ISR is stale-while-revalidate for whole routes (shown here with cacheComponents off, the default). A route with export const revalidate = 10 is prerendered by next build (○ /isr 10s 1y) and served from the server cache with Cache-Control: s-maxage=10, stale-while-revalidate=31535990.
Inside the window, responses are x-nextjs-cache: HIT. Once the entry is older than 10 seconds, nothing happens until a request arrives. That request still gets the old page, marked STALE, and starts a regeneration in the background. When it finishes, the new page replaces the entry and the next request is a HIT with new data. So freshness lags by one request, and nobody waits for a render. Per the docs (not run here), if regeneration throws, the last good page keeps being served and the next request retries.
On-demand revalidation skips the timer. A route handler (or Server Action) calls revalidatePath('/isr') or revalidateTag('v', 'max'). In my runs, after revalidatePath the next request was a MISS that rendered fresh data; after revalidateTag(…, 'max') the next requests were STALE until the background render finished.
Across instances, it doesn't propagate. The cache and its tag state live in each process: memory in front of the .next folder on disk. A call reaches only the instance that received it. I ran three next start processes: A and B on the same build folder, C on a copy. After revalidatePath on A, A served the new page; B and C kept serving the old one with HIT. To make revalidation global you configure a shared cacheHandler whose tag index lives in Redis or similar, so a call on one instance invalidates entries on all.
All of this was Next.js 16.4.5, next build && next start, Node 24.19.0.