Migrating to RTK 2.0 and Redux 5.0
- What's changed in Redux Toolkit 2.0, Redux core 5.0, Reselect 5.0, and Redux Thunk 3.0, including breaking changes and new features
Introduction
Redux Toolkit has been available since 2019, and today it's the standard way to write Redux apps. We've gone 4+ years without any breaking changes. Now, RTK 2.0 gives us a chance to modernize the packaging, clean up deprecated options, and tighten up some edge cases.
Redux Toolkit 2.0 is accompanied by major versions of all the other Redux packages: Redux core 5.0, React-Redux 9.0, Reselect 5.0, and Redux Thunk 3.0.
This page lists known potentially breaking changes in each of those packages, as well as new features in Redux Toolkit 2.0. As a reminder, you should not need to actually install or use the core redux
package directly - RTK wraps that, and re-exports all methods and types.
In practice, most of the "breaking" changes should not have an actual effect on end users, and we expect that many projects can just update the package versions with very few code changes needed.
The changes most likely to need app code updates are:
- Object syntax removed for
createReducer
andcreateSlice.extraReducers
configureStore.middleware
must be a callbackMiddleware
type changed - Middlewareaction
andnext
are typed asunknown
Packaging Changes (all)
We've made updates to the build packaging for all of the Redux-related libraries. These are technically "breaking", but should be transparent to end users, and actually enable better support for scenarios such as using Redux via ESM files under Node.
Addition of exports
field in package.json
We've migrated the package definitions to include the exports
field for defining which artifacts to load, with a modern ESM build as the primary artifact (with CJS still included for compatibility purposes).
We've done local testing of the package, but we ask the community to try out this in your own projects and report any breakages you find!
Build Artifact Modernization
We've updated the build output in several ways:
- Build output is no longer transpiled! Instead we target modern JS syntax (ES2020)
- Moved all build artifacts to live under
./dist/
, instead of separate top-level folders - The lowest Typescript version we test against is now TS 4.7.
Dropping UMD builds
Redux has always shipped with UMD build artifacts. These are primarily meant for direct import as script tags, such as in a CodePen or a no-bundler build environment.
For now, we're dropping those build artifacts from the published package, on the grounds that the use cases seem pretty rare today.
We do have a browser-ready ESM build artifact included at dist/$PACKAGE_NAME.browser.mjs
, which can be loaded via a script tag that points to that file on Unpkg.
If you have strong use cases for us continuing to include UMD build artifacts, please let us know!
Breaking Changes
Core
Action types must be strings
We've always specifically told our users that actions and state must be serializable, and that action.type
should be a string. This is both to ensure that actions are serializable, and to help provide a readable action history in the Redux DevTools.
store.dispatch(action)
now specifically enforces that action.type
must be a string and will throw an error if not, in the same way it throws an error if the action is not a plain object.
In practice, this was already true 99.99% of the time and shouldn't have any effect on users (especially those using Redux Toolkit and createSlice
), but there may be some legacy Redux codebases that opted to use Symbols as action types.
createStore
Deprecation
In Redux 4.2.0, we marked the original createStore
method as @deprecated
. Strictly speaking, this is not a breaking change, nor is it new in 5.0, but we're documenting it here for completeness.
This deprecation is solely a visual indicator that is meant to encourage users to migrate their apps from legacy Redux patterns to use the modern Redux Toolkit APIs.
The deprecation results in a visual strikethrough when imported and used, like , but with no runtime errors or warnings.createStore
createStore
will continue to work indefinitely, and will not ever be removed. But, today we want all Redux users to be using Redux Toolkit for all of their Redux logic.
To fix this, there are three options:
- Follow our strong suggestion to switch over to Redux Toolkit and
configureStore
- Do nothing. It's just a visual strikethrough, and it doesn't affect how your code behaves. Ignore it.
- Switch to using the
legacy_createStore
API that is now exported, which is the exact same function but with no@deprecated
tag. The simplest option is to do an aliased import rename, likeimport { legacy_createStore as createStore } from 'redux'
Typescript rewrite
In 2019, we began a community-powered conversion of the Redux codebase to TypeScript. The original effort was discussed in #3500: Port to TypeScript, and the work was integrated in PR #3536: Convert to TypeScript.
However, the TS-converted code sat around in the repo for several years, unused and unpublished, due to concerns about possible compatibility issues with the existing ecosystem (as well as general inertia on our part).
Redux core v5 is now built from that TS-converted source code. In theory, this should be almost identical in both runtime behavior and types to the 4.x build, but it's very likely that some of the changes may cause types issues.
Please report any unexpected compatibility issues on Github!
AnyAction
deprecated in favour of UnknownAction
The Redux TS types have always exported an AnyAction
type, which is defined to have {type: string}
and treat any other field as any
. This makes it easy to write uses like console.log(action.whatever)
, but unfortunately does not provide any meaningful type safety.
