Mark Erikson on Redux Toolkit 3.0, the React API That Stalled, and His Biggest Mistake

November 20, 2026

In 2023, to get Redux Toolkit 2.0 out the door, Mark Erikson put his project manager hat on and made a call: nothing in RTK Query would change. He could tell it needed work. He did not yet know what that work was, and finding out would have pushed the release back another year or two. It could wait for a 3.0, maybe a year later.

Three years later, Redux Toolkit is still on 2.x.

That is the kind of decision I wanted to dig into when Mark came on the podcast. He has maintained Redux since 2016, a year after Dan Abramov and Andrew Clark created it, alongside his day job at Replay. Every option he adds and every release he cuts lands in a huge number of codebases, mine included. So I asked him what is planned, what he is waiting on, and what he got wrong.

The plan for 3.0

RTK Query has grown the way most libraries grow. People asked for options, the options got added one at a time, and nobody designed them as a whole. The result works, but it has inconsistent internals and some sharp edges in the API.

Mark's goal for 3.0 is not to change what the library does. It is to review every behavior and option, decide how each one is supposed to work, rebuild the internals to match, and unlock features he has wanted for years but could never get the TypeScript types right for. The next majors will probably be ESM only as well. The 2.0 release was mostly about supporting both CommonJS and ESM, which he said was even more painful than he expected, and Node can now import across the two formats on its own.

The part I found most useful for anyone on RTK today: when he planned the work out, a lot of it turned out to be bug fixes and features that do not need a breaking change. So the roadmap he posted a few weeks before we recorded runs through a series of 2.x releases first, including features people have requested for years. Only after that does it get to 3.0 and the deliberate breaking changes.

The React API that never shipped

I asked how changes in React itself affect Redux. Part of the answer is what you would expect. Server Components and framework data loading mean many apps need less client state. Mixing RTK Query with Server Components is still an open question, and Mark was upfront that TanStack Query has done more work there and that Redux has no official answer yet.

The other part was news to me. Every external state library, Redux and Zustand included, tells React to re-render through one hook: useSyncExternalStore. Mark was its alpha tester. The React team used React Redux's code as the reference, and he swapped early builds into a prototype branch and reported back when half his tests failed.

That hook has a limitation by design. React's concurrent features depend on React owning the state, so it can pause a low priority render and resume it later. If an external store updates in the middle of one, React throws the partial work away and falls back to a synchronous render. That is why no external store is fully compatible with transitions and Suspense, and why Mark keeps answering the same GitHub issue the same way: there is nothing the library can do until React ships a better primitive.

Someone did try to build it. Jordan Eldredge, on the Relay team at Meta, wrote a concurrent store hook with a userland polyfill and opened a draft PR against React last December. Mark thought it was exactly the right idea. Then:

Nothing. Nobody commented on that PR at all.

Between reorganizations at Meta and internal debate about whether this was the right direction, the effort stalled. Mark recently raised it with Seth Webster of the React Foundation, which is taking over React's development process and held a contributor summit in London the week we recorded. He is hopeful, but for now he is waiting.

The biggest mistake: React Redux v6

I asked what he would do differently, and he did not need time to think. It was shipping React Redux version 6.

In 2018 the React team got on a call with the maintainers of Redux, MobX, and Apollo to warn them about concurrent rendering and a problem called tearing, where a paused render resumes with newer store data and two parts of the UI show different values. Mark wanted to get ahead of it. In v6, instead of every connected component subscribing to the store, only the Provider subscribed and passed the state down through the new context API, so React could guarantee the whole tree saw the same value.

I was making essentially the worst possible architectural decision for React performance that I could have possibly made.

Every store update became a state update at the root of the tree, and React had to walk down from there every time. His benchmarks said it was a little slower. Within a month or two of the December 2018 release, users were reporting apps that had become much slower. He rewrote the internals with hooks, went back to per-component subscriptions, and shipped v7 a few months later. That rewrite is also what made useSelector and useDispatch possible.

I had good intentions. I didn't think through all the possibilities. I learned my lesson.

I have a lot of respect for a maintainer who will say that on a recording about a library with this many users.

