useSuspenseQuery shows loading state after the first error, and triggers ErrorBoundary for the second

[Guuz]

Hi @TkDodo ! :)
Not sure if this is a bug or just something that i completely misunderstand 😅

We have various screens in our React Native app that load data from an API and use polling (refetchInterval) to keep this data up to date. We've seen some odd things so i'm debugging what is happening exactly when error occur. I'm using this code to debug:

  let c = 0 // Global variable, outside of React Component
  
  // React component
  const { data: foo, error: fooError } = useSuspenseQuery({
    queryKey: ['foo'],
    refetchInterval: 5000,
    queryFn: async () => {
      c++
      console.log('foo called', c)
      await new Promise(resolve => setTimeout(resolve, 1500))

      // if (c % 4 === 0) { // 1 error
      if (c % 4 === 0 || c % 4 === 1) { // 2 consecutive errors
        console.log('foo error')
        return Promise.reject(new Error('foo error internal'))
      }
      console.log('foo success')
      return Promise.resolve('foo')
    },
  })
  console.log({ foo, fooError })

This refetches every 5 seconds. Then for the 4th and 5th refetch it rejects with an error.

I expect that the user always see the succesfully fetched data. This is described in the docs:

Not all errors are thrown to the nearest Error Boundary per default - we're only throwing errors if there is no other data to show. That means if a Query ever successfully got data in the cache, the component will render, even if data is stale.
https://tanstack.com/query/v5/docs/framework/react/guides/suspense#throwonerror-default

You also mention it here: #8985 (comment)
And i totally agree with your statement; you want to show the (old) data, not the error.

What i'm seeing is that the data is loaded and shown normally. Polling works 👍
Then when c%4 === 0 triggers i see the error is "hidden" and the existing data stays on the screen. All good!

But with the next refetch things start "misbehaving":
The spinner is shown...
Then when i enable c%4 === 1 (so 2 consecutive errors) the error is thrown to the boundary...

I made a short movie that shows how the above code behaves:
https://github.com/user-attachments/assets/25a39ee5-1c0e-4d20-9ce5-d02091b8d289

I would expect, after reading the docs and your comments, that any existing data is always shown when refetching/polling fails.
Is this expectation correct? Is there a bug? What causes this behaviour?

[TkDodo]

I would expect, after reading the docs and your comments, that any existing data is always shown when refetching/polling fails.

Yes, I’d expect that too. Can you create a minimal reproduction with stackblitz please?