EPIC: Separate all SDK modules into standalone go modules

[aaronc]

Overview

This issue outlines a roadmap for breaking up all of the x/... SDK modules and simapp into standalone go modules using the app wiring design .

• Update the ADR for reflecting the implementation (first draft: docs: add ADR 057 App Wiring Part I #11873)

Problem Definition

The key culprit of the SDK not being decomposable into separate go modules is the dependency on simapp. simapp imports every module but every module and most packages also import simapp into their tests. Overall there are 184 references to simapp in other packages (run Code > Analyze Code > Backwards Dependencies… on simapp in Goland or IntelliJ to create this report). The reason for such widespread usage of simapp in tests is because there needs to be a way to do integration tests in the SDK and simapp provides the key test fixture. Unfortunately, this creates a Gordian knot where all the modules depend on each other and breaking them up is almost impossible. App wiring provides a solution to this problem by creating a way to quickly spin up integration test fixtures locally in each module rather than using a single SDK wide test fixture that brings all these problems.

Phase 1: App Wiring MVP

In this first phase, we create an MVP of app wiring that has these requirements:

• can refactor a single module (we’re choosing x/nft) to not depend on simapp at all for testing or any other reason
• can refactor part of simapp/app.go to use the new app wiring while still wiring up everything else as before - this is required to make the app wiring migration gradual and opt-in

For now, the app wiring framework that we will create will forego dealing with any of the messy issues related to protobuf generated code and semantic versioning. This means basically that app wiring will use the global protobuf registry for decoding the app config and the existing gogo proto-based codec package will be used elsewhere.

• Single module (x/nft) tests using app.yaml instead of simapp (full feature branch in feat: minimum viable app wiring #11900), being broken down into separate PRs:
  • fix(container): issue with multiple dependencies coming from an out struct #11912
  • fix(container)!: fix issue providing a module-scoped dep from a module #11913
  • feat: add core module with app config support #11914
  • feat(container): automatic debugging on errors #11915
  • feat: start app wiring with runtime and x/params modules #11924
  • additional app wiring setups from feat: minimum viable app wiring #11900
    • feat(x/auth): app wiring setup #12019
    • feat(x/bank): app wiring migration #12032
    • migrate x/staking to use app wiring #12036
    • migrate x/genutil to use app wiring #12035
    • migrate x/nft to use app wiring #12038
  • x/nft cli tests and testutil/network from feat: minimum viable app wiring #11900
• Setup vanity URL for cosmossdk.io/core go module #11904
• Rename container go module cosmossdk.io/depinject #11905
• Rename api module cosmossdk.io/api #11906
• Consider renaming container.Option #11917
• Remove container graphviz dependency #11925
• Consider renaming container.AutoGroupType #11935
• Consider renaming the container module #11937

Phase 2: Decouple modules and packages from simapp

Following the patterns demonstrated in x/nft do the same for each module and package which depends on simapp in the SDK, essentially:

• create a protobuf Module config message and register a module implementation with appmodule.Register
• inject the new module into simapp
• migrate CLI tests to use appconfig Superseded by EPIC: Unit Testing of Modules via Mocks #12398
• migrate simulations and any other code requiring simapp Superseded by EPIC: Unit Testing of Modules via Mocks #12398

Modules should be migrated starting with those with fewer dependencies on other modules first.

• migrate x/params to use app wiring #12023
• migrate x/groups to use app wiring  #12024
• migrate x/authz to use app wiring #12084
• migrate x/feegrant to use app wiring #12073
• migrate x/crisis to use app wiring #12194
• migrate x/capability to use app wiring #12069
• migrate x/upgrade to use app wiring #12298
• migrate x/evidence to use app wiring #12238
• migrate x/auth to use app wiring #12302
• migrate x/bank to app wiring #12083
• migrate x/staking to use app wiring #12036
• migrate x/slashing to use app wiring #12195
• migrate x/distribution to app wiring #12292
• migrate x/mint to use app wiring #12234
• x/genutil
• migrate x/gov to use app wiring #12274
• migrate vesting to use app wiring #12300