We now export an UnknownAction
type, which treats all fields other than action.type
as unknown
. This encourages users to write type guards that check the action object and assert its specific TS type. Inside of those checks, you can access a field with better type safety.
UnknownAction
is now the default any place in the Redux source that expects an action object.
AnyAction
still exists for compatibility, but has been marked as deprecated.
Note that Redux Toolkit's action creators have a .match()
method that acts as a useful type guard:
if (todoAdded.match(someUnknownAction)) {
// action is now typed as a PayloadAction<Todo>
}
You can also use the new isAction
util to check if an unknown value is some kind of action object.
Middleware
type changed - Middleware action
and next
are typed as unknown
Previously, the next
parameter is typed as the D
type parameter passed, and action
is typed as the Action
extracted from the dispatch type. Neither of these are a safe assumption:
next
would be typed to have all of the dispatch extensions, including the ones earlier in the chain that would no longer apply.- Technically it would be mostly safe to type
next
as the default Dispatch implemented by the base redux store, however this would causenext(action)
to error (as we cannot promiseaction
is actually anAction
) - and it wouldn't account for any following middlewares that return anything other than the action they're given when they see a specific action.
- Technically it would be mostly safe to type
action
is not necessarily a known action, it can be literally anything - for example a thunk would be a function with no.type
property (soAnyAction
would be inaccurate)
We've changed next
to be (action: unknown) => unknown
(which is accurate, we have no idea what next
expects or will return), and changed the action
parameter to be unknown
(which as above, is accurate).
In order to safely interact with values or access fields inside of the action
argument, you must first do a type guard check to narrow the type, such as isAction(action)
or someActionCreator.match(action)
.
This new type is incompatible with the v4 Middleware
type, so if a package's middleware is saying it's incompatible, check which version of Redux it's getting its types from! (See overriding dependencies later in this page.)
PreloadedState
type removed in favour of Reducer
generic
We've made tweaks to the TS types to improve type safety and behavior.
First, the Reducer
type now has a PreloadedState
possible generic:
type Reducer<S, A extends Action, PreloadedState = S> = (
state: S | PreloadedState | undefined,
action: A,
) => S
Per the explanation in #4491:
Why the need for this change? When the store is first created by createStore
/configureStore
, the initial state is set to whatever is passed as the preloadedState
argument (or undefined
if nothing is passed). That means that the first time that the reducer is called, it is called with the preloadedState
. After the first call, the reducer is always passed the current state (which is S
).
For most normal reducers, S | undefined
accurately describes what can be passed in for the preloadedState
. However the combineReducers
function allows for a preloaded state of Partial<S> | undefined
.
The solution is to have a separate generic that represents what the reducer accepts for its preloaded state. That way createStore
can then use that generic for its preloadedState
argument.
Previously, this was handled by a $CombinedState
type, but that complicated things and led to some user-reported issues. This removes the need for $CombinedState
altogether.
This change does include some breaking changes, but overall should not have a huge impact on users upgrading in user-land:
- The
Reducer
,ReducersMapObject
, andcreateStore
/configureStore
types/function take an additionalPreloadedState
generic which defaults toS
. - The overloads for
combineReducers
are removed in favor of a single function definition that takes theReducersMapObject
as its generic parameter. Removing the overloads was necessary with these changes, since sometimes it was choosing the wrong overload. - Enhancers that explicitly list the generics for the reducer will need to add the third generic.
Toolkit only
Object syntax for createSlice.extraReducers
and createReducer
removed
RTK's createReducer
API was originally designed to accept a lookup table of action type strings to case reducers, like { "ADD_TODO": (state, action) => {} }
. We later added the "builder callback" form to allow more flexibility in adding "matchers" and a default handler, and did the same for createSlice.extraReducers
.
We have removed the "object" form for both createReducer
and createSlice.extraReducers
in RTK 2.0, as the builder callback form is effectively the same number of lines of code, and works much better with TypeScript.
As an example, this:
const todoAdded = createAction('todos/todoAdded')
createReducer(initialState, {
[todoAdded]: (state, action) => {},
})
createSlice({
name,
initialState,
reducers: {
/* case reducers here */
},
extraReducers: {
[todoAdded]: (state, action) => {},
},
})
should be migrated to:
createReducer(initialState, (builder) => {
builder.addCase(todoAdded, (state, action) => {})
})
createSlice({
name,
initialState,
reducers: {
/* case reducers here */
},
extraReducers: (builder) => {
builder.addCase(todoAdded, (state, action) => {})
},
})
Codemods
To simplify upgrading codebases, we've published a set of codemods that will automatically transform the deprecated "object" syntax into the equivalent "builder" syntax.
