Single ISR Function Invocation Triggering Multiples RSC Renders

[ViniciusHack]

Summary

In the problematic traces (ISR Function Invocations), the Vercel Runtime is triggering the exact same RSC GET and prerender route sequence multiple times within a single function invocation. It doesn't have a pattern, sometimes it only prerender once and it's good, other can call RSC GET up to 7 times, taking more than 17 seconds.

Image with multiples prerender route (6 calls):

Good example (1 time call):

Normal Trace (~2.7s):
1x RSC GET -> 1x prerender route -> waitUntil -> Done.
This is the expected behavior.
Problematic Trace (13s+):
The trace shows a loop: RSC GET (Start) -> prerender (End) -> Gap -> RSC GET (Starts Again) -> prerender -> ...
This sequence repeats 7 times sequentially in a single trace.

There is no traffic spike. I am the only user accessing this path.

Is there an internal mechanism in the App Router or Vercel's ISR handler that triggers a "soft retry" if the initial stream hangs or encounters a specific non-fatal error?

Why does the runtime decide to discard the first completed prerender step and start over immediately?

The "Wait" time between these loops is significant. Is the runtime waiting for a socket timeout before retrying the render?

Additional information

No response

Example

No response

[icyJoseph]

Hi, what Next.js version have you got? I am fairly sure the team had been doing some investigation into similar traces and some fixes landed in Next.js 16.1.6

Please do try out with latest next@16.1.6 and post back your results.

[ViniciusHack]

Hi!
I'm using version 16.0.8. I'll try out 16.1.6

[ViniciusHack]

Thanks @icyJoseph, I do notice an improvement. However, I still found a few situations where an RSC GET is followed by another GET in sequence for some reason. Furthermore, there are a multiple RSC GETs (as before), but they are almost instant, taking <5ms instead of >1.5 second.
Also, I see an RSC POST at the end, which I didn't see in the previous version.

Plus, Something that happened in the previous version and persists in this new one is an unexplained long time on the prerender. The fetches take around 300-800ms, but the full prerender usually takes 1.4-2 seconds. Something that is not tracked lasts over 1 second after the requests. I wonder if this is just the React endering process or the result being saved into the cache. It's hard to believe since it's a significant amount of time and there is no Vercel tracing log (like clientComponentLoading) related to it.