Add in-place store migrations (#8485)

* Add 1st version of migrate

* Put migration logic into Configurator

* add test to bank store migration

* add test for configurator

* Error if no migration found

* Remove RunMigrations from Configurator interface

* Update spec

* Rename folders

* copy-paste from keys.go

* Fix nil map

* rename function

* Update simapp/app.go

Co-authored-by: Robert Zaremba <robert@zaremba.ch>

* Update simapp/app_test.go

Co-authored-by: Robert Zaremba <robert@zaremba.ch>

* Adderss reviews

* Fix tests

* Update testutil/context.go

Co-authored-by: Robert Zaremba <robert@zaremba.ch>

* Update docs for ConsensusVersion

* Rename to forVersion

* Fix tests

* Check error early

* Return 1 for intiial version

* Use MigrationKeeper

* Fix test

* Revert adding marshaler to Configurator

* Godoc updates

* Update docs

Co-authored-by: Robert Zaremba <robert@zaremba.ch>
Co-authored-by: Aaron Craelius <aaron@regen.network>

simapp/app.go

@@ -198,6 +198,9 @@ type SimApp struct {
 
 	// simulation manager
 	sm *module.SimulationManager
+
+	// the configurator
+	configurator module.Configurator
 }
 
 func init() {
@@ -393,7 +396,8 @@ func NewSimApp(
 
 	app.mm.RegisterInvariants(&app.CrisisKeeper)
 	app.mm.RegisterRoutes(app.Router(), app.QueryRouter(), encodingConfig.Amino)
-	app.mm.RegisterServices(module.NewConfigurator(app.MsgServiceRouter(), app.GRPCQueryRouter()))
+	app.configurator = module.NewConfigurator(app.MsgServiceRouter(), app.GRPCQueryRouter())
+	app.mm.RegisterServices(app.configurator)
 
 	// add test gRPC service for testing gRPC queries in isolation
 	testdata.RegisterQueryServer(app.GRPCQueryRouter(), testdata.QueryImpl{})
@@ -598,6 +602,28 @@ func (app *SimApp) RegisterTendermintService(clientCtx client.Context) {
 	tmservice.RegisterTendermintService(app.BaseApp.GRPCQueryRouter(), clientCtx, app.interfaceRegistry)
 }
 
+// RunMigrations performs in-place store migrations for all modules. This
+// function MUST be only called by x/upgrade UpgradeHandler.
+//
+// `migrateFromVersions` is a map of moduleName to fromVersion (unit64), where
+// fromVersion denotes the version from which we should migrate the module, the
+// target version being the module's latest ConsensusVersion.
+//
+// Example:
+//   cfg := module.NewConfigurator(...)
+//   app.UpgradeKeeper.SetUpgradeHandler("store-migration", func(ctx sdk.Context, plan upgradetypes.Plan) {
+//       err := app.RunMigrations(ctx, module.MigrationMap{
+//           "bank": 1,     // Migrate x/bank from v1 to current x/bank's ConsensusVersion
+//           "staking": 8,  // Migrate x/staking from v8 to current x/staking's ConsensusVersion
+//      })
+//      if err != nil {
+//           panic(err)
+//      }
+//   })
+func (app *SimApp) RunMigrations(ctx sdk.Context, migrateFromVersions module.MigrationMap) error {
+	return app.mm.RunMigrations(ctx, app.configurator, migrateFromVersions)
+}
+
 // RegisterSwaggerAPI registers swagger route with API Server
 func RegisterSwaggerAPI(ctx client.Context, rtr *mux.Router) {
 	statikFS, err := fs.New()

simapp/app_test.go

@@ -6,10 +6,13 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/require"
+	abci "github.com/tendermint/tendermint/abci/types"
 	"github.com/tendermint/tendermint/libs/log"
+	tmproto "github.com/tendermint/tendermint/proto/tendermint/types"
 	dbm "github.com/tendermint/tm-db"
 