The codemods package is available on NPM as @reduxjs/rtk-codemods
. More details are available here.
To run the codemods against your codebase, run npx @reduxjs/rtk-codemods <TRANSFORM NAME> path/of/files/ or/some**/*glob.js.
Examples:
npx @reduxjs/rtk-codemods createReducerBuilder ./src
npx @reduxjs/rtk-codemods createSliceBuilder ./packages/my-app/**/*.ts
We also recommend re-running Prettier on the codebase before committing the changes.
These codemods should work, but we would greatly appreciate feedback from more real-world codebases!
configureStore.middleware
must be a callback
Since the beginning, configureStore
has accepted a direct array value as the middleware
option. However, providing an array directly prevents configureStore
from calling getDefaultMiddleware()
. So, middleware: [myMiddleware]
means there is no thunk middleware added (or any of the dev-mode checks).
This is a footgun, and we've had numerous users accidentally do this and cause their apps to fail because the default middleware never got configured.
As a result, we've now made the middleware
only accept the callback form. If for some reason you still want to replace all of the built-in middleware, do so by returning an array from the callback:
const store = configureStore({
reducer,
middleware: (getDefaultMiddleware) => {
// WARNING: this means that _none_ of the default middleware are added!
return [myMiddleware]
// or for TS users, use:
// return new Tuple(myMiddleware)
},
})
But note that we consistently recommend not replacing the default middleware entirely, and that you should use return getDefaultMiddleware().concat(myMiddleware)
.
configureStore.enhancers
must be a callback
Similarly to configureStore.middleware
, the enhancers
field must also be a callback, for the same reasons.
The callback will receive a getDefaultEnhancers
function that can be used to customise the batching enhancer that's now included by default.
For example:
const store = configureStore({
reducer,
enhancers: (getDefaultEnhancers) => {
return getDefaultEnhancers({
autoBatch: { type: 'tick' },
}).concat(myEnhancer)
},
})
It's important to note that the result of getDefaultEnhancers
will also contain the middleware enhancer created with any configured/default middleware. To help prevent mistakes, configureStore
will log an error to console if middleware was provided and the middleware enhancer wasn't included in the callback result.
const store = configureStore({
reducer,
enhancers: (getDefaultEnhancers) => {
return [myEnhancer] // we've lost the middleware here
// instead:
return getDefaultEnhancers().concat(myEnhancer)
},
})
Standalone getDefaultMiddleware
and getType
removed
The standalone version of getDefaultMiddleware
has been deprecated since v1.6.1, and has now been removed. Use the function passed to the middleware
callback instead, which has the correct types.
We have also removed the getType
export, which was used to extract a type string from action creators made with createAction
. Instead, use the static property actionCreator.type
.
RTK Query behavior changes
We've had a number of reports where RTK Query had issues around usage of dispatch(endpoint.initiate(arg, {subscription: false}))
. There were also reports that multiple triggered lazy queries were resolving the promises at the wrong time. Both of these had the same underlying issue, which was that RTKQ wasn't tracking cache entries in these cases (intentionally). We've reworked the logic to always track cache entries (and remove them as needed), which should resolve those behavior issues.
We also have had issues raised about trying to run multiple mutations in a row and how tag invalidation behaves. RTKQ now has internal logic to delay tag invalidation briefly, to allow multiple invalidations to get handled together. This is controlled by a new invalidationBehavior: 'immediate' | 'delayed'
flag on createApi
. The new default behavior is 'delayed'
. Set it to 'immediate'
to revert to the behavior in RTK 1.9.
In RTK 1.9, we reworked RTK Query's internals to keep most of the subscription status inside the RTKQ middleware. The values are still synced to the Redux store state, but this is primarily for display by the Redux DevTools "RTK Query" panel. Related to the cache entry changes above, we've optimized how often those values get synced to the Redux state for perf.
reactHooksModule
custom hook configuration
Previously, custom versions of React Redux's hooks (useSelector
, useDispatch
, and useStore
) could be passed separately to reactHooksModule
, usually to enable using a different context to the default ReactReduxContext
.
In practicality, the react hooks module needs all three of these hooks to be provided, and it became an easy mistake to only pass useSelector
and useDispatch
, without useStore
.
The module has now moved all three of these under the same configuration key, and will check that all three are provided if the key is present.
// previously
const customCreateApi = buildCreateApi(
coreModule(),
reactHooksModule({
useDispatch: createDispatchHook(MyContext),
useSelector: createSelectorHook(MyContext),
useStore: createStoreHook(MyContext),
}),
)
// now
const customCreateApi = buildCreateApi(
coreModule(),
reactHooksModule({
hooks: {
useDispatch: createDispatchHook(MyContext),
useSelector: createSelectorHook(MyContext),
useStore: createStoreHook(MyContext),
},
}),
)
Error message extraction
Redux 4.1.0 optimized its bundle size by extracting error message strings out of production builds, based on React's approach. We've applied the same technique to RTK. This saves about 1000 bytes from prod bundles (actual benefits will depend on which imports are being used).
configureStore
field order for middleware
matters
If you are passing both the middleware
and enhancers
fields to configureStore
, the middleware
field must come first in order for internal TS inference to work properly.
