nginx: add Cognito buffer directives to chart ConfigMap (16k)

[elamaran11]

Summary

Add proxy_buffer_size, proxy_buffers, and large_client_header_buffers to helm/accelbench/templates/web-configmap.yaml at 16k, matching the values already in docker/nginx.conf from #45.

Why

PR #45 added the buffer directives to docker/nginx.conf — the config baked into the web image. But the chart mounts a ConfigMap at /etc/nginx/conf.d/default.conf via a subPath mount in helm/accelbench/templates/web-deployment.yaml:32-34, which overrides the baked file. At runtime, nginx reads the ConfigMap version, which had no buffer directives → defaults → 400 Bad Request — Request Header Or Cookie Too Large once Cognito sessions accumulated state.

Confirmed before this PR:

$ kubectl get configmap accelbench-nginx -n accelbench -o jsonpath='{.data.default\.conf}' | grep buffer
(no output)

After applying the chart change to the live cluster:

$ kubectl exec -n accelbench accelbench-web-... -- grep -E "buffer_size|large_client" /etc/nginx/conf.d/default.conf
proxy_buffer_size           16k;
large_client_header_buffers 4 16k;

That's why the 400s recurred a week after #45 merged — the fix never reached the running nginx.

Test plan

• helm upgrade accelbench helm/accelbench -n accelbench --reuse-values succeeds
• kubectl rollout restart deployment/accelbench-web rolls cleanly (2/2 ready)
• kubectl exec into a new web pod shows the buffer directives in the live nginx config
• curl https://modelbench.elamaras.people.aws.dev/ returns 200

🤖 Generated with Claude Code

[elamaran11]

Quick context on why this is still needed despite #45:

PR #45 added the buffer directives to docker/nginx.conf — the config baked into the web image. But the Helm chart mounts a ConfigMap at /etc/nginx/conf.d/default.conf via a subPath mount in helm/accelbench/templates/web-deployment.yaml:32-34, which overrides the baked file. At runtime, nginx reads the ConfigMap version.

The ConfigMap template (helm/accelbench/templates/web-configmap.yaml) on main has no buffer directives, so the live nginx runs on the 4k/8k defaults. Confirmed against the cluster before this PR:

$ kubectl get configmap accelbench-nginx -n accelbench -o jsonpath='{.data.default\.conf}' | grep buffer
(no output)

This PR adds the directives to the ConfigMap template — the place that actually reaches the running nginx. After applying:

$ kubectl exec -n accelbench accelbench-web-... -- grep -E "buffer_size|large_client" /etc/nginx/conf.d/default.conf
proxy_buffer_size           32k;
large_client_header_buffers 4 32k;

That's why the 400s recurred a week after #45 merged — the fix never reached the running nginx.