Which server sent this error page? A lookup table
An error page tells you which server sent it before it tells you what broke. Here are the default error page texts and headers of Apache, nginx, Varnish, HAProxy, CloudFront, Cloudflare and Google, captured live, and what each one means.You can usually tell which server sent an error page from its wording and its response headers: every web server, cache, load balancer and CDN ships its own default error text, and most of them also name themselves in the Server header. Knowing which layer answered tells you where to look, because the layer that shows the error is often not the one that broke.
A 503 from a cache in front of your site means the cache could not reach your server. A 503 from your server itself means the server is up and refusing work. Same status code, different hop, different fix.
Every text below was captured on the day we wrote this: Apache 2.4.69, nginx 1.29.8, Varnish 7.7.3 and HAProxy 3.2.26 running in containers with their default error pages, each pointed at a backend that was not running, plus live requests to CloudFront, Cloudflare and Google.
The lookup table
| Page text you see | Status | Headers that confirm it | Who sent it |
|---|---|---|---|
| "The server is temporarily unable to service your request due to maintenance downtime or capacity problems. Please try again later." | 503 | Server: Apache/... |
Apache. With mod_proxy, it means Apache could not reach the app behind it |
| "Your browser sent a request that this server could not understand." | 400 | Server: Apache/... |
Apache rejected a malformed request, for example a bad Host header |
| "The requested URL was not found on this server." | 404 | Server: Apache/... |
Apache |
| "502 Bad Gateway" with "nginx/1.29.8" printed under it | 502 | Server: nginx/1.29.8 |
nginx could not get a response from its upstream |
| "Error 503 Backend fetch failed" and "Guru Meditation: XID: ..." | 503 | Server: Varnish, X-Varnish, Via: ... (Varnish/7.7) |
Varnish, common in front of Magento and other busy sites, could not fetch from its backend |
| "No server is available to handle this request." | 503 | none: HAProxy sent no Server header |
HAProxy has no healthy backend server |
| "ERROR: The request could not be satisfied" | 403, 502, 504... | Server: CloudFront, X-Cache: Error from cloudfront, X-Amz-Cf-Pop |
Amazon CloudFront |
| "error code: 1001" (plain text to non-browsers) | 409 | Server: cloudflare, CF-RAY |
Cloudflare edge, before your server was involved |
| "404. That's an error. ... That's all we know." | 404 | the page names Google itself | A Google server, not yours |
Two details stand out. Varnish puts its own status text in the status line, not only in the body: it answered HTTP/1.1 503 Backend fetch failed rather than the usual 503 Service Unavailable. And HAProxy was the only one that did not name itself at all, so a bare 503 with "No server is available" and no Server header points to HAProxy by elimination.
How to read the headers in one command
The page text can be customised; the headers usually are not. This prints the status line and every header that names a layer:
curl -s -o /dev/null -D - https://example.com/ \
| grep -iE '^(HTTP|server|via|x-cache|x-varnish|cf-ray|x-amz-cf-pop)'Against the CloudFront edge we tested, with a hostname that no distribution was configured for, it returned:
HTTP/1.1 403 Forbidden
Server: CloudFront
X-Cache: Error from cloudfront
Via: 1.1 2d60fba3896e0efac8b6f5da1b702c20.cloudfront.net (CloudFront)
X-Amz-Cf-Pop: BUD50-P3X-Cache: Error from cloudfront is the useful line: CloudFront generated the error itself. A response that came from your origin through CloudFront would carry your origin's status and a Miss from cloudfront or Hit from cloudfront value instead.
What each one means for your fix
- The server is temporarily unable to service your request due to maintenance downtime or capacity problems.: despite the wording, nobody declared maintenance. Apache prints this default text for any 503 it generates itself; behind a reverse proxy setup it usually means the application server Apache forwards to is down or full.
- Your browser sent a request that this server could not understand.: a 400 caused by the request itself: a malformed Host header in our test, oversized cookies or headers in many real reports. It blames the browser, but a proxy in between can be the one sending the bad request.
- 502 Bad Gateway: nginx is up; the app behind
proxy_passis not answering (502) or not answering in time (504). The 503 vs 504 breakdown for nginx, Cloudflare and ALB covers what triggers each. - Error 503 Backend fetch failed: Varnish generated the 503 because the fetch from the backend failed at the connection level. If your backend itself had returned a 503, Varnish would have passed that through instead. The Guru Meditation XID is a request ID for
varnishlog, not a diagnosis. - No server is available to handle this request.: every server in the backend is marked down by health checks or none is configured. Look at the HAProxy stats page or logs for why the checks fail.
- ERROR: The request could not be satisfied: the status above the text says which case: 403 is often a WAF rule or a hostname not added to the distribution, 502 means CloudFront could not connect to the origin or got an invalid response, 504 means the origin timed out.
- error code: 1001: its own codes (1xxx for refusals, 52x for origin problems) identify the case exactly; see Cloudflare 5xx errors and Cloudflare 1xxx errors. Note that curl and other non-browser clients get a one-line
error code: 1001instead of the branded HTML page. - 404. That's an error.: a Google server returned the error, not yours. If a check against your own domain gets this page, your domain most likely points at a Google-hosted service that has nothing configured for that URL.
AWS Application Load Balancers name themselves too: AWS documents that an ALB adds server: awselb/2.0 to responses unless the routing.http.response.server.enabled attribute is turned off. A 502 or 504 with that header came from the load balancer, so check target health in the target group first.
When the error page lies about the layer
Two cases make the page misleading:
- A custom error page set up on one layer can show for errors from another. A branded "we'll be right back" page on the CDN tells you nothing about which hop failed; the headers still do.
- A layer can pass through another layer's error untouched. An nginx 502 can reach the visitor through a CDN unchanged, with the CDN adding only its own headers. Read the body for the generator and the headers for every hop it went through.
Make your monitoring tell them apart
An uptime check that only looks at the status code sees "503" in every case above. WebPixie's uptime monitoring lets you set the expected status code and a body keyword, so a check can require a word that only your real page contains. A branded CDN or proxy page that answers with 200, such as a maintenance page, then fails the check on the missing keyword instead of passing as up. When a check fails, the table above tells you which layer to look at.