Non-default middleware/enhancers must use Tuple
We've seen many cases where users passing the middleware
parameter to configureStore have tried spreading the array returned by getDefaultMiddleware()
, or passed an alternate plain array. This unfortunately loses the exact TS types from the individual middleware, and often causes TS problems down the road (such as dispatch
being typed as Dispatch<AnyAction>
and not knowing about thunks).
getDefaultMiddleware()
already used an internal MiddlewareArray
class, an Array
subclass that had strongly typed .concat/prepend()
methods to correctly capture and retain the middleware types.
We've renamed that type to Tuple
, and configureStore
's TS types now require that you must use Tuple
if you want to pass your own array of middleware:
import { configureStore, Tuple } from '@reduxjs/toolkit'
configureStore({
reducer: rootReducer,
middleware: (getDefaultMiddleware) => new Tuple(additionalMiddleware, logger),
})
(Note that this has no effect if you're using RTK with plain JS, and you could still pass a plain array here.)
This same restriction applies to the enhancers
field.
Entity adapter type updates
createEntityAdapter
now has an Id
generic argument, which will be used to strongly type the item IDs anywhere those are exposed. Previously, the ID field type was always string | number
. TS will now try to infer the exact type from either the .id
field of your entity type, or the selectId
return type. You could also fall back to passing that generic type directly. If you use the EntityState<Data, Id>
type directly, you must supply both generic arguments!
The .entities
lookup table is now defined to use a standard TS Record<Id, MyEntityType>
, which assumes that each item lookup exists by default. Previously, it used a Dictionary<MyEntityType>
type, which assumed the result was MyEntityType | undefined
. The Dictionary
type has been removed.
If you prefer to assume that the lookups might be undefined, use TypeScript's noUncheckedIndexedAccess
configuration option to control that.
Reselect
createSelector
Uses weakMapMemoize
As Default Memoizer
createSelector
now uses a new default memoization function called weakMapMemoize
. This memoizer offers an effectively infinite cache size, which should simplify usage with varying arguments, but relies exclusively on reference comparisons.
If you need to customize equality comparisons, customize createSelector
to use the original lruMemoize
method instead:
createSelector(inputs, resultFn, {
memoize: lruMemoize,
memoizeOptions: { equalityCheck: yourEqualityFunction },
})
defaultMemoize
Renamed to lruMemoize
Since the original defaultMemoize
function is no longer actually the default, we've renamed it to lruMemoize
for clarity. This only matters if you specifically imported it into your app to customize selectors.
createSelector
Dev-Mode Checks
createSelector
now does checks in development mode for common mistakes, like input selectors that always return new references, or result functions that immediately return their argument. These checks can be customized at selector creation or globally.
This is important, as an input selector returning a materially different result with the same parameters means that the output selector will never memoize correctly and be run unnecessarily, thus (potentially) creating a new result and causing rerenders.
const addNumbers = createSelector(
// this input selector will always return a new reference when run
// so cache will never be used
(a, b) => ({ a, b }),
({ a, b }) => ({ total: a + b }),
)
// instead, you should have an input selector for each stable piece of data
const addNumbersStable = createSelector(
(a, b) => a,
(a, b) => b,
(a, b) => ({
total: a + b,
}),
)
This is done the first time the selector is called, unless configured otherwise. More details are available in the Reselect docs on dev-mode checks.
Note that while RTK re-exports createSelector
, it intentionally does not re-export the function to configure this check globally - if you wish to do so, you should instead depend on reselect
directly and import it yourself.
ParametricSelector
Types Removed
The ParametricSelector
and OutputParametricSelector
types have been removed. Use Selector
and OutputSelector
instead.
React-Redux
Requires React 18
React-Redux v7 and v8 worked with all versions of React that supported hooks (16.8+, 17, and 18). v8 switched from internal subscription management to React's new useSyncExternalStore
hook, but used the "shim" implementation to provide support for React 16.8 and 17, which did not have that hook built in.
React-Redux v9 switches to requiring React 18, and does not support React 16 or 17. This allows us to drop the shim and save a small bit of bundle size.