-	abci "github.com/tendermint/tendermint/abci/types"
+	sdk "github.com/cosmos/cosmos-sdk/types"
+	"github.com/cosmos/cosmos-sdk/types/module"
 )
 
 func TestSimAppExportAndBlockedAddrs(t *testing.T) {
@@ -45,3 +48,87 @@ func TestGetMaccPerms(t *testing.T) {
 	dup := GetMaccPerms()
 	require.Equal(t, maccPerms, dup, "duplicated module account permissions differed from actual module account permissions")
 }
+
+func TestRunMigrations(t *testing.T) {
+	db := dbm.NewMemDB()
+	encCfg := MakeTestEncodingConfig()
+	app := NewSimApp(log.NewTMLogger(log.NewSyncWriter(os.Stdout)), db, nil, true, map[int64]bool{}, DefaultNodeHome, 0, encCfg, EmptyAppOptions{})
+
+	// Create a new configurator for the purpose of this test.
+	app.configurator = module.NewConfigurator(app.MsgServiceRouter(), app.GRPCQueryRouter())
+
+	testCases := []struct {
+		name         string
+		moduleName   string
+		forVersion   uint64
+		expRegErr    bool // errors while registering migration
+		expRegErrMsg string
+		expRunErr    bool // errors while running migration
+		expRunErrMsg string
+		expCalled    int
+	}{
+		{
+			"cannot register migration for version 0",
+			"bank", 0,
+			true, "module migration versions should start at 1: invalid version", false, "", 0,
+		},
+		{
+			"throws error on RunMigrations if no migration registered for bank",
+			"", 1,
+			false, "", true, "no migrations found for module bank: not found", 0,
+		},
+		{
+			"can register and run migration handler for x/bank",
+			"bank", 1,
+			false, "", false, "", 1,
+		},
+		{
+			"cannot register migration handler for same module & forVersion",
+			"bank", 1,
+			true, "another migration for module bank and version 1 already exists: internal logic error", false, "", 0,
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			var err error
+
+			// Since it's very hard to test actual in-place store migrations in
+			// tests (due to the difficulty of maintaing multiple versions of a
+			// module), we're just testing here that the migration logic is
+			// called.
+			called := 0
+
+			if tc.moduleName != "" {
+				// Register migration for module from version `forVersion` to `forVersion+1`.
+				err = app.configurator.RegisterMigration(tc.moduleName, tc.forVersion, func(sdk.Context) error {
+					called++
+
+					return nil
+				})
+
+				if tc.expRegErr {
+					require.EqualError(t, err, tc.expRegErrMsg)
+
+					return
+				}
+			}
+			require.NoError(t, err)
+
+			err = app.RunMigrations(
+				app.NewContext(true, tmproto.Header{Height: app.LastBlockHeight()}),
+				module.MigrationMap{
+					"auth": 1, "authz": 1, "bank": 1, "staking": 1, "mint": 1, "distribution": 1,
+					"slashing": 1, "gov": 1, "params": 1, "ibc": 1, "upgrade": 1, "vesting": 1,
+					"feegrant": 1, "transfer": 1, "evidence": 1, "crisis": 1, "genutil": 1, "capability": 1,
+				},
+			)
+			if tc.expRunErr {
+				require.EqualError(t, err, tc.expRunErrMsg)
+			} else {
+				require.NoError(t, err)
+				require.Equal(t, tc.expCalled, called)
+			}
+		})
+	}
+}

testutil/context.go

@@ -0,0 +1,25 @@
+package testutil
+
+import (
+	"github.com/tendermint/tendermint/libs/log"
+	tmproto "github.com/tendermint/tendermint/proto/tendermint/types"
+	dbm "github.com/tendermint/tm-db"
+
+	"github.com/cosmos/cosmos-sdk/store"
+	sdk "github.com/cosmos/cosmos-sdk/types"
+)
+
+// DefaultContext creates a sdk.Context with a fresh MemDB that can be used in tests.
+func DefaultContext(key sdk.StoreKey, tkey sdk.StoreKey) sdk.Context {
+	db := dbm.NewMemDB()
+	cms := store.NewCommitMultiStore(db)
+	cms.MountStoreWithDB(key, sdk.StoreTypeIAVL, db)
+	cms.MountStoreWithDB(tkey, sdk.StoreTypeTransient, db)
+	err := cms.LoadLatestVersion()
+	if err != nil {
+		panic(err)
+	}
+	ctx := sdk.NewContext(cms, tmproto.Header{}, false, log.NewNopLogger())
+
+	return ctx
+}

types/context_test.go

@@ -8,13 +8,11 @@ import (
 	"github.com/golang/mock/gomock"
 	"github.com/stretchr/testify/suite"
 	abci "github.com/tendermint/tendermint/abci/types"
-	"github.com/tendermint/tendermint/libs/log"
 	tmproto "github.com/tendermint/tendermint/proto/tendermint/types"
-	dbm "github.com/tendermint/tm-db"
 