These other issues can also be addressed incrementally during phase 2:

• Add SimulationManager handling to runtime.App #12054
• Handle additional x/auth module app config parameters #12055
• App wiring setup for tx's and ante/post handlers #12056
• Create legacy simapp to test along new app wiring simapp #12068
• depinject module upgrades based to make this work easier (nice to have):
  • Refactor container Run to Build #11907
  • Provide useful error when an interface can't be resolved but there are implementing types in depinject #11908
  • Provide useful error when a type can't be resolved but there are matching types in unreferenced modules #11909
  • Allow depinject interface types to be resolved using a key #11910
  • Allow depinject interface type resolution to be configured per module #11911
  • Better error for unexported fields with container.In/Out #11943
  • Write container README.md docs #11944
  • Should all injectable types require an interface? #12252
• Move simapp.DefaultConsensusParams() out of simapp #11903
• Extend core/appconfig documentation #12237
• Improve upgrade and mint app-wiring module configuration  #12400

As this work proceeds, simapp itself can be refactored to use its app.yaml more and more - this can be done incrementally as each module is migrated or at the end when simapp is standalone.

Other packages besides the x/* modules depend on simapp as well, usually to call simapp.MakeTestEncodingConfig() or simapp.DefaultConsensusParams(). DefaultConsensusParams should have been moved in Phase 1, but MakeTestEncodingConfig may require app wiring because it used to build a codec with some modules registered. A strategy for decoupling the following packages from simapp will need to be created:

• Remove dependency from baseapp -> simapp #12535
• Remove dependency client -> simapp #12544
• Remove dependency from codec -> simapp  #12528
• Remove dependency from crypto -> simapp #12612
• Remove dependency server -> simapp #12619
• Remove dependency on testutil -> simapp #12546
• Remove dependency types -> simapp #12584

At this point, a review of all the migrated modules should be performed. The review should verify the tests coverage and the test behaviors is kept unchanged compared to before the migration.

Phase 3: Make simapp a standalone go module

Instructions in https://github.com/golang/go/wiki/Modules#is-it-possible-to-add-a-module-to-a-multi-module-repository must be followed. Also, we want to make tags off of a release branch, not main. It is expected that these steps will be required:

• make simapp a standalone go module github.com/cosmos/cosmos-sdk/simapp which imports github.com/cosmos/cosmos-sdk probably using a replace tag to start. Any remaining dependencies of the rest of the SDK on simapp must be removed by now

Phase 4: Make all modules standalone

Again following the instructions from https://github.com/golang/go/wiki/Modules#is-it-possible-to-add-a-module-to-a-multi-module-repository and starting with the modules with the fewest dependencies to the most.
List of To-dos when creating a go.mod: https://github.com/cosmos/cosmos-sdk/blob/main/CONTRIBUTING.md#gomod

• x/nft
• x/params
• extract x/group to its own module #17936
• x/authz
• x/feegrant
• x/crisis
• x/capability
• x/upgrade
• x/evidence
• x/auth
• x/bank
• x/staking
• x/slashing
• x/distribution
• Extract x/mint to its own go.mod #17813
• go.mod: carve out runtime Go module #18411

Phase 5

• audit dependency graph
• remove cosmossdk.io/x/* dependence in the root go module #18290
• cleanup intermodule dependency

NOTE: it may not be possible to move all modules out of the core go module with this strategy alone because of other cyclic dependencies between each other (such as x/auth and x/bank). Other strategies may be needed or there may be an inseparable set of modules that stay in the SDK for now and are decoupled later if needed.

[julienrbrt]

The work on this EPIC should use the branch epic/app-wiring as base.

[aaronc]

We've never used feature branches and have no process for them. Let's just consider this epic paused

[tac0turtle]

closing this as most of the work is completed and it will continue in #20279