Imported from openai/plugins (plugins/vercel/skills/cdn-caching/SKILL.md). Install upstream with npx skills add openai/plugins --skill cdn-caching. Copyright stays with the author.

Vercel Caching

You are an expert in understanding Vercel's caching infrastructure, and how the CDN Cache, ISR, and PPR work.

Core Knowledge

• ISR (and PPR, a rendering strategy built on it) is a framework feature — Next.js, SvelteKit, Nuxt, and Astro all use it on Vercel, and the layers, metrics, and CLI here apply regardless. (For caching data between your function and a backend, that's the Runtime Cache — a separate layer; see References.)
• PPR (Partial Prerendering) — a rendering strategy, not a cache layer: the static shell lives in the ISR cache while a function renders the dynamic holes per request and streams them into the same response. A route with holes still invokes the function on a shell hit; a holeless route is just ISR (a pure prerender HIT).

How caching works

Vercel caches at multiple layers between the visitor and your backend. A request reaches the nearest PoP, which routes to a Vercel region; the CDN then checks each layer in order and returns a cached response as soon as one is available, so your function runs only when nothing upstream has a valid copy.

Cache layers

• CDN cache — regional, ephemeral. On a hit the region returns the response with no function call. Reads/writes are free.
• ISR cache — durable, in a single Function region. On a CDN miss, Vercel reads here before invoking your function (cache shielding), then replicates the result back to the CDN. Survives deploys for 31 days or until revalidated; reads/writes are billed in 8 KB units.
• Function invocation — runs only if neither cache has a valid copy. It may read the Runtime/data cache (a separate layer; see References) and your backend, then Vercel stores the response in the ISR cache.
• Image cache — optimized images, cached on the CDN after the first transform.
• Purges propagate globally in ~300 ms.

Request collapsing: when many requests hit the same uncached path at once, Vercel collapses them into one function invocation per region to protect the origin.

Key concepts

• Cache hit rate — share served from cache (HIT/STALE/PRERENDER) versus origin (MISS/REVALIDATED). Measure it over cacheable requests — exclude BYPASS and (not set) (redirects, errors, uncacheable methods), or they drag the ratio down for non-cache reasons. Low hit rate means more origin load and higher latency.

• Revalidation — refreshing cached content. Time-based runs automatically after an interval; on-demand runs when you call an API. Both use stale-while-revalidate: visitors keep getting the cached version while the new one regenerates in the background.

• Invalidate vs. dangerously-delete — two ways to clear content, with very different blast on hit rate:

  • Invalidate (invalidateByTag, Next.js revalidateTag/revalidatePath) = stale-while-revalidate. Keeps serving stale while refreshing in the background → response shows x-vercel-cache: STALE.
  • Dangerously-delete (dangerouslyDeleteByTag, Next.js updateTag or a revalidate with no lifetime) = hard removal. The next request blocks in the foreground to regenerate → x-vercel-cache: REVALIDATED.

• Cache tags & blast radius — tags group cached entries so one call can clear many. A coarse tag attached to thousands of paths has a large blast radius: a single write drops them all and the hit rate collapses until they re-warm. Prefer granular tags (product-${id}) plus a roll-up tag.

• Cache status (x-vercel-cache response header) — the outcome:

  Value	Meaning
  HIT	Served from cache; no function ran
  MISS	Not cached; origin/function ran
  STALE	Served stale while revalidating in background (SWR / invalidate)
  PRERENDER	Served a prerendered ISR/PPR shell
  REVALIDATED	Foreground revalidation after a delete (or Pragma: no-cache)
  BYPASS	Caching skipped (no-store, private, cookies, etc.)

• Cache reason (cacheReason) — the finer explanation of that outcome for a single request. The cache_result metric lumps all MISSes (and all STALEs) together; the reason is the only thing that tells them apart. Nine values, three per group:

  cacheReason	Refines	Meaning
  cold	MISS	Cache empty for this key/variant (first request or evicted); the function ran
  collapsed	MISS	Concurrent requests to one uncached path collapsed into a single invocation
  error	MISS	An error prevented serving from cache
  draft_mode	→ BYPASS	Next.js Draft Mode active — bypassed so editors see live content
  prerender_bypass	→ BYPASS	Prerender-bypass cookie/token present
  crawler	→ BYPASS	SEO-crawler UA — full response served so bots index real content
  stale_time	STALE	Time-based revalidate interval elapsed; regenerating in background (SWR)
  stale_tag	STALE	Tag invalidated (revalidateTag / invalidateByTag); regenerating
  stale_error	STALE	A revalidation attempt failed; serving the last-good copy (a bug signal)

  A raw MISS with reason draft_mode / prerender_bypass / crawler is displayed as BYPASS (all usually expected). The three stale_* reasons separate a healthy time refresh (stale_time) from a broad-tag blast (stale_tag) from a failing regen (stale_error). Read cacheReason from vercel logs or the dashboard Logs "Reason" row — the x-vercel-cache-reason header is internal-only and not visible via curl.

