Envoy Gateway 504 Upstream Request Timeout After 15 Seconds

Envoy Gateway v1.9.2 gives an HTTP request about 15 seconds when the HTTPRoute does not set a timeout:

• A backend that sleeps for 3 seconds still returns HTTP 200
• A backend that sleeps for 16 seconds returns HTTP 504 and the body upstream request timeout

The proxy stops the wait. It does not wait for the app.

That 15 second value is timeouts.request, an absolute deadline for the client call. It is not the gap between upstream reads. If you are moving an ingress-nginx proxy-read-timeout, read that migration before you copy the number onto this field.

Why the request returns 504 at 15 seconds

A 3 second sleep fits the default clock. Send it and read the status. The application answered, and the body was slept=3.0:

curl -sS -D - --max-time 25 -o slept.txt -H 'Host: app.example.com' \
  'http://127.0.0.1:19080/?seconds=3'

HTTP/1.1 200 OK
server: BaseHTTP/0.6 Python/3.12.15
date: Sat, 03 Oct 2026 19:18:13 GMT
content-type: text/plain
content-length: 10

A 16 second sleep does not fit. The 504 is timestamped 19:18:28, about 15 seconds after that 200, not when a 16 second sleep would have finished:

curl -sS -D - --max-time 30 -o slept.txt -H 'Host: app.example.com' \
  'http://127.0.0.1:19080/?seconds=16'

HTTP/1.1 504 Gateway Timeout
content-length: 24
content-type: text/plain
date: Sat, 03 Oct 2026 19:18:28 GMT

The body is upstream request timeout. In this test that plain-text line came from Envoy. It is not an nginx HTML 504 page.

These captures used a local address for the Gateway. Send the same Host and path to the hostname your Gateway publishes.

Set timeouts.request

Gateway API puts the deadline on the HTTPRoute rule. The values are durations:

• 60s is sixty seconds. A bare 60 is not valid here.
• 0s disables this request timeout. Use that only when the call is allowed to run without this deadline. This page did not apply 0s.

60s below is a stand-in. Use the longest call your clients should wait for:

apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
  name: web
spec:
  parentRefs:
  - name: eg
  hostnames:
  - app.example.com
  rules:
  - timeouts:
      request: 60s
    backendRefs:
    - name: web
      port: 80

A shorter check used request: 4s and backendRequest: 4s:

• A 3 second sleep returned HTTP 200 and slept=3.0 at 19:19:13
• A 5 second sleep returned HTTP 504 and upstream request timeout at 19:19:17, about 4 seconds later, not 15

Prefer these HTTPRoute fields for ordinary HTTP:

• A BackendTrafficPolicy request timeout is a different control. Use it when the route cannot express the timeout you need, such as some gRPC streams.
• Do not expect that policy to override an HTTPRoute that already sets timeouts.request.

An upgraded WebSocket is not this clock. WebSocket disconnects are the idle stream, not timeouts.request.

Set timeouts.backendRequest

timeouts.request and timeouts.backendRequest are different budgets:

• request is the whole client call
• backendRequest is one call from the gateway to a backend
• backendRequest cannot be longer than request. The API rejects that HTTPRoute.

timeouts:
  request: 10s
  backendRequest: 2s

Client request budget     10s
Backend attempt budget     2s
                           deadline for that backend call

With those two values, a 3 second sleep returned HTTP 504 and upstream request timeout. The 10 second client budget did not keep the wait open. The 2 second backend clock ended it.

Use the shorter field only when you need it:

• backendRequest when one upstream try must be shorter than the overall client deadline
• request when you only need that overall deadline

Troubleshooting

Summary

With no HTTPRoute timeout, Envoy Gateway v1.9.2 returned HTTP 200 for a 3 second sleep and HTTP 504 for a 16 second sleep. The 504 body was upstream request timeout, and the wait stopped at about 15 seconds.

timeouts.request is that deadline. At 4 seconds, a 3 second sleep succeeded and a 5 second sleep returned 504. timeouts.backendRequest of 2 seconds, under a 10 second request budget, returned 504 for a 3 second sleep. backendRequest cannot be longer than request.

This is not proxy-read-timeout, and it is not the timer that closes an idle keep-alive connection after the response is done.

References

• Envoy Gateway HTTP timeouts (v1.9)
• Gateway API HTTPRoute timeouts