Custom context typing
React Redux supports creating hooks
(and connect
) with a custom context, but typing this has been fairly non-standard. The pre-v9 types required Context<ReactReduxContextValue>
, but the context default value was usually initialised with null
(as the hooks use this to make sure they actually have a provided context). This, in "best" cases, would result in something like the below:
import { createContext } from 'react'
import {
ReactReduxContextValue,
createDispatchHook,
createSelectorHook,
createStoreHook,
} from 'react-redux'
import { AppStore, RootState, AppDispatch } from './store'
const context = createContext<ReactReduxContextValue>(null as any)
export const useStore = createStoreHook(context).withTypes<AppStore>()
export const useDispatch = createDispatchHook(context).withTypes<AppDispatch>()
export const useSelector = createSelectorHook(context).withTypes<RootState>()
In v9, the types now match the runtime behavior. The context is typed to hold ReactReduxContextValue | null
, and the hooks know that if they receive null
they'll throw an error so it doesn't affect the return type.
The above example now becomes:
import { createContext } from 'react'
import {
ReactReduxContextValue,
createDispatchHook,
createSelectorHook,
createStoreHook,
} from 'react-redux'
import { AppStore, RootState, AppDispatch } from './store'
const context = createContext<ReactReduxContextValue | null>(null)
export const useStore = createStoreHook(context).withTypes<AppStore>()
export const useDispatch = createDispatchHook(context).withTypes<AppDispatch>()
export const useSelector = createSelectorHook(context).withTypes<RootState>()
Redux Thunk
Thunk Uses Named Exports
The redux-thunk
package previously used a single default export that was the middleware, with an attached field named withExtraArgument
that allowed customization.
The default export has been removed. There are now two named exports: thunk
(the basic middleware) and withExtraArgument
.
If you are using Redux Toolkit, this should have no effect, as RTK already handles this inside of configureStore
.
New Features
These features are new in Redux Toolkit 2.0, and help cover additional use cases that we've seen users ask for in the ecosystem.
combineSlices
API with slice reducer injection for code-splitting
The Redux core has always included combineReducers
, which takes an object full of "slice reducer" functions and generates a reducer that calls those slice reducers. RTK's createSlice
generates slice reducers + associated action creators, and we've taught the pattern of exporting individual action creators as named exports and the slice reducer as a default export. Meanwhile, we've never had official support for lazy-loading reducers, although we've had sample code for some "reducer injection" patterns in our docs.
This release includes a new combineSlices
API that is designed to enable lazy-loading of reducers at runtime. It accepts individual slices or an object full of slices as arguments, and automatically calls combineReducers
using the sliceObject.name
field as the key for each state field. The generated reducer function has an additional .inject()
method attached that can be used to dynamically inject additional slices at runtime. It also includes a .withLazyLoadedSlices()
method that can be used to generate TS types for reducers that will be added later. See #2776 for the original discussion around this idea.
For now, we are not building this into configureStore
, so you'll need to call const rootReducer = combineSlices(.....)
yourself and pass that to configureStore({reducer: rootReducer})
.
Basic usage: a mixture of slices and standalone reducers passed to combineSlices
const stringSlice = createSlice({
name: 'string',
initialState: '',
reducers: {},
})
const numberSlice = createSlice({
name: 'number',
initialState: 0,
reducers: {},
})
const booleanReducer = createReducer(false, () => {})
const api = createApi(/* */)
const combinedReducer = combineSlices(
stringSlice,
{
num: numberSlice.reducer,
boolean: booleanReducer,
},
api,
)
expect(combinedReducer(undefined, dummyAction())).toEqual({
string: stringSlice.getInitialState(),
num: numberSlice.getInitialState(),
boolean: booleanReducer.getInitialState(),
api: api.reducer.getInitialState(),
})
Basic slice reducer injection
// Create a reducer with a TS type that knows `numberSlice` will be injected
const combinedReducer =
combineSlices(stringSlice).withLazyLoadedSlices<
WithSlice<typeof numberSlice>
>()
// `state.number` doesn't exist initially
expect(combinedReducer(undefined, dummyAction()).number).toBe(undefined)
// Create a version of the reducer with `numberSlice` injected (mainly useful for types)
const injectedReducer = combinedReducer.inject(numberSlice)
// `state.number` now exists, and injectedReducer's type no longer marks it as optional
expect(injectedReducer(undefined, dummyAction()).number).toBe(
numberSlice.getInitialState(),
)
// original reducer has also been changed (type is still optional)
expect(combinedReducer(undefined, dummyAction()).number).toBe(
numberSlice.getInitialState(),
)
selectors
field in createSlice
The existing createSlice
API now has support for defining selectors
directly as part of the slice. By default, these will be generated with the assumption that the slice is mounted in the root state using slice.name
as the field, such as name: "todos"
-> rootState.todos
. Additionally, there's now a slice.selectSlice
method that does that default root state lookup.
