really just two different, but similar modules with different APIs. - we want to add a new revamped API for an existing module and it's just too cumbersome to add it to the existing package,
so putting it in
v2is cleaner for users. In this case, care should be made to not deprecate support forv1if it is actively used in immutable smart contracts.
The following guidelines are recommended for marking packages as alpha or beta:
- marking something as
alphaorbetashould be a last resort and just putting something in the stable package (i.e.v1orv2) should be preferred - a package should be marked as
alphaif and only if there are active discussions to remove or significantly alter the package in the near future - a package should be marked as
betaif and only if there is an active discussion to significantly refactor/rework the functionality in the near future but not remove it - modules can and should have types in both stable (i.e.
v1orv2) and unstable (alphaorbeta) packages.
alpha and beta should not be used to avoid responsibility for maintaining compatibility.
Whenever code is released into the wild, especially on a blockchain, there is a high cost to changing things. In some
cases, for instance with immutable smart contracts, a breaking change may be impossible to fix.
When marking something as alpha or beta, maintainers should ask the questions:
- what is the cost of asking others to change their code vs the benefit of us maintaining the optionality to change it?
- what is the plan for moving this to
v1and how will that affect users?
alpha or beta should really be used to communicate "changes are planned".
As a case study, gRPC reflection is in the package grpc.reflection.v1alpha. It hasn't been changed since
2017 and it is now used in other widely used software like gRPCurl. Some folks probably use it in production services
and so if they actually went and changed the package to grpc.reflection.v1, some software would break and
they probably don't want to do that... So now the v1alpha package is more or less the de-facto v1. Let's not do that.
The following are guidelines for working with non-stable packages:
- Buf's recommended version suffix
(ex.
v1alpha1) should be used for non-stable packages - non-stable packages should generally be excluded from breaking change detection
- immutable smart contract modules (i.e. CosmWasm) should block smart contracts/persistent
scripts from interacting with
alpha/betapackages
Instead of using Buf's recommended version suffix,
we can omit v1 for packages that don't actually have a second version. This
allows for more concise names for common use cases like cosmos.bank.Send.
Packages that do have a second or third version can indicate that with .v2
or .v3.
Top-level packages should adopt a short name that is known to not collide with
other names in common usage within the Cosmos ecosystem. In the near future, a
registry should be created to reserve and index top-level package names used
within the Cosmos ecosystem. Because the Cosmos SDK is intended to provide
the top-level types for the Cosmos project, the top-level package name cosmos
is recommended for usage within the Cosmos SDK instead of the longer cosmos_sdk.
ICS specifications could consider a
short top-level package like ics23 based upon the standard number.
Sub-package depth should be increased with caution. Generally a single
sub-package is needed for a module or a library. Even though x or modules
is used in source code to denote modules, this is often unnecessary for .proto
files as modules are the primary thing sub-packages are used for. Only items which
are known to be used infrequently should have deep sub-package depths.
For the Cosmos SDK, it is recommended that that we simply write cosmos.bank,
cosmos.gov, etc. rather than cosmos.x.bank. In practice, most non-module
types can go straight in the cosmos package or we can introduce a
cosmos.base package if needed. Note that this naming will not change
go package names, i.e. the cosmos.bank protobuf package will still live in
x/bank.
Message type names should be as concise possible without losing clarity. sdk.Msg
types which are used in transactions will retain the Msg prefix as that provides
helpful context.
ADR 021 specifies that modules should
implement a gRPC query service. We should consider the principle of conciseness
for query service and RPC names as these may be called from persistent script
modules such as CosmWasm. Also, users may use these query paths from tools like
gRPCurl. As an example, we can shorten
/cosmos_sdk.x.bank.v1.QueryService/QueryBalance to
/cosmos.bank.Query/Balance without losing much useful information.
RPC request and response types should follow the ServiceNameMethodNameRequest/
ServiceNameMethodNameResponse naming convention. i.e. for an RPC method named Balance
on the Query service, the request and response types would be QueryBalanceRequest
and QueryBalanceResponse. This will be more self-explanatory than BalanceRequest
and BalanceResponse.
Instead of Buf's default service suffix recommendation,
we should simply use the shorter Query for query services.
For other types of gRPC services, we should consider sticking with Buf's default recommendation.
Get and Query should be omitted from Query service names because they are
redundant in the fully-qualified name. For instance, /cosmos.bank.Query/QueryBalance
just says Query twice without any new information.
A registry of top-level package names should be created to coordinate naming across the ecosystem, prevent collisions, and also help developers discover useful schemas. A simple starting point would be a git repository with community-based governance.
- names will be more concise and easier to read and type
- all transactions using
Anywill be at shorter (_sdk.xand.v1will be removed) .protofile imports will be more standard (without"third_party/proto"in the path)- code generation will be easier for clients because .proto files will be
in a single
proto/directory which can be copied rather than scattered throughout the SDK
.protofiles will need to be reorganized and refactored- some modules may need to be marked as alpha or beta