 	"github.com/cosmos/cosmos-sdk/crypto/keys/secp256k1"
-	"github.com/cosmos/cosmos-sdk/store"
 	"github.com/cosmos/cosmos-sdk/tests/mocks"
+	"github.com/cosmos/cosmos-sdk/testutil"
 	"github.com/cosmos/cosmos-sdk/types"
 )
 
@@ -26,23 +24,14 @@ func TestContextTestSuite(t *testing.T) {
 	suite.Run(t, new(contextTestSuite))
 }
 
-func (s *contextTestSuite) defaultContext(key types.StoreKey) types.Context {
-	db := dbm.NewMemDB()
-	cms := store.NewCommitMultiStore(db)
-	cms.MountStoreWithDB(key, types.StoreTypeIAVL, db)
-	s.Require().NoError(cms.LoadLatestVersion())
-	ctx := types.NewContext(cms, tmproto.Header{}, false, log.NewNopLogger())
-	return ctx
-}
-
 func (s *contextTestSuite) TestCacheContext() {
 	key := types.NewKVStoreKey(s.T().Name() + "_TestCacheContext")
 	k1 := []byte("hello")
 	v1 := []byte("world")
 	k2 := []byte("key")
 	v2 := []byte("value")
 
-	ctx := s.defaultContext(key)
+	ctx := testutil.DefaultContext(key, types.NewTransientStoreKey("transient_"+s.T().Name()))
 	store := ctx.KVStore(key)
 	store.Set(k1, v1)
 	s.Require().Equal(v1, store.Get(k1))
@@ -64,7 +53,7 @@ func (s *contextTestSuite) TestCacheContext() {
 
 func (s *contextTestSuite) TestLogContext() {
 	key := types.NewKVStoreKey(s.T().Name())
-	ctx := s.defaultContext(key)
+	ctx := testutil.DefaultContext(key, types.NewTransientStoreKey("transient_"+s.T().Name()))
 	ctrl := gomock.NewController(s.T())
 	s.T().Cleanup(ctrl.Finish)
 

types/module/configurator.go

@@ -1,6 +1,11 @@
 package module
 
-import "github.com/gogo/protobuf/grpc"
+import (
+	"github.com/gogo/protobuf/grpc"
+
+	sdk "github.com/cosmos/cosmos-sdk/types"
+	sdkerrors "github.com/cosmos/cosmos-sdk/types/errors"
+)
 
 // Configurator provides the hooks to allow modules to configure and register
 // their services in the RegisterServices method. It is designed to eventually
@@ -15,16 +20,34 @@ type Configurator interface {
 	// QueryServer returns a grpc.Server instance which allows registering services
 	// that will be exposed as gRPC services as well as ABCI query handlers.
 	QueryServer() grpc.Server
+
+	// RegisterMigration registers an in-place store migration for a module. The
+	// handler is a migration script to perform in-place migrations from version
+	// `forVersion` to version `forVersion+1`.
+	//
+	// EACH TIME a module's ConsensusVersion increments, a new migration MUST
+	// be registered using this function. If a migration handler is missing for
+	// a particular function, the upgrade logic (see RunMigrations function)
+	// will panic. If the ConsensusVersion bump does not introduce any store
+	// changes, then a no-op function must be registered here.
+	RegisterMigration(moduleName string, forVersion uint64, handler MigrationHandler) error
 }
 
 type configurator struct {
 	msgServer   grpc.Server
 	queryServer grpc.Server
+
+	// migrations is a map of moduleName -> forVersion -> migration script handler
+	migrations map[string]map[uint64]MigrationHandler
 }
 
 // NewConfigurator returns a new Configurator instance
 func NewConfigurator(msgServer grpc.Server, queryServer grpc.Server) Configurator {
-	return configurator{msgServer: msgServer, queryServer: queryServer}
+	return configurator{
+		msgServer:   msgServer,
+		queryServer: queryServer,
+		migrations:  map[string]map[uint64]MigrationHandler{},
+	}
 }
 
