next-intl without middleware

[jgcjordi]

Hi! I'm starting a new project with Next.js and Vercel.
Now I’ve noticed that with only around 2k page visits per week, I already hit the free tier’s “fluid CPU” limit. I'm new on this but 2k not seems to me a high number of visits to reach the limit.

At first, I set all pages with ISR, but most of this usage now comes from the middleware. I’m trying to remove this middleware somehow, since it’s only being used for this library. Honestly, the only useful part for me is the language detection on the home page.

While trying this, I ran into another issue:
I’d like to keep localePrefix: 'as-needed', but according to the docs, it seems that’s not possible.
Is there any workaround for this?

Related discussions:
👉 #975

[amannn]

Hmm, that looks a bit odd indeed. Can you share how many middleware invocations you have?

Based on my local tests with the App Router example, the middleware invocation takes a few ms (typically between 4-8ms).

Let's say we take 8ms as an example and 2h 1m of compute time, that would account for 907,500 invocations!

Is your middleware potentially called on unrelated routes? One thing to be aware of is that Next.js prefetches pages for all links on a page. So if you render a page and it has links to 100 pages, then you don't get 1 middleware invocation per "visit", but 101.

To address your question from above, localePrefix: 'as-needed' indeed requires middleware invocations—at least for your default locale since rewrites are necessary.

[amannn]

As a side note: I found that Vercel observability tools previously reported more rewrites than necessary. This will be fixed in #2051.

[jgcjordi]

Thanks for the answer!
I can see that in the last 12 hours I had around 2k middleware invocations.
If this rate is constant, that would mean around 4k per day and 120k per month.
To reach 2 hours of compute time, the middleware would need to take around 60 ms per invocation — I’m not sure if that’s possible.
It could also be that during the other 12 hours there’s some kind of demand spike. I'll keep an eye on that.

On the other hand, I’ll try disabling prefetch on all links, since it seems that could solve the main part of my problem.
Over the next few days, I’ll share whether it worked and if the “fluid CPU” usage decreases. Thanks a lot!