Investigating cache issues

Reach for the Vercel CLI. vercel metrics gives aggregate numbers (requires Observability Plus); vercel logs shows per-request behavior.

Metrics need to be queried by team and project (-S <team> -p <project>). Filter production with -f "environment eq 'production'" (there is no --prod flag). Run vercel metrics schema <metric> to discover dimensions; use -F json for machine-readable output. With -g, remember --limit is per time bucket — omit -g when you need totals across the whole window.

Cache hit rate

Start here for an overall picture of how well caching is working.

Step 1 — overall split. Group vercel.request.count by cache_result. Treat HIT, STALE, and PRERENDER as cache-served; focus investigation on MISS. Exclude BYPASS and (not set) when computing a hit rate over cacheable traffic (see Debugging BYPASS traffic). STALE means stale-while-revalidate is working — dig into revalidation frequency in Analyzing ISR costs, not here.

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "environment eq 'production'" --group-by cache_result --since 24h

Step 2 — where misses concentrate. Split the MISS bucket (and optionally STALE) by path_type, then by route or request_path:

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "environment eq 'production' and cache_result eq 'MISS'" \
  --group-by path_type --since 24h

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "environment eq 'production' and cache_result eq 'MISS' and path_type eq 'prerender'" \
  --group-by request_path --since 24h

What to expect: prerender routes (static shells, ISR pages) should show a high share of HIT/PRERENDER. A prerender path with a disproportionate MISS count is your short list for per-path header inspection (curl above) and code review.

streaming_func routes render dynamically by default, but you can still cache them with Cache-Control headers — matching requests are cached on the CDN. Each cache entry varies by Vary headers (cookies, RSC, etc.) as well as path and query parameters, so expect more cache keys and a lower hit rate than a fully static prerender route.

Analyzing ISR costs

Once you know hit rate, quantify ISR spend and whether revalidation — not traffic volume — is driving it.

Utilization vs. ISR billing. Utilization is vercel.request.count — total request volume. ISR cost is billed separately in 8 KB units: read_units when the regional CDN misses and falls through to the ISR cache, and write_units on every revalidation/regeneration. The regional CDN shields ISR heavily — most requests never touch the ISR layer, so read_units will be far below request count. Do not compare read_units to write_units as a utilization check; focus on write_units (revalidation cost) and how they relate to total traffic.

vercel metrics vercel.request.count -S <team> -p <project> -a sum --since 24h
vercel metrics vercel.isr_operation.write_units -S <team> -p <project> -a sum --since 24h

Write utilization = cache serves ÷ ISR writes — cached reads per regeneration.

# numerator: cache serves — sum the HIT + STALE + PRERENDER buckets
vercel metrics vercel.request.count -S <team> -p <project> \
  -f "environment eq 'production' and (cache_result eq 'HIT' or cache_result eq 'STALE')" \
  --group-by route -a sum --since 24h
# denominator: ISR writes
vercel metrics vercel.isr_operation.write_units -S <team> -p <project> \
  -f "environment eq 'production'" --group-by route -a sum --since 24h

High is good; near or below ~1 means you regenerate about as fast as the page is read (wasted writes) → lengthen the revalidate interval or move time-based to on-demand tag revalidation.

Which routes revalidate most. Break write units down by route and request_path to find paths that regenerate often relative to traffic:

vercel metrics vercel.isr_operation.write_units -S <team> -p <project> \
  -a sum --group-by route --since 24h

vercel metrics vercel.isr_operation.write_units -S <team> -p <project> \
  -a sum --group-by request_path --since 24h

Regeneration vs. serving. Group write units by path_type — concentration in background_func confirms revalidation (not per-request dynamic work) is the cost driver.

Time-based vs. tag-based revalidation. Time-based intervals regenerate on a schedule whether or not content changed — often inefficient. Tag-based on-demand revalidation is usually better, but an overly broad tag has a large blast radius: one invalidate drops every entry that carries it.

• Tag blast radius — group write units by cache_tags. If many unrelated routes show near-identical write counts, a shared hot tag is invalidating them in lockstep (e.g. every blog post rewriting at the same rate because they share one broad blogPost tag):

vercel metrics vercel.isr_operation.write_units -S <team> -p <project> \
  -a sum --group-by cache_tags --since 24h

• What triggered revalidation — group vercel.request.count by triggering_tag to see which tags fire most often (triggering_tag is on request count only, not ISR operation metrics. It is one of the tags that triggered the page to be stale):

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "triggering_tag ne null" --group-by triggering_tag --since 24h

Tags with a large blast radius that revalidate frequently are the usual root cause of high write_units. Prefer granular tags (product-${id}) and on-demand invalidation over short time-based intervals for event-driven content.