 var _ Configurator = configurator{}
@@ -38,3 +61,51 @@ func (c configurator) MsgServer() grpc.Server {
 func (c configurator) QueryServer() grpc.Server {
 	return c.queryServer
 }
+
+// RegisterMigration implements the Configurator.RegisterMigration method
+func (c configurator) RegisterMigration(moduleName string, forVersion uint64, handler MigrationHandler) error {
+	if forVersion == 0 {
+		return sdkerrors.Wrap(sdkerrors.ErrInvalidVersion, "module migration versions should start at 1")
+	}
+
+	if c.migrations[moduleName] == nil {
+		c.migrations[moduleName] = map[uint64]MigrationHandler{}
+	}
+
+	if c.migrations[moduleName][forVersion] != nil {
+		return sdkerrors.Wrapf(sdkerrors.ErrLogic, "another migration for module %s and version %d already exists", moduleName, forVersion)
+	}
+
+	c.migrations[moduleName][forVersion] = handler
+
+	return nil
+}
+
+// runModuleMigrations runs all in-place store migrations for one given module from a
+// version to another version.
+func (c configurator) runModuleMigrations(ctx sdk.Context, moduleName string, fromVersion, toVersion uint64) error {
+	// No-op if toVersion is the initial version.
+	if toVersion <= 1 {
+		return nil
+	}
+
+	moduleMigrationsMap, found := c.migrations[moduleName]
+	if !found {
+		return sdkerrors.Wrapf(sdkerrors.ErrNotFound, "no migrations found for module %s", moduleName)
+	}
+
+	// Run in-place migrations for the module sequentially until toVersion.
+	for i := fromVersion; i < toVersion; i++ {
+		migrateFn, found := moduleMigrationsMap[i]
+		if !found {
+			return sdkerrors.Wrapf(sdkerrors.ErrNotFound, "no migration found for module %s from version %d to version %d", moduleName, i, i+1)
+		}
+
+		err := migrateFn(ctx)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}

types/module/module.go

@@ -40,6 +40,7 @@ import (
 	"github.com/cosmos/cosmos-sdk/codec"
 	codectypes "github.com/cosmos/cosmos-sdk/codec/types"
 	sdk "github.com/cosmos/cosmos-sdk/types"
+	sdkerrors "github.com/cosmos/cosmos-sdk/types/errors"
 )
 
 //__________________________________________________________________________________________
@@ -174,6 +175,12 @@ type AppModule interface {
 	// RegisterServices allows a module to register services
 	RegisterServices(Configurator)
 
+	// ConsensusVersion is a sequence number for state-breaking change of the
+	// module. It should be incremented on each consensus-breaking change
+	// introduced by the module. To avoid wrong/empty versions, the initial version
+	// should be set to 1.
+	ConsensusVersion() uint64
+
 	// ABCI
 	BeginBlock(sdk.Context, abci.RequestBeginBlock)
 	EndBlock(sdk.Context, abci.RequestEndBlock) []abci.ValidatorUpdate
@@ -208,6 +215,9 @@ func (gam GenesisOnlyAppModule) LegacyQuerierHandler(*codec.LegacyAmino) sdk.Que
 // RegisterServices registers all services.
 func (gam GenesisOnlyAppModule) RegisterServices(Configurator) {}
 
+// ConsensusVersion implements AppModule/ConsensusVersion.
+func (gam GenesisOnlyAppModule) ConsensusVersion() uint64 { return 1 }
+
 // BeginBlock returns an empty module begin-block
 func (gam GenesisOnlyAppModule) BeginBlock(ctx sdk.Context, req abci.RequestBeginBlock) {}
 
@@ -328,6 +338,30 @@ func (m *Manager) ExportGenesis(ctx sdk.Context, cdc codec.JSONMarshaler) map[st
 	return genesisData
 }
 
+// MigrationHandler is the migration function that each module registers.
+type MigrationHandler func(store sdk.Context) error
+
+// MigrationMap is a map of moduleName -> version, where version denotes the
+// version from which we should perform the migration for each module.
+type MigrationMap map[string]uint64
+
+// RunMigrations performs in-place store migrations for all modules.
+func (m Manager) RunMigrations(ctx sdk.Context, cfg Configurator, migrateFromVersions MigrationMap) error {
+	c, ok := cfg.(configurator)
+	if !ok {
+		return sdkerrors.Wrapf(sdkerrors.ErrInvalidType, "expected %T, got %T", configurator{}, cfg)
+	}
+
+	for moduleName, module := range m.Modules {
+		err := c.runModuleMigrations(ctx, moduleName, migrateFromVersions[moduleName], module.ConsensusVersion())
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
 // BeginBlock performs begin block functionality for all modules. It creates a
 // child context with an event manager to aggregate events emitted from all
 // modules.