Timeout to service that was working

I'm at my wits end here. I have a service exposed via Gateway API using Envoy Gateway. When first deployed it works fine, then after some time it starts returning:
`upstream connect error or disconnect/reset before headers. reset reason: connection timeoutupstream connect error or disconnect/reset before headers. reset reason: connection timeout`
If I curl the service from within the cluster, it responds immediately with the expected response, so the service is fine. But accessing from a browser returns the above. It's just this one service, I have other services in the cluster that all work fine. The only difference with this one is it's the only one on the apex domain. Gateway etc yaml is:
```
---
apiVersion: cert-manager.io/v1
kind: Certificate
metadata:
name: example
spec:
secretName: example-tls
issuerRef:
group: cert-manager.io
name: letsencrypt
kind: ClusterIssuer
dnsNames:
- "example.com"
- "www.example.com"
---
apiVersion: gateway.networking.k8s.io/v1
kind: Gateway
metadata:
name: example
labels:
app.kubernetes.io/name: envoy
app.kubernetes.io/instance: envoy-example
annotations:
kubernetes.io/tls-acme: 'true'
spec:
gatewayClassName: envoy
listeners:
- name: http
protocol: HTTP
port: 80
- name: https
protocol: HTTPS
port: 443
tls:
mode: Terminate
certificateRefs:
- kind: Secret
name: example-tls
---
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
name: example-tls-redirect
spec:
parentRefs:
- name: example
sectionName: http
hostnames:
- "example.com"
- "www.example.com"
rules:
- filters:
- type: RequestRedirect
requestRedirect:
scheme: https
---
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
name: example
labels:
app.kubernetes.io/name: envoy
app.kubernetes.io/instance: envoy-example
spec:
parentRefs:
- name: example
sectionName: https
hostnames:
- "example.com"
- "www.example.com"
rules:
- matches:
- path:
type: PathPrefix
value: /
backendRefs:
- name: example-service
port: 80
```
Nothing in pod logs except the log version of the response at the top (timeout).

If it just never worked that would be one thing. But it starts off working and then at some point soon after breaks. Anyone seen anything like it before?