# Qwen3.8-27B load balancer — metadata-only request log.
# Runs as the `lb` container; reaches replicas by compose service name.
worker_processes 1;
error_log /var/log/qwen/lb-error.log warn;
pid /tmp/nginx.pid;
events { worker_connections 1024; }

http {
  # ---------------------------------------------------------------------------
  # PRIVACY: METADATA ONLY.
  # The team's prompts contain proprietary code. This log records timings, which
  # replica served, status and byte counts — and nothing else. Never add
  # $request_body, never log the response, never log $http_authorization.
  # vLLM's own --enable-log-requests writes prompt TEXT to the container log and
  # is deliberately NOT enabled anywhere in this stack.
  # ---------------------------------------------------------------------------
  #
  # TTFT WARNING: nginx cannot measure time-to-first-token on a streaming
  # response. $upstream_header_time is time-to-response-headers, which vLLM emits
  # immediately when the SSE stream opens — long before the first token. Real TTFT
  # comes from the Prometheus histogram vllm:time_to_first_token_seconds, collected
  # per replica by telemetry/scrape.py. Do not read header_time as TTFT.
  # 64K-context prompts exceed nginx's 1 MB default and are rejected with 413
  # BEFORE reaching vLLM. Sized for max-model-len 65536 with headroom.
  client_max_body_size 32m;

  log_format qwen escape=json
    '{"ts":"$time_iso8601","replica":"$upstream_addr","status":$status,'
    '"request_time":$request_time,"header_time":"$upstream_header_time",'
    '"upstream_time":"$upstream_response_time","bytes_out":$body_bytes_sent,'
    '"bytes_in":$request_length,"uri":"$uri","method":"$request_method"}';

  access_log /var/log/qwen/requests.jsonl qwen;

  upstream qwen {
    # Round-robin (nginx default), equal weight — DELIBERATE.
    #
    # least_conn routes toward whichever replica is currently faster. During an
    # A/B that means the faster arm also receives MORE traffic, so the two arms
    # end up with unequal sample sizes drawn over different slices of the day.
    # Per-request rates stay valid either way, but equal-n arms make the
    # crossover analysis clean and the confidence interval honest.
    server replica-gpu0:18020 max_fails=3 fail_timeout=30s;
    server replica-gpu1:18021 max_fails=3 fail_timeout=30s;
    keepalive 32;
  }

  server {
    listen 18000;

    location / {
      proxy_pass http://qwen;
      proxy_http_version 1.1;
      proxy_buffering off;              # required for streaming
      proxy_request_buffering off;
      proxy_read_timeout 1800s;         # a 192K-context prefill can take ~6 min
      proxy_send_timeout 1800s;
      proxy_set_header Connection "";
      # Authorization passes through untouched (and is never logged).
    }

    location = /lb-health { access_log off; return 200 "ok\n"; }
  }
}