You can call sliceObject.getSelectors(selectSliceState)
to generate the selectors with an alternate location, similar to how entityAdapter.getSelectors()
works.
const slice = createSlice({
name: 'counter',
initialState: 42,
reducers: {},
selectors: {
selectSlice: (state) => state,
selectMultiple: (state, multiplier: number) => state * multiplier,
},
})
// Basic usage
const testState = {
[slice.name]: slice.getInitialState(),
}
const { selectSlice, selectMultiple } = slice.selectors
expect(selectSlice(testState)).toBe(slice.getInitialState())
expect(selectMultiple(testState, 2)).toBe(slice.getInitialState() * 2)
// Usage with the slice reducer mounted under a different key
const customState = {
number: slice.getInitialState(),
}
const { selectSlice, selectMultiple } = slice.getSelectors(
(state: typeof customState) => state.number,
)
expect(selectSlice(customState)).toBe(slice.getInitialState())
expect(selectMultiple(customState, 2)).toBe(slice.getInitialState() * 2)
createSlice.reducers
callback syntax and thunk support
One of the oldest feature requests we've had is the ability to declare thunks directly inside of createSlice
. Until now, you've always had to declare them separately, give the thunk a string action prefix, and handle the actions via createSlice.extraReducers
:
// Declare the thunk separately
const fetchUserById = createAsyncThunk(
'users/fetchByIdStatus',
async (userId: number, thunkAPI) => {
const response = await userAPI.fetchById(userId)
return response.data
},
)
const usersSlice = createSlice({
name: 'users',
initialState,
reducers: {
// standard reducer logic, with auto-generated action types per reducer
},
extraReducers: (builder) => {
// Add reducers for additional action types here, and handle loading state as needed
builder.addCase(fetchUserById.fulfilled, (state, action) => {
state.entities.push(action.payload)
})
},
})
Many users have told us that this separation feels awkward.
We've wanted to include a way to define thunks directly inside of createSlice
, and have played around with various prototypes. There were always two major blocking issues, and a secondary concern:
- It wasn't clear what the syntax for declaring a thunk inside should look like.
- Thunks have access to
getState
anddispatch
, but theRootState
andAppDispatch
types are normally inferred from the store, which in turn infers it from the slice state types. Declaring thunks insidecreateSlice
would cause circular type inference errors, as the store needs the slice types but the slice needs the store types. We weren't willing to ship an API that would work okay for our JS users but not for our TS users, especially since we want people to use TS with RTK. - You can't do synchronous conditional imports in ES modules, and there's no good way to make the
createAsyncThunk
import optional. EithercreateSlice
always depends on it (and adds that to the bundle size), or it can't usecreateAsyncThunk
at all.
We've settled on these compromises:
- In order to create async thunks with
createSlice
, you specifically need to set up a custom version ofcreateSlice
that has access tocreateAsyncThunk
. - You can declare thunks inside of
createSlice.reducers
, by using a "creator callback" syntax for thereducers
field that is similar to thebuild
callback syntax in RTK Query'screateApi
(using typed functions to create fields in an object). Doing this does look a bit different than the existing "object" syntax for thereducers
field, but is still fairly similar. - You can customize some of the types for thunks inside of
createSlice
, but you cannot customize thestate
ordispatch
types. If those are needed, you can manually do anas
cast, likegetState() as RootState
.
In practice, we hope these are reasonable tradeoffs. Creating thunks inside of createSlice
has been widely asked for, so we think it's an API that will see usage. If the TS customization options are a limitation, you can still declare thunks outside of createSlice
as always, and most async thunks don't need dispatch
or getState
- they just fetch data and return. And finally, setting up a custom createSlice
allows you to opt into createAsyncThunk
being included in your bundle size (though it may already be included if used directly or as part of RTK Query - in either of these cases there's no additional bundle size).
Here's what the new callback syntax looks like:
const createAppSlice = buildCreateSlice({
creators: { asyncThunk: asyncThunkCreator },
})
const todosSlice = createAppSlice({
name: 'todos',
initialState: {
loading: false,
todos: [],
error: null,
} as TodoState,
reducers: (create) => ({
// A normal "case reducer", same as always
deleteTodo: create.reducer((state, action: PayloadAction<number>) => {
state.todos.splice(action.payload, 1)
}),
// A case reducer with a "prepare callback" to customize the action
addTodo: create.preparedReducer(
(text: string) => {
const id = nanoid()
return { payload: { id, text } }
},
// action type is inferred from prepare callback
(state, action) => {
state.todos.push(action.payload)
},
),
// An async thunk
fetchTodo: create.asyncThunk(
// Async payload function as the first argument
async (id: string, thunkApi) => {
const res = await fetch(`myApi/todos?id=${id}`)
return (await res.json()) as Item
},
// An object containing `{pending?, rejected?, fulfilled?, settled?, options?}` second
{
pending: (state) => {
state.loading = true
},
rejected: (state, action) => {
state.error = action.payload ?? action.error
},
fulfilled: (state, action) => {
state.todos.push(action.payload)
},
// settled is called for both rejected and fulfilled actions
settled: (state, action) => {
state.loading = false
},
},
),
}),
})
// `addTodo` and `deleteTodo` are normal action creators.