Confirm in code. Metrics tell you which tag is hot; the repo tells you why. Grep for the tag's invalidation call site — revalidateTag(, invalidateByTag(, updateTag(, dangerouslyDeleteByTag( — and read the trigger. A CMS webhook or a sync cron that invalidates a broad tag on every event (instead of a specific ${type}:${id}) is the classic amplifier.

Debugging BYPASS traffic

The largest legitimate sources of BYPASS are Draft Mode and SEO crawlers. Draft Mode must bypass cache so editors see live content. SEO bots must receive the full response — especially on PPR routes where the static shell and dynamic holes are assembled at request time — so crawlers index what users actually see. That BYPASS is expected, not a misconfiguration.

Before tuning headers or revalidate intervals, confirm what's left after those two buckets:

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "cache_result eq 'BYPASS'" --group-by bot_category --since 24h

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "cache_result eq 'BYPASS'" --group-by user_agent --since 24h

vercel metrics vercel.request.count -S <team> -p <project> \
  -f "cache_result eq 'BYPASS'" --group-by request_method --since 24h

The Firewall/WAF with the vercel-firewall skill can be used to manage verified SEO crawlers, block abusive bots, and rate-limit junk traffic before it distorts your hit-rate picture.

Reducing ISR cost

• Prefer tag-based over time-based revalidation. Replace short revalidate intervals with on-demand revalidateTag / invalidateByTag when content changes — time-based regeneration runs whether or not anything changed. If using Cache Components, analyze cacheLife calls with the next-cache-components skill.
• Scope tags to specific IDs. Invalidate blogPost:<id>, not a generic blogPost/page tag — one broad invalidate regenerates everything that carries it.
• Tune the revalidate interval where your framework declares it (Next.js revalidate / cacheLife, SvelteKit isr, Nuxt routeRules, Astro). For Next.js Cache Components, see the next-cache-components skill.
• Use CDN-Cache-Control headers to cache dynamic functions.

Inspect one path

curl -sSI https://<host>/<path> | grep -iE 'x-vercel-cache|x-matched-path|cache-control|vary|age|set-cookie'

This zero-dependency first reach shows the status (x-vercel-cache), the cache directives (Cache-Control / CDN-Cache-Control / Vercel-CDN-Cache-Control), and — crucially — x-matched-path, which reveals rewrites like /precomputed/exp~.../... that expose experiment/flag precomputation. vary flags personalization (RSC, cookies); set-cookie forces BYPASS. For a per-phase timing breakdown, vercel httpstat /some/path (CLI v48.9.0+; needs the httpstat tool installed) adds latency stats. A path that should cache but shows MISS/BYPASS usually has private, no-store, max-age=0, a per-request input (cookies/headers/searchParams), or an uncacheable method (see FAQ).

Inspect one request. When metrics or headers give you a request ID, pull the full log record:

vercel logs --request-id <request-id> --json

Use --json so the agent can parse cache status, path, and timing fields programmatically.

FAQ

• What are prerender variant misses? When a route uses a dynamic param, each distinct cache-key variant is prerendered and cached separately, so each variant misses on its first hit per region and low-traffic ones rarely stay warm. The most common modern cause is feature-flag / experiment precomputation — middleware picks a variant per request (/precomputed/exp~.../... paths), and flags × routes × PPR segments multiply into thousands of ISR entries (also a middleware-invocation cost). Fix: collapse the variant matrix (retire finished experiments), or accept the cost.
• Does PPR avoid function invocations? No — a PPR route has dynamic holes by definition, so the cached shell hit still runs the function to fill them. (A route with no holes is just ISR and serves a pure prerender HIT — see Key concepts.)
• Why are there more function invocations than PPR requests? PPR requests have a static shell and a dynamic function invocation. When the static shell needs to be regenerated, it incurs a function invocation on top of the dynamic function for the content.

Related skills

• vercel-firewall — manage verified SEO crawlers, block abusive bots, and rate-limit junk BYPASS traffic.
• runtime-cache — caching data between your function and a backend (per-region key-value / data cache). A different layer from the CDN/ISR caches; use it to cache an API response or query result inside a function.
• next-cache-components — Next.js use cache, cacheLife, cacheTag, and revalidate tuning (one framework's ISR/PPR controls).

References:

• Caching overview: https://vercel.com/docs/caching
• ISR: https://vercel.com/docs/incremental-static-regeneration
• Partial Prerendering (PPR): https://vercel.com/docs/partial-prerendering
• Cache-Control headers: https://vercel.com/docs/caching/cache-control-headers
• Diagnosing and fixing cache issues (full runbook): https://vercel.com/docs/caching/cdn-cache/debug-cache-issues
• vercel metrics CLI: https://vercel.com/docs/cli/metrics
• vercel logs CLI: https://vercel.com/docs/cli/logs