# xLB configuration.

[server]
bind = "0.0.0.0:8006"
mode = "thin-l7"               # thin-l7 (default) | l4
max_connections = 100000
tcp_nodelay = true
reuseport = false
max_header_bytes = 65536       # thin-l7 only

# Optional TLS termination for client connections. When enabled, clients speak
# HTTPS to xLB; xLB still forwards plaintext HTTP to the gateways.
# [server.tls]
# enabled = false
# cert = "/etc/xlb/tls/fullchain.pem"
# key = "/etc/xlb/tls/privkey.pem"

[timeouts]
connect = "5s"
# first_byte = "60s"           # optional TTFB budget before first response byte (defaults to idle)
idle = "600s"
total = "3600s"
fin = "30s"

[pool]
max_idle_per_backend = 64
# max_conns_per_backend = 0    # 0 = unlimited; positive caps concurrent in-flight connections per backend
# acquire_timeout = "5s"       # max wait for a backend concurrency slot (defaults to timeouts.connect)
idle_timeout = "60s"

[backends]
algorithm = "leastconn"        # leastconn | roundrobin
health_check_path = "/health"
# health_check_headers = ["Authorization: Bearer sk-..."]  # extra probe headers for an auth-protected health_check_path (e.g. LiteLLM /health; see docs/gateways.md)
health_interval = "2s"
health_check_timeout = "2s"
unhealthy_threshold = 2
healthy_threshold = 2
# passive_unhealthy_threshold = 0  # 0 = disabled; positive ejects backend after N consecutive TCP dial failures
# eject_cooldown = "5s"            # cooldown before passively ejected backend is re-admitted

[[backends.servers]]
addr = "10.56.88.32:8001"
weight = 1

[[backends.servers]]
addr = "10.56.88.33:8002"
weight = 1

[retry]
max_retries = -1               # -1 = try all remaining healthy backends; 0 = disabled
# timeout = "15s"              # optional wall-clock budget across all pre-first-byte retry attempts (<= timeouts.total, starts at request arrival)
retry_statuses = [402, 404, 429, 500, 502, 503, 504]
# non_idempotent_retry_statuses = [402, 404, 429, 503]  # optional narrower status list for POST/PUT/PATCH/DELETE (avoids replaying in-doubt 500/502/504 after upstream acceptance)
# respect_idempotency_key = false                       # set true only when upstreams share a deduplication ledger for Idempotency-Key / X-Idempotency-Key
max_body_bytes = 67108864      # 64 MiB max replayable request body for thin-l7 status rerouting / stale-socket retry

[metrics]
enabled = true
bind = "0.0.0.0:9106"