// `fetchTodo` is the async thunk
export const { addTodo, deleteTodo, fetchTodo } = todosSlice.actions
Codemod
Using the new callback syntax is entirely optional (the object syntax is still standard), but an existing slice would need to be converted before it can take advantage of the new capabilities this syntax provides. To make this easier, a codemod is provided.
npx @reduxjs/rtk-codemods createSliceReducerBuilder ./src/features/todos/slice.ts
"Dynamic middleware" middleware
A Redux store's middleware pipeline is fixed at store creation time and can't be changed later. We have seen ecosystem libraries that tried to allow dynamically adding and removing middleware, potentially useful for things like code splitting.
This is a relatively niche use case, but we've built our own version of a "dynamic middleware" middleware. Add it to the Redux store at setup time, and it lets you add middleware later at runtime. It also comes with a React hook integration that will automatically add a middleware to the store and return the updated dispatch method..
import { createDynamicMiddleware, configureStore } from '@reduxjs/toolkit'
const dynamicMiddleware = createDynamicMiddleware()
const store = configureStore({
reducer: {
todos: todosReducer,
},
middleware: (getDefaultMiddleware) =>
getDefaultMiddleware().prepend(dynamicMiddleware.middleware),
})
// later
dynamicMiddleware.addMiddleware(someOtherMiddleware)
configureStore
adds autoBatchEnhancer
by default
In v1.9.0, we added a new autoBatchEnhancer
that delays notifying subscribers briefly when multiple "low-priority" actions are dispatched in a row. This improves perf, as UI updates are typically the most expensive part of the update process. RTK Query marks most of its own internal actions as "low-pri" by default, but you have to have the autoBatchEnhancer
added to the store to benefit from that.
We've updated configureStore
to add the autoBatchEnhancer
to the store setup by default, so that users can benefit from the improved perf without needing to manually tweak the store config themselves.
entityAdapter.getSelectors
accepts a createSelector
function
entityAdapter.getSelectors()
now accepts an options object as its second argument. This allows you to pass in your own preferred createSelector
method, which will be used to memoize the generated selectors. This could be useful if you want to use one of Reselect's new alternate memoizers, or some other memoization library with an equivalent signature.
Immer 10.0
Immer 10.0 is now final, and has several major improvements and updates:
- Much faster update perf
- Much smaller bundle size
- Better ESM/CJS package formatting
- No default export
- No ES5 fallback
We've updated RTK to depend on the final Immer 10.0 release.
Next.js Setup Guide
We now have a docs page that covers how to set up Redux properly with Next.js. We've seen a lot of questions around using Redux, Next, and the App Router together, and this guide should help provide advice.
(At this time, the Next.js with-redux
example is still showing outdated patterns - we're going to file a PR shortly to update that to match our docs guide.)
Overriding dependencies
It will take a while for packages to update their peer dependencies to allow for Redux core 5.0, and in the meantime changes like the Middleware type will result in perceived incompatibilities.
It's likely that most libraries will not actually have any practices that are incompatible with 5.0, but due to the peer dependency on 4.0 they end up pulling in old type declarations.
This can be solved by manually overriding the dependency resolution, which is supported by both npm
and yarn
.
npm
- overrides
NPM supports this through an overrides
field in your package.json
. You can override the dependency for a specific package, or make sure that every package that pulls in Redux receives the same version.
{
"overrides": {
"redux-persist": {
"redux": "^5.0.0"
}
}
}
{
"overrides": {
"redux": "^5.0.0"
}
}
yarn
- resolutions
Yarn supports this through a resolutions
field in your package.json
. Just like with NPM, you can override the dependency for a specific package, or make sure that every package that pulls in Redux receives the same version.
{
"resolutions": {
"redux-persist/redux": "^5.0.0"
}
}
{
"resolutions": {
"redux": "^5.0.0"
}
}
Recommendations
Based on changes in 2.0 and previous versions, there have been some shifts in thinking that are good to know about, if non-essential.
Alternatives to actionCreator.toString()
As part of RTK's original API, action creators made with createAction
have a custom toString()
override that returns the action type.
This was primarily useful for the (now removed) object syntax for createReducer
:
const todoAdded = createAction<Todo>('todos/todoAdded')
createReducer(initialState, {
[todoAdded]: (state, action) => {}, // toString called here, 'todos/todoAdded'
})
While this was convenient (and other libraries in the Redux ecosystem such as redux-saga
and redux-observable
have supported this to various capacities), it didn't play well with Typescript and was generally a bit too "magic".
const test = todoAdded.toString()
// ^? typed as string, rather than specific action type
Over time, the action creator also gained a static type
property and match
method which were more explicit and worked better with Typescript.
const test = todoAdded.type
// ^? 'todos/todoAdded'
// acts as a type predicate
if (todoAdded.match(unknownAction)) {
unknownAction.payload
// ^? now typed as PayloadAction<Todo>
}
For compatibility, this override is still in place, but we encourage considering using either of the static properties for more understandable code.
