zhifu-platform/ops/haproxy/replicas.cfg

global
    maxconn 2048
    log stdout format raw local0

defaults
    log global
    option httplog
    mode http
    timeout connect 2s
    timeout client 65s
    timeout server 65s
    timeout queue 2s
    retries 2
    retry-on conn-failure
    option redispatch 1

frontend platform
    bind :8080
    default_backend platform_replicas

backend platform_replicas
    balance roundrobin
    option httpchk GET /health/ready
    http-check expect status 200
    http-response set-header X-Zhifu-Replica %[srv_name]
    server api1 host.docker.internal:18001 check inter 1s fall 2 rise 2
    server api2 host.docker.internal:18002 check inter 1s fall 2 rise 2

frontend embeddings
    bind :8200
    default_backend embedding_replicas

backend embedding_replicas
    balance leastconn
    option httpchk GET /health/ready
    http-check expect status 200
    server embedding1 host.docker.internal:8200 check inter 1s fall 2 rise 2
    server embedding2 host.docker.internal:8201 check inter 1s fall 2 rise 2

frontend monitoring
    bind :8404
    http-request use-service prometheus-exporter if { path /metrics }
    stats enable
    stats uri /stats