RTK Query is already in your package.json

This is my own soapbox. On at least three teams I have seen Redux Toolkit installed for state, with SWR or TanStack Query installed next to it for data fetching. My guess is that most of those teams never knew RTK Query ships inside the package they already have.

Mark's guess was different: TanStack Query is simply the popular default now, or people believe Redux is only for client state. Either way, you can do it, but you are skipping something built on the Redux pieces already in your bundle, with DevTools support included.

So I asked whether RTK Query should become its own library to make it more visible. "That's definitely not going to happen." The question he usually gets is the reverse, why Redux Toolkit is not just merged into the core redux package. His design is one install with separate entry points, so you only pay for what you import:

// UI-agnostic core: endpoints and the thunks that do the fetching
import { createApi } from '@reduxjs/toolkit/query'

// The same thing, plus generated React hooks
import { createApi } from '@reduxjs/toolkit/query/react'

The same split runs through the whole stack. A Redux store is plain JavaScript that you could write in about 25 lines. Mark figures more than 95 percent of users pair it with React, but the core has been kept UI-agnostic on purpose.

Where Redux stands, in his own numbers

I still reach for Redux Toolkit when I start a new project, so I wanted his honest read on where the library sits today. He did not dress it up.

By his rough estimate, Redux peaked around 2017 or 2018, when 60 to 65 percent of React apps used it. Today he puts it at 20 to 25 percent, and a lot of those are apps that picked it years ago. npm downloads keep climbing, but only because the whole ecosystem keeps growing.

He also pointed out that the maintainers never asked for that peak. The community decided that "if you use React, you have to use Redux," copied the deliberately verbose patterns from the docs, and put Redux in places it never belonged. Then TanStack Query, Apollo, and SWR showed that most apps mainly need to fetch server data and cache it, and Zustand covered the small amount of client state left over.

His position is that both things are true. Redux is older and no longer the default. It is also mature, battle tested, and still a good fit for large apps, and he has talked to teams at very large companies who told him they could not have scaled their codebase without it.

Not everything belongs in the store

We had François Best, the creator of nuqs, on the show earlier, so I asked Mark what he thinks of keeping state in the URL. He is fully for it. Shareable links, deep links to a specific modal or filter, and state that survives a refresh are all real benefits, and nuqs makes it feel almost like useState.

That fits a bigger correction he made. "Put all your state in Redux" was an early sales pitch that people took more literally than they should have, which is how every keystroke of a form ended up in the store. Form state can stay local, or in something like React Hook Form, with one action dispatched on submit. Deciding what state lives where, and which tool fits each kind, is the actual job.

From "never" to all of his code

Mark used to swear he would never use AI to write code. In March 2025 he went through what he described as a month-long depressive episode over what AI was doing to the industry and to the part of programming he enjoys. That September he tried it properly. Now he writes all of his code with it, and in April he published two long posts about the journey and his setup.

What stood out is the limit he sets for himself. He understands the logic that leads from one agent session to a hundred running overnight, and he is choosing not to follow it.

I have built my career on understanding the code.

He keeps three or four sessions active, makes the decisions himself, and still estimates he is five to ten times more productive. A new React Redux hook built on signals and proxies, and all of the 3.0 prototyping and roadmap planning, came out of that workflow. When the agents produced hundreds of thousands of words of research notes, he had them boiled down to the ten decisions he needed to make.

For the rest of us, the practical note is that agents already know Redux Toolkit, createSlice, and RTK Query. Where he has seen them slip is reaching for createAsyncThunk with an inline fetch when the codebase uses RTK Query, so tell them which one to use. The RTK package also ships its own skills files with best practices for agents to pick up from node_modules.

What I took from it

Most of what Mark described was a decision about scope. Leave RTK Query out of 2.0 so the release could ship. Keep it inside the Redux Toolkit package but behind its own entry point. Roll back v6 within months instead of defending it. Wait for React to ship the right primitive instead of faking one. Cap his own AI use at the point where he still understands the code.

Those calls are not visible from the outside when you run npm install. Hearing the reasoning behind them is the reason I do these interviews.

The full conversation with Mark Erikson is out now on Talks with Ido Evergreen. Watch it on YouTube.