For example, with redux-observable
:
// before (works in runtime, will not filter types properly)
const epic = (action$: Observable<Action>) =>
action$.pipe(
ofType(todoAdded),
map((action) => action),
// ^? still Action<any>
)
// consider (better type filtering)
const epic = (action$: Observable<Action>) =>
action$.pipe(
filter(todoAdded.match),
map((action) => action),
// ^? now PayloadAction<Todo>
)
With redux-saga
:
// before (still works)
yield takeEvery(todoAdded, saga)
// consider
yield takeEvery(todoAdded.match, saga)
// or
yield takeEvery(todoAdded.type, saga)
Future plans
Custom slice reducer creators
With the addition of the callback syntax for createSlice, the suggestion was made to enable custom slice reducer creators. These creators would be able to:
- Modify reducer behavior by adding case or matcher reducers
- Attach actions (or any other useful functions) to
slice.actions
- Attach provided case reducers to
slice.caseReducers
The creator would need to first return a "definition" shape when createSlice
is first called, which it then handles by adding any necessary reducers and/or actions.
An API for this is not set in stone, but the existing create.asyncThunk
creator implemented with a potential API could look like:
const asyncThunkCreator = {
type: ReducerType.asyncThunk,
define(payloadCreator, config) {
return {
type: ReducerType.asyncThunk, // needs to match reducer type, so correct handler can be called
payloadCreator,
...config,
}
},
handle(
{
// the key the reducer was defined under
reducerName,
// the autogenerated action type, i.e. `${slice.name}/${reducerName}`
type,
},
// the definition from define()
definition,
// methods to modify slice
context,
) {
const { payloadCreator, options, pending, fulfilled, rejected, settled } =
definition
const asyncThunk = createAsyncThunk(type, payloadCreator, options)
if (pending) context.addCase(asyncThunk.pending, pending)
if (fulfilled) context.addCase(asyncThunk.fulfilled, fulfilled)
if (rejected) context.addCase(asyncThunk.rejected, rejected)
if (settled) context.addMatcher(asyncThunk.settled, settled)
context.exposeAction(reducerName, asyncThunk)
context.exposeCaseReducer(reducerName, {
pending: pending || noop,
fulfilled: fulfilled || noop,
rejected: rejected || noop,
settled: settled || noop,
})
},
}
const createSlice = buildCreateSlice({
creators: {
asyncThunk: asyncThunkCreator,
},
})
We're not sure how many people/libraries would actually make use of this though, so any feedback over on the Github issue is welcome!
createSlice.selector
selector factories
There have been some concerns raised internally about whether createSlice.selectors
supports memoized selectors sufficiently. You can provide a memoized selector to your createSlice.selectors
configuration, but you're stuck with that one instance.
const todoSlice = createSlice({
name: 'todos',
initialState: {
todos: [] as Todo[],
},
reducers: {},
selectors: {
selectTodosByAuthor = createSelector(
(state: TodoState) => state.todos,
(state: TodoState, author: string) => author,
(todos, author) => todos.filter((todo) => todo.author === author),
),
},
})
export const { selectTodosByAuthor } = todoSlice.selectors
With createSelector
's default cache size of 1, this can cause caching issues if called in multiple components with different arguments. One typical solution for this (without createSlice
) is a selector factory:
export const makeSelectTodosByAuthor = () =>
createSelector(
(state: RootState) => state.todos.todos,
(state: RootState, author: string) => author,
(todos, author) => todos.filter((todo) => todo.author === author),
)
function AuthorTodos({ author }: { author: string }) {
const selectTodosByAuthor = useMemo(makeSelectTodosByAuthor, [])
const todos = useSelector((state) => selectTodosByAuthor(state, author))
}
Of course, with createSlice.selectors
this is no longer possible, as you need the selector instance when creating your slice.
In 2.0.0 we have no set solution for this - a few APIs have been floated (PR 1, PR 2) but nothing was decided upon. If this is something you'd like to see supported, consider providing feedback in the Github discussion!
3.0 - RTK Query
RTK 2.0 was largely focused on core and toolkit changes. Now that 2.0 is released, we would like to shift our focus to RTK Query, as there are still some rough edges to iron out - some of which may require breaking changes, necessitating a 3.0 release.
If you have any feedback for what that could look like, please consider chiming in at the RTK Query API pain points and rough spots feedback thread!