Stirling-PDF/testing/compose/multinode/nginx.conf

worker_processes 1;
events { worker_connections 1024; }

http {
    # Round-robin across the app nodes. Add a node here to scale out.
    # max_fails=1 marks a node down after a single failure so a dead node drains fast.
    upstream stirling_nodes {
        server stirling-1:8080 max_fails=1 fail_timeout=10s;
        server stirling-2:8080 max_fails=1 fail_timeout=10s;
    }

    # Large uploads (SYSTEM_MAXFILESIZE=100MB) plus headroom.
    client_max_body_size 200m;

    map $http_upgrade $connection_upgrade {
        default upgrade;
        ''      close;
    }

    server {
        listen 8080;

        location / {
            proxy_pass http://stirling_nodes;

            # Graceful failover: retries the other node only on connection-level failures (unreachable/timeout), never on 5xx, so a POST a node already started is never re-sent.
            proxy_next_upstream error timeout;
            proxy_next_upstream_tries 2;
            proxy_connect_timeout 3s;

            proxy_set_header Host              $host;
            proxy_set_header X-Real-IP         $remote_addr;
            proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
            proxy_set_header X-Forwarded-Proto $scheme;

            # WebSocket / SSE upgrade support (policy run streaming).
            proxy_http_version 1.1;
            proxy_set_header Upgrade    $http_upgrade;
            proxy_set_header Connection $connection_upgrade;

            # Stream server-sent events straight through, don't buffer.
            proxy_buffering off;
            proxy_read_timeout 3600s;

            # Surface which app node served the request, so the validate script can prove the LB is spreading load.
            add_header X-Served-By $upstream_addr always;